The Unmetered Rack

Eighteen months ago I stopped thinking about GPU hours and put the whole business on monetizing AI compute by the token. That call was right and I would make it again. What I got wrong was the step after it. The token is the right unit to count. It is the wrong unit to put on an invoice. This piece is what I found when I went looking for why, and why the contract I now sell is an unmetered rack.

The same company that pays a flat monthly fee against roughly 325,000 GPUs turns around and sells a million tokens for fifty dollars. And NVIDIA, which also writes the inference software, gives the models away and runs hosted endpoints where it could meter anything it wanted, still prices every one of those in capacity units: per GPU, per year, per gigawatt. Every capital-intensive network in the record kept a meter while supply was short, then put a capacity charge in front of it once unit cost collapsed in a market that was still growing. The wholesale layer of AI already runs unmetered. The retail layer does not. That gap is the subject.

The argument is not that the token stops mattering. The token is the unit of output of the factory. What moves is the invoice. The buyer stops paying per token and starts paying for the machine that makes them.

I have tried to make this checkable rather than persuasive. It is built on an evidence ledger of more than 900 facts, and every load-bearing number in it was verified independently against the filing, the pricing page or the paper it came from. Where the evidence cuts against me, I say so and keep going. Where two sources disagree, both are on the page.

---

#0. Two invoices (cold open)

What Anthropic pays

$1.25B/month

for approximately 325,000 GPUs

• UnitThe machine
• MeterNone
• Quiet monthSame price
• Busy monthSame price
• Implied$5.27 per GPU-hour

SpaceX Form S-1, May 20 2026

the same company

What Anthropic charges

$50/million tokens

Claude Fable 5.1 output, list price

• UnitThe token
• MeterAlways on
• Quiet monthYou pay less
• Busy monthYou pay more
• Capacity tierNo longer for sale

Anthropic pricing page, September 2026

Eighteen months ago I stopped thinking about GPU hours. I put the whole business on one idea instead: monetize AI compute by the token.

That was the right call and I would make it again tomorrow. A GPU hour tells you how long you rented a box. It tells you nothing about what the box produced. The token is the unit of output of the factory. Energy goes in, tokens come out, and that conversion is the business. Once I started measuring the work in tokens, every question I had about the economics got an answer.

Then I did the obvious next thing, and that is where it went wrong. I assumed that because the token is how you measure the work, the token is also how you bill for it.

Counting and billing are two different jobs. The token is the right unit to count. It is the wrong unit to put on an invoice.

I did not work that out from a spreadsheet. I worked it out from customers who kept asking me the same question in different words. They wanted to know what the thing was going to cost them next quarter, and I could not tell them, because the answer depended on how many times their own people ran their own work. Every improvement they made to their systems showed up on my invoice as a bigger number. I was charging them for using the machine they had already decided to buy.

Then I read two invoices that belong to the same company, and the whole thing came into focus.

#The invoice Anthropic pays

SpaceX filed a Form S-1 on May 20, 2026. Inside it are Cloud Services Agreements under which Anthropic pays $1.25 billion per month through May 2029, terminable by either party on 90 days' notice.¹ The filing's own words: "the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee. The agreements may be terminated by either party upon 90 days' notice." The financial notes date the agreement to May 3, 2026.¹

The June 3, 2026 amendment restates the contract as capacity across COLOSSUS and COLOSSUS II covering approximately 325,000 NVIDIA GPUs, with the 90-day termination right available only after an initial three-month period.²

Look at the shape of that. A flat monthly fee against approximately 325,000 GPUs, with no meter attached. Nobody is counting tokens on that invoice. Nobody is counting requests or cache writes either. The number is the same in a quiet month and a busy one.

The arithmetic prices out at about $5.27 per GPU-hour, and about $45 billion over 36 months if the contract is never canceled.²

Two things keep that from being a clean story, and both belong here. SpaceX describes the arrangement as monetizing a portion of its compute while retaining the right to reallocate that capacity to its own initiatives, so it is not a dedicated whole-facility rental.² And the term is contested in public. The S-1 frames a three-year commitment through May 2029, while Musk publicly called it "a 180 day lease with 90 day notice mutual cancellation thereafter".³ The buyer's exposure is real but it is short.

What is not contested is the unit. The buyer pays for the machine. The machine does not send a bill per token.

#The invoice Anthropic sends

Now open the pricing page of the company writing that check.

Those are the live list prices.⁴ A million Fable 5.1 output tokens cost $50. Cache writes are their own line item, $12.50 for five minutes and $20 for one hour on Fable 5.1.⁴ Batch work is 50% off.⁴ The increase to $3/$15 that had been scheduled for September 1, 2026 was canceled.⁴

Every one of those lines is a meter. The price moves with how much the customer uses. There is no version of that table where a buyer knows next quarter's number.

#The capacity product with a banner on it

You would expect the fix to be a capacity product. Buy the machine, stop counting.

It exists, and it is closed. Anthropic's service-tiers documentation says it plainly: "Priority Tier capacity commitments are no longer available for purchase. Organizations with an existing commitment can continue to use Priority Tier through their contract end date".⁵ Anyone who needs guaranteed capacity is directed to sales.⁵

The same wall stands at the cloud. On Amazon Bedrock, the list of models supported for Provisioned Throughput stops at Claude 3.5 Sonnet v2 from October 2024 and Llama 3.2. No Claude 4.x or Claude 5 model can be purchased as provisioned capacity.⁶

What a buyer can get instead is a ceiling. Anthropic's rate-limit page caps Scale-tier spend at $200,000 per month and sets Opus 5 output at 2,000,000 tokens per minute, and it states that all limits are maximums, not guaranteed minimums.⁷

So the same company buys compute flat, on a fixed monthly fee against roughly 325,000 GPUs, and sells it by the token. The one product that would let a customer do what it did is marked no longer for sale.

I do not think that is hypocrisy. I think it tells you which side of the transaction understands the asset. The party that knows exactly what compute costs buys the machine. The party that does not know is handed a meter.

The question this piece is going to answer is why the meter is still there at all, and how long it lasts.

---

#1. What a meter is for

I am a tech guy at heart. Most of my working life has been spent near machines that cost a fortune to build and almost nothing to run for one more minute. That work taught me something simple. A meter is not a principle. It is a rationing device.

Select a point on the line.

You install a meter when the next unit is scarce or expensive. You take it out, or shrink it to nothing, when the next unit is cheap and you want more of it consumed. Every capital-intensive network I can find in the record has run that same play. The meter comes first. The capacity charge comes after.

#The first electricity contract had no meter

Thomas Edison's first contract for electricity, in 1881, was a charge per lamp installed.⁸ You paid for the sockets you were entitled to light. Nobody counted the electrons.

The meter arrived with the industry, and so did the argument about what it should charge for. In 1892 the British engineer John Hopkinson differentiated fixed and variable costs in electricity rates, and that two-part rate was promoted in the United States by Samuel Insull.⁸ In the mid-1890s Arthur Wright introduced demand charges based on maximum actual demand, after controversy over price discrimination in special contracts drove a search for a consistent pricing scheme.⁸

Insull learned of demand-metered billing in Brighton, England in 1894. By 1897 he was offering Chicago customers two-tiered electric rates, and many homeowners found their bills lowered by 32% within a year.⁹ That is the whole idea in one sentence. Charge for the machine standing ready. Charge very little for the unit it produces. Watch consumption climb.

The two-part rate dates to 1892. It was not invented by a software company, and it was not invented to be generous. It was invented because a utility with heavy fixed costs and falling marginal costs makes more money when the meter stops governing behavior.

#December 1996

The clearest modern run of this experiment is AOL. Before December 1996, AOL charged $9.95 per month for five hours of access, with additional time at $2.95 per hour.¹⁰ Then the plan became $19.95 per month for unlimited access.¹⁰

What happened next is the part people forget. Average usage went from seven hours per month to 23 hours per month.¹¹ Andrew Odlyzko, studying the same episode, records that usage per person tripled over the following year, and that it took a year only because AOL could not expand capacity quickly enough to satisfy demand.¹²

The transition was ugly. About 15 percent of AOL subscribers got a busy signal during the most popular evening hours, and daily sessions rose by one-third to 9 million.¹³ A spring 1997 study found AOL had a 60.3% call failure rate, roughly double competitors, and AOL roughly doubled its modems under pressure from attorneys general.¹⁴ AOL was not even first. AT&T WorldNet offered unlimited dial-up at $19.95 in early 1996, months ahead of AOL.¹⁴

So the flat rate broke the network for a while. It still won. In February 1998 AOL raised the flat rate 10% to $21.95, with more than 11 million customers.¹¹ Customers absorbed a price increase on a service that had been failing to answer the phone, because the invoice no longer punished them for using it.

Odlyzko tells one story I have never been able to shake. An AOL member insisted she was being cheated by hourly pricing even after Steve Case showed her she would pay more under the $19.95 flat rate. "I don't care," she told an incredulous Case. "I am being cheated by you." ¹² Buyers will pay a premium to stop counting. Counting has a cost, and it lands on the person doing the counting.

#The rule Odlyzko wrote down

Odlyzko's 2001 survey of communications pricing history reaches a conclusion that reads like a design note for anyone selling compute today. Even tiny charges based on utilization decrease usage substantially, and in a rapidly growing market it is in the service providers' interest to encourage usage, which argues for simple, preferably flat rate, pricing.¹²

He also states the condition, and the condition is the part that matters for this essay. When service costs decrease, those arguments prevail over the need to operate a network at high utilization levels and to extract the highest possible revenues.¹²

Two conditions, then. Unit cost is falling. The market is still growing. When both hold, the seller puts a capacity charge in front of the meter and makes the unit price close to nothing.

#The same move, over and over

Netflix introduced its monthly subscription with unlimited rentals and no due dates in September 1999, and dropped the per-rental model by early 2000.¹⁵ Blockbuster held the meter for five more years. On December 14, 2004 it announced the end of late fees effective January 1, 2005, estimating that late fees would otherwise have contributed approximately $250 to $300 million to 2005 operating income.¹⁶ A quarter of a billion dollars of income was sitting inside the penalty for using the product wrong. It went away anyway, because a competitor had removed the reason to count.

Broadband ran the same arc in the other direction and back. Comcast imposed a 250GB monthly data cap in 2008, suspended it in 2012, and trialed a 300GB allowance with $10 per 50GB overage as streaming grew.¹⁷ On June 26, 2025 Comcast eliminated home data caps on Xfinity Internet, with all plans unlimited and one-year or five-year price locks.¹⁸

There is a buyer-side version of this too, and it is not flattering. DellaVigna and Malmendier found that health-club members on flat monthly contracts over $70 attended 4.3 times a month. They paid more than $17 per visit when a $10 per-visit pass was available, forgoing about $600 over their membership.¹⁹ People overpay for flat rates. Sellers know it. That is one reason the capacity charge survives contact with a finance department.

#Cloud already sells the machine, not the minute

Enterprise buyers have been signing capacity contracts for years without calling them that. AWS Reserved Instances are one- or three-year commitments that cannot be canceled. You are billed a discounted hourly rate for every hour within the term, regardless of whether the Reserved Instance is being used.²⁰ Savings Plans advertise savings of up to 72% compared to On-Demand prices in exchange for a consistent hourly spend commitment.²¹

The aggregate is large. As of June 30, 2026 Amazon reported approximately $496 billion of customer commitments, primarily AWS, not yet recognized as revenue, with a weighted-average remaining contract life of 6.4 years.²² The on-demand meter is the shop window. The book of business behind it is a capacity book with a six-year average life.

#Paying for a machine to exist

Power markets went further. They built a product whose entire content is the promise of capacity. PJM describes capacity as a commitment of resources to deliver when needed, particularly in case of a grid emergency. It is procured three years forward through auction, and PJM likens the capacity payment to an insurance policy paid on top of energy.²³

Tolling agreements do the same thing at the level of a single plant. The toller has the right but not the obligation to have the plant convert its fuel to electricity at a set heat rate. The toller pays a capacity payment for dispatch rights, and keeps paying it even when delivered power falls short of contract.²⁴ Accounting guidance treats those contracts as leases of the plant, because the toller obtains dispatch rights and bears facility-specific risk through fixed capacity payments, with only a capacity-payment reduction as remedy for non-performance.²⁴ When a buyer pays that way, the accountants stop calling it a purchase of output and start calling it a lease of the machine.

#The counter, and it is a real one

Every flat-rate win above happened while supply was getting cheaper and more plentiful. Where supply is short, the meter does not just survive. It runs hot.

Look at the same PJM capacity market. The 2027/2028 Base Residual Auction, held December 17, 2025, cleared at the FERC-approved cap of $333.44 per MW-day UCAP across the whole RTO. Cleared supply times price came to $16.4 billion, 6,623 MW short of the reliability requirement. PJM's own simulation shows the uncapped price would have been $529.80 per MW-day and $26.3 billion.²⁵ The 2028/2029 auction, held July 14, 2026, again cleared at the cap, $325 per MW-day UCAP, for $16.4 billion, 6,831 MW short of the reliability requirement.²⁶ PJM said these were the first auctions in its history in which the entire RTO fell short, and cited continued additions of large data center loads.²⁶ Enel North America reports the 2027/2028 price as $121,705 per MW-year, the third consecutive year at the price cap. Data centers accounted for nearly the entirety of the 5,250 MW peak-load increase over the prior auction.²⁷

That is what scarcity does to a capacity price. It pins it to the ceiling, three years running.

Software does the same thing when the vendor holds the constraint. IBM mainframe software is priced on MSUs metered by a rolling four-hour average peak, and customers pay IBM monthly based on the peak R4HA achieved for each software product.²⁸ IBM did eventually offer the alternative. In May 2019, generally available June 21, 2019, it introduced Tailored Fit Pricing. One option, the Enterprise Consumption model, uses hourly MSUs aggregated annually. The other, the Enterprise Capacity model, offers full-capacity licensing at a predictable monthly cost.²⁸ It took decades, and it took competitive pressure.

Sometimes the meter simply moves out of sight. Spotify sells a flat retail subscription, $12.99 for the individual plan as of February 2026, to 300 million paying subscribers as of June 2026. It pays about 70% of its total revenue to rights holders on a pro-rata, per-play basis.²⁹ The listener never counts a play. Somebody still does.

And flat pricing has a cost the buyer should see. Flexera's 2026 State of the Cloud report, with 753 respondents, was published March 18, 2026. It found wasted cloud spend rose to 29%, the first increase in five years, and attributed that to surging cloud-based AI workloads.³⁰ Remove the meter from the invoice and you have to keep it on the dashboard. The instrument stays. The bill changes.

So the objection to everything in this section is short and fair. Flat rate wins under abundance, and scarcity keeps the meter. I accept that framing completely. It means the argument about AI pricing is not an argument about pricing philosophy at all. It is an argument about supply.

Which leaves one question worth the rest of this essay. Is compute scarce or is it being manufactured into abundance, and who is doing the manufacturing.

---

#2. NVIDIA's price list has no token on it

When I want to know what a company believes, I go read what it charges for.

Not the keynote. Not the analyst deck. The licensing guide.

It is a short document with a table in it. There is no token on that table. There is no token anywhere on the page.

#The objection to get out of the way first

NVIDIA builds machines, so of course its price list is not written in tokens. A turbine maker does not bill per kilowatt-hour either. If that were all this was, it would prove nothing.

But NVIDIA is not only the manufacturer here, and that is the part worth sitting with. It writes the inference software that decides how many tokens a machine can produce. It publishes the models. It runs hosted endpoints where it serves those models itself. Any one of those is a place a company could put a meter on output, and one of them is a place NVIDIA actually does, which I come back to below.

Everywhere else, on the layer it sells to people who own the machine, the unit is a GPU and a year. That is a choice, not a limitation of being a manufacturer.

#The license is written in GPUs

NVIDIA AI Enterprise is licensed per GPU. The guide is blunt about it: "NVIDIA AI Enterprise software is licensed on a per-GPU basis. A software license is required for every GPU installed on the server or workstation".³¹ There is no token-based tier and no rack-level tier on that page.

The prices themselves run like this.

Sources:.³² The five-year line is not a typo. Five years costs the same as four. The perpetual license includes five years of support. The cloud line is "$1 / hour / GPU + CSP Instance Cost(s)", with no mention of tokens.³²

A five-year AI Enterprise subscription is also included with NVIDIA H100, H200 NVL and A800 PCIe GPUs.³² Buy the silicon, the software rides along.

Every unit on that table is a piece of capacity. A GPU. An hour of a GPU. A year of a GPU. Not one of them is a unit of output.

#What that looks like on one rack

NVIDIA counts a Vera Rubin NVL72 rack, previously known as NVL144, as 72 Vera Rubin GPUs made up of 144 GPU dies.³³ At $4,500 per GPU per year, the annual AI Enterprise bill on that rack is about $324,000.

That number does not move.

It is $324,000 if the rack serves one token this year. It is $324,000 if it serves a trillion. Whether NVIDIA counts the package or the die changes the size of the number. It does not change the shape of it.

That is a capacity charge. It is the grammar a utility uses for a demand charge and a landlord uses for rent.

#The token machine itself is free

Here is the part that took me a while to absorb. The software that produces the tokens is given away.

Dynamo, which describes itself as "The open-source, datacenter-scale inference stack," is Apache 2.0 on GitHub.³⁴ NVIDIA's own March 16, 2026 release calls it free, open source software and says it "boosted the inference performance of NVIDIA Blackwell GPUs by up to 7x, lowering token cost".³⁵ NVIDIA built the thing that makes tokens cheaper and then handed it out.

The models are the same story. The NVIDIA Nemotron Open Model License, last modified December 15, 2025, grants a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable license. It states the works are commercially usable, and says "NVIDIA does not claim ownership to any outputs generated using the Works or Derivative Works".³⁶ NVIDIA's own Nemotron page answers the question directly: "Yes, you can download and run NVIDIA Nemotron models from Hugging Face for free in production".³⁷

The serving layer has one seam in it, and the seam is worth naming. Plain NIM and the NIM Certified Feature Branch are "Available at no charge under the applicable license terms." Only the NIM Certified Production Branch requires an active NVIDIA AI Enterprise subscription, and that page carries no token pricing either.³⁸

So the whole path is free until you want a supported production build, and then the charge is per GPU per year. A company that wanted a cut of your output would not do it this way.

#How NVIDIA merchandises itself to its own investors

The capacity grammar goes all the way up to the earnings call.

On the August 26, 2026 call, CFO Colette Kress put it in the prepared remarks: "Since Hopper, our revenue opportunity has grown from roughly $18 billion per gigawatt to $25 billion with Blackwell, to $40 billion with Vera Rubin".³⁹ Huang gave the same three figures again in the Q&A: "In the case of Hopper, we were at about $18 billion per gigawatt. For Grace Blackwell, we're about $25 billion per gigawatt. And for Vera Rubin, it's about $40 billion per gigawatt".³⁹

Dollars per gigawatt. That is a power unit, which is to say a capacity unit.

The quarter itself was $96.2 billion of revenue, up 106%, with data center revenue of $89.0 billion, up 117%.⁴⁰ The company disclosed no separate software revenue figure. The purchase side runs in the same units. Supply commitments went from "$119 billion last quarter to $279 billion as of July 26, 2026," with total purchase and capacity commitments of $366 billion.⁴¹

The largest AI company on earth buys in capacity and sells in capacity.

#Where NVIDIA does put a meter on tokens

I want to be exact here, because there is a per-token price list at NVIDIA and I will not pretend there is not.

It sits at build.nvidia.com, where NVIDIA hosts the models itself. A third-party tracker, CostBench, verified on August 19, 2026, reports the hosted NIM APIs as a forever-free credit tier plus usage-based pricing from about $0.10 to $1.20 per million tokens.⁴² Another tracker citing the same catalog lists hosted Nemotron prices from $0.04 input and $0.16 output per million tokens on Nano 9B up to $0.50 to $0.60 input and $2.20 to $2.40 output on Nemotron 3 Ultra.⁴³ NVIDIA's own FAQ describes that catalog as "The NVIDIA API catalog is a trial experience" with model-dependent rate limits, and says production use requires an AI Enterprise license.⁴⁴

There is a rule underneath that. NVIDIA meters tokens in exactly one circumstance, which is when NVIDIA is running the model on its own machines for you. The moment the machine is yours, the price reverts to per GPU.

The DGX Cloud Lepton story fits the same rule and is worth reporting because the record changed. The documentation pages that described per-token serverless billing, per-minute marketplace billing and $0.153 per GB per month storage all return HTTP 404 today. The 87 live Lepton doc pages contain no billing page and no "serverless" or "billed" text.⁴⁵ The Lepton product page now points "serverless endpoints" at build.nvidia.com, "including serverless endpoints, prebuilt NVIDIA NIM™ microservices, and GPU-backed compute," with no pricing shown.⁴⁶ Lepton was announced in May 2025 as a marketplace where developers "purchase GPU capacity directly from participating cloud providers," on demand or long term, at prices NVIDIA does not set.⁴⁷ GPU capacity, sold by the unit of capacity.

#The second bite, and how long it lasted

NVIDIA did try to attach a usage meter to somebody else's revenue this year. It is a useful test of the thesis, so here is the whole thing.

On July 1, 2026, NVIDIA announced "a revenue-sharing and credit-support model" in a blog by Colette Kress and Raj Mirpuri. It named Sharon AI for up to 40,000 GB300 GPUs, and Firmus for a Batam campus scaling to 360 MW and up to 170,000 GPUs.⁴⁸ Kress described the structure as a "take-or-pay commitment on a portion of the facility's capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project, and in exchange, we share in a portion of the NeoCloud's revenue earned above that floor".³⁹ Her summary of the economics: "We get paid twice, once on the hardware sale and again, through the share of rental revenue, a highly reoccurring stream layered on top of a one-time equipment purchase".³⁹

The reported terms set "a base hourly rate covering the provider's costs, and Nvidia would receive 50% of any revenue above that threshold." They also required providers to vet customers through an approval process, and pushed utilization across multiple smaller AI firms rather than one large buyer.⁴⁹

Read the unit in that sentence. The share is on a base hourly rate. It is a cut of rent, not a cut of tokens.

Then it stopped. WSJ, as relayed by Reuters on August 27, 2026, reported that "Nvidia has paused some deals that were part of its new financing initiative that offered credit support to AI cloud companies in exchange for a share of revenue".⁵⁰ The same report said some NVIDIA employees had "expressed concerns to current and potential customers that the initiative could draw antitrust scrutiny".⁵⁰ NVIDIA's on-record response was that the model introduced in July "is still in place and continues to evolve due to high demand".⁵¹ July 1 to August 27 is under two months.

Nothing has been collected yet. The Form 10-Q for the quarter ended July 26, 2026 says: "If certain criteria are met, we will participate in revenue share generated by the AI clouds from third-party customers, which may contribute to revenue in the future".⁵² Kress sized it as having "the potential to drive billions in revenue over the medium to long term".³⁹ Against $96.2 billion in a single quarter, that is a rounding line. The commitments behind it total $36 billion, under 10% of NVIDIA's $366 billion of purchase and capacity commitments.⁴¹ One analysis of the model notes that "NVIDIA has not disclosed the revenue share percentage or how usage is measured".⁵³

While I am correcting the record on filings, one more. Press reports repeat WIRED's March 11, 2026 report that NVIDIA will spend $26 billion over five years on open models, with executives Bryan Catanzaro and Kari Briski confirming it in interviews. The filing those reports point to is NVIDIA's Form 10-Q for the quarter ended October 26, 2025, which discloses $26 billion of multi-year cloud service agreement commitments and makes no reference to open models at all.⁵² The FY2026 10-K carries the same commitment line one year later at $27 billion, again with no mention of open models.⁵⁴ The open-model characterization comes from the interviews, not from either document.

#The objection: NVIDIA told telcos to meter tokens

This is the strongest thing anyone can throw at the argument, so here it is in full.

On May 21, 2026, NVIDIA's technical blog told telco AI-factory operators to change their meter: "moving from Compute-as-a-Service to Token-as-a-Service shifts monetization from GPU hours to token-metered AI services such as vertical applications, model APIs, and inference endpoints billed by tokens, requests, or workflows".⁵⁵ The post carries a worked example. An H100 sustaining 30 million billable tokens an hour at $1 per million tokens: "If you charge 1 USD per 1 million tokens, that GPU has 30 USD per hour of token revenue potential. At 60% 'token-active' utilization, that yields about 18 USD of realized token revenue per hour".⁵⁵ That works out to roughly $157,680 per GPU per year, against about $18,400 a year renting the same GPU at $3 an hour at 70% utilization. NVIDIA is telling that operator the token meter earns more than eight times what renting the box earns.

Huang went further at GTC in March 2026 and published a rate card. Free at high throughput and low speed, then $3, then $6, then $45 per million tokens, with a $150 premium tier. He also suggested how to split the plant: "Suppose I take 25 percent of my power, and use it in the free tier; 25 percent in the medium tier; 25 percent in the high tier; and 25 percent in the premium tier".⁵⁶

Both documents are real and I am not going to soften them. NVIDIA wants token meters to exist. It profits when they do, because a token meter justifies buying more racks.

But look at who the audience is. Both are addressed to the owner of the factory, telling that owner how to price what comes out of it. Neither one changes what NVIDIA charges the factory owner. That is still $4,500 per GPU per year, still $1 per GPU-hour in a cloud, still $18 billion, $25 billion and $40 billion per gigawatt.

NVIDIA builds and sells the factory. It is agnostic about the meter downstream, and it will sell you the design for one if you want it. The person who decides whether the enterprise sees a meter is whoever owns the racks.

SAVRN owns the racks and runs NVIDIA's stack on them. We chose the other option on that list, which is no meter at all.

One party in the chain has chosen the meter and says so out loud. Their pricing pages explain why.

---

#3. The labs' meter is a rationing device

I read pricing pages for a living now. Not because I enjoy it. A price list is the one place a company cannot hide what it believes about its own supply.

The spread is not a discount schedule. It is the price of scheduling rights: when you want the tokens, and how you ask for them.

Source: Anthropic pricing documentation

The labs are not hiding it. They have said out loud what the meter is for, and their documentation says what it does when supply gets tight.

#The plan, stated out loud

At BlackRock's Infrastructure Summit in Washington on Wednesday March 11, 2026, Sam Altman described the shape of the whole industry. "Fundamentally our business and I think the business of every other model provider is going to look like selling tokens".⁵⁷ He put the buyer's side of it just as plainly. "We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for".⁵⁷

In the same speech he said the goal is to move away from being capacity constrained, and he reached for the old line about making intelligence too cheap to meter.⁵⁸ Both are in one talk. Sell it on a meter. Get to where you would not need one.

Four months earlier he had written the other half. In a November 2025 post he said OpenAI and others "have to rate limit our products and not offer new features and models because we face such a severe compute constraint".⁵⁹

That is this section's argument in the founder's own words. A meter is what you use while you are short.

#One model, six prices

Take one million input tokens going into Claude Opus 5. Same weights, same request. Here is what it costs, depending on nothing but how you buy it.

The base, cache-write and cache-hit figures come from Anthropic's live pricing page. It lists Opus 5 at $5 base input, $6.25 five-minute cache write, $10 one-hour cache write, $0.50 cache hit and $25 output per million.⁴ The Batch API discounts input and output 50%, putting Opus 5 batch at $2.50 and $12.50, with most batches finishing in under an hour.⁶⁰ Fast mode, a research preview, prices Opus 5 and Opus 4.8 at $10 input and $50 output, twice standard, for up to 2.5x higher output tokens per second.⁶¹

Bottom of the ladder to the top is twenty times. The model never changed. What changed is how badly you wanted the answer now.

Multipliers stack on top. US-only inference on Claude 4.6 and later carries a 1.1x multiplier on all token categories,⁶⁰ and caching and data-residency multipliers stack on fast mode.⁶¹

OpenAI runs the same instrument with different dials. Prompts over 272K input tokens on GPT-5.6 Sol are priced at 2x input and 1.5x output for the full request, and cache writes bill at 1.25x the uncached input rate.⁶² Regional data-residency endpoints on models released on or after March 5, 2026 carry a 10% uplift.⁶³ Flex processing prices tokens at Batch API rates in exchange for slower responses and occasional 429 Resource Unavailable errors.⁶⁴

Look at what the dials measure. Urgency. Patience. Willingness to eat a 429. None of those is a property of the model. They are properties of a queue. A price list built that way is a queue management system with an invoice attached.

#Prices are falling and the caps are tightening at the same time

The obvious objection here is that token prices are collapsing, so nobody is being rationed. The price data supports the first half.

On July 30, 2026 OpenAI cut GPT-5.6 Luna by 80%, from $1.00/$6.00 to $0.20/$1.20, and Terra by 20%, from $2.50/$15.00 to $2.00/$12.00, three days after Moonshot released Kimi K3 free.⁶⁵ On August 21, 2026 it cut GPT-5.6 Sol from $5/$30 to a promotional $4/$20, a 20% cut on input and 33% on output, guaranteed through November 21. That was the third pricing move on the same model family in four weeks, framed as a response to Anthropic and to lower-priced Chinese open models.⁶⁶ Anthropic did its own version, making the $2/$10 Sonnet 5 introductory price standard and canceling the scheduled September 1, 2026 increase to $3/$15.⁴

Prices do not only fall. GPT-5.5, released April 23, 2026, doubled per-token prices against GPT-5.4, from $2.50/$15.00 to $5.00/$30.00, and OpenAI's live page still carries both.⁶⁷ A better model arrived and the meter went up.

Now put the caps beside the cuts.

#Your limit is their supply

On May 6, 2026 Anthropic announced it would use all the compute at SpaceX's Colossus 1, more than 300 megawatts and over 220,000 NVIDIA GPUs. On the strength of it, Anthropic doubled Claude Code's five-hour limits, dropped the peak-hours reduction and raised Opus API rate limits.⁶⁸ New power came online and customer limits moved the same day.

They move the other way too. In late March 2026 Anthropic engineer Thariq Shihipar confirmed session limits were deliberately reduced during weekday peak hours, affecting about 7% of users.⁶⁹ On April 4, 2026 Anthropic blocked Claude subscriptions from powering third-party agents such as OpenClaw, citing "capacity and service issues". The trigger was that a $20 Pro subscription could be used to run agents that would cost hundreds of dollars on a standard API key.⁷⁰ A temporary 50% boost to Claude Code weekly limits began May 13, 2026, a same-day counter to OpenAI offering two free months of Codex to switchers. It was extended four times, to September 13, 2026.⁷¹ On August 29, 2026 Anthropic announced a "permanent 25% increase" to Claude Code weekly limits effective September 14. Because the temporary 50% boost expires September 13, that works out to a 17% reduction from current levels.⁷² Anthropic deleted the original thread and then confirmed the 17% figure.

The billing side moved, then stopped. Anthropic paused its June 15, 2026 Agent SDK billing change on the day it took effect. Agent SDK, claude -p, GitHub Actions and third-party app usage still draw from subscription limits, with the monthly credit unavailable.⁷³

The contract language is what I would read first. Claude Max is $100 a month at 5x or $200 a month at 20x. A session limit resets every five hours, and a weekly limit runs across all models. Anthropic also reserves the right to "limit your usage in other ways, such as weekly and monthly caps or model and feature usage, at our discretion".⁷⁴ On the developer side, the rate-limit page caps Scale-tier spend at $200,000 per month and sets Opus 5 output at 2,000,000 tokens per minute. It then states plainly: "All limits described here represent maximum allowed usage, not guaranteed minimums".⁷

Maximum allowed usage, not guaranteed minimums. That is a supply disclosure, written as a terms-of-service line.

#The capacity products sit above the meter, not below it

Both labs sell something that looks like capacity. Neither prices it under the meter.

OpenAI's Scale Tier sells GPT-5 input units of 25,000 tokens per minute at $75.00 per unit per day, and output units of 2,500 tokens per minute at $60.00 per unit per day. The minimum term is 30 days, with a 99.9% uptime SLA.⁷⁵ Run that input unit flat out and you get 36 million tokens a day, so $75 buys a million tokens for $2.08. GPT-5's list price is $1.25 per million input.⁶³ Reserving it costs about 1.67x list at perfect utilization, which nobody reaches. GPT-5.5 Scale Tier is a combined 50,000 tokens per minute at $750.00 per unit per day, a 2.08x premium to the $5.00 input list at perfect utilization.⁷⁵ Overage is measured in 15-minute windows and, as of July 2026, spills over automatically into Fast mode, billed at double standard rates.⁷⁵

Reserved Tier, for GPT-5.6 onward, is denominated in dollars per minute and burns down at list price. OpenAI's own worked example: "Reserved Tier purchases burn down at list prices; discounts are applied at the time of $/min purchase. For example, if you have a 5% discount, when purchasing $1/min in Reserved Tier ($1440/day list price), you'd actually pay $1368/day".⁷⁶ The unit sold is a dollar. The dollar is spent at the meter.

On May 19, 2026 OpenAI launched Guaranteed Capacity, with one-, two- or three-year commitments at discounts that increase with term, offered until the current allocation sells out.⁷⁷ The reason given was demand for certainty: "Customers are increasingly asking us for certainty on capacity. As models get better, we expect that the world will be capacity-constrained for some time".⁷⁷ A discount on committed spend is still spend measured in tokens.

Anthropic's version is off the shelf. Its service-tiers documentation now reads: "Priority Tier capacity commitments are no longer available for purchase. Organizations with an existing commitment can continue to use Priority Tier through their contract end date".⁵ Anyone who needs guaranteed capacity is directed to sales.⁵ Priority Tier is also not supported on Claude Fable 5.1, Mythos 5.1, Mythos 5, Mythos Preview, Opus 5 or Sonnet 5.⁵

The only flat-capacity product either lab shipped for frontier models is the one you can no longer buy.

#What the meter is actually paying for

The margin data explains why the meter stays.

SemiAnalysis puts the true blended realized price for Claude Opus 4.7 on agentic tasks at $0.99 per million tokens, against a $5/$25 sticker. The drivers are roughly 300:1 input to output ratios and cache hit rates above 90%.⁷⁸ The same analysis reports Anthropic's inference-infrastructure gross margins rising from 38% to over 70% while annualized revenue went from $9 billion to $44 billion.⁷⁸ As summarized elsewhere, SemiAnalysis puts Anthropic's blended gross margin in the mid-60% range with API gross margin above 80%.⁷⁹

Serving is profitable. The company is not, yet. PitchBook estimates Anthropic on track for gross margins of roughly 44%, having spent $0.71 on computing power per dollar of revenue in Q1 2026 and a projected $0.56 in Q2.⁸⁰

Put those two together and you can see where the money goes. Compute runs at a projected 56% of Q2 2026 revenue. Inference-infrastructure gross margin is over 70%, so inference cost is at most about 30% of revenue. Training and other non-serving compute is therefore roughly 26% or more of Anthropic revenue, about 46% or more of its compute spend. At most about 54% of Anthropic's compute serves paid inference.⁸⁰ On the projected 56 cents of compute per revenue dollar, that is at most about 30 cents actually serving a token somebody paid for.

OpenAI's split is wider. Epoch AI, citing The Information, says OpenAI spent around $4.5 billion serving paid users in 2025 out of more than $8 billion spent on inference. That implies around $4 billion went to serving free users, so 40 to 50% of its inference compute served people who paid nothing.⁸¹ For 2026, The Information reports plans for about $32 billion of training compute against about $14 billion of inference compute, a figure OpenAI has not confirmed.⁸² If the 2025 paid share of inference held into 2026, and that is an assumption rather than a reported figure, roughly 17% of OpenAI's total 2026 compute would serve a paying token.⁸²

The serving business itself looks healthy. OpenAI's compute margin is the revenue left after the cost of running models for paying corporate and consumer users. It reached 70% as of October 2025, up from 52% at the end of 2024 and roughly double its January 2024 level, per The Information as reported by Fortune. OpenAI declined to comment.⁸³

So the token I buy is priced well above what it costs to serve, and the surplus goes to the training run and the free tier. That is a reasonable way to fund a research company. It is a bad way to price an input a business runs at industrial volume, because the buyer's bill scales with the seller's research budget and the seller's free users.

#The objection: prices are falling, so where is the rationing

The objection is correct on its face and it is the strongest one against this section. Luna fell 80%. Sol fell 20% and 33%. Sonnet's increase was canceled. If the labs were rationing, list prices would be rising.

BCG looked at the same period and described the mechanism. "When AI compute capacity tightened, providers responded with rate limits, queue delays, or outright throttling" rather than raising prices.⁸⁴ That is how rationing looks when a competitor is giving the product away three days before your price review. You cannot raise the headline number while Kimi K3 is free. You can cut the headline number and take the supply back through the cap, which is exactly the sequence Anthropic ran between May 13 and September 14, 2026.

BCG also puts a stake in the ground on the format itself. "The era of flat-fee pricing models for enterprise users of AI compute has ended." It cites Anthropic's April 2026 move to a $20-per-seat base plus pay-per-token API rates, and OpenAI's move of Codex to token metering. It forecasts dynamic token pricing that resembles wholesale electricity markets.⁸⁴ In the same analysis BCG reports H100 rates of about $8 an hour in early 2024, $1.96 by late 2025 and $2.64 in April 2026, and notes some enterprise GPU clusters running at 5% utilization.⁸⁴

Hold those two BCG findings next to each other. The flat-fee era has ended for the enterprise buying tokens, and the hardware underneath rents for about a third of its early-2024 rate while some of it runs at 5% utilization.

Which leaves the question of who is buying that hardware, and on what terms. The labs sign for it in gigawatts.

---

#4. Same racks

When I want to know what a market believes, I read what it signed. The labs are the most sophisticated buyers of compute alive. Look at what they put their name to in 2025 and 2026, and the pattern shows up fast. They buy the machine, not the meter. And the machine is not exotic. It is the same NVIDIA hardware line an enterprise can order.

#What the labs actually signed

Start with Anthropic on Azure. Microsoft's November 18, 2025 announcement has Anthropic committing $30 billion of Azure compute and up to 1 gigawatt on NVIDIA Grace Blackwell and Vera Rubin systems. The same announcement adds joint engineering to optimize Claude for NVIDIA architectures.⁸⁵

One correction before anyone repeats a number I have seen repeated wrong. The Azure ND GB300 v6 record of about 1.1 million tokens per second on a single GB300 NVL72 rack was announced November 4, 2025. It was set on Meta's Llama 2 70B at FP4 and validated by Signal65, with no Claude or Anthropic involvement.⁸⁶ It is a rack record, not a Claude record. Use it as the first and not the second.

Then West Virginia. The governor's office describes the Nscale agreement as approximately $45 billion over six years for approximately 460 megawatts using NVIDIA Vera Rubin technology.⁸⁷ The headline itself says "Reported," so that is a relayed figure and not a contract value the state holds.

Lambda came next at $35 billion. The report derived from the Wall Street Journal states that the term, GPU count, payment schedule and computing capacity were not disclosed.⁸⁸ A six-year term is reported by trade press, which also notes that the companies have not disclosed how residual capacity risk is allocated after Anthropic's contract expires.⁸⁹

Underneath the silicon sit the buildings. TeraWulf's July 6, 2026 8-K describes a 20-year lease with two five-year renewal options for approximately 401 MW of critical IT load at Hawesville, Kentucky.⁹⁰ The company's release puts it at $19 billion of contracted lease revenue over the initial term, ramping to 401 MW by early 2028.⁹¹ Riot's Q2 2026 release describes a 20-year, 191 MW Rockdale lease worth $9.1 billion in initial contract revenue, about $16.1 billion with both five-year extensions, running through June 2048.⁹² Riot did not name the tenant. CNBC's David Faber confirmed on August 11, 2026 that the tenant is Anthropic.⁹³

OpenAI signs the same shape at a larger size. Microsoft told the SEC in its 10-Q subsequent-events note that OpenAI has contracted to purchase an incremental $250 billion of Azure services. The same note says Microsoft no longer holds a right of first refusal as OpenAI's compute provider.⁹⁴ Ohio is a lease. On August 17, 2026 NVIDIA agreed to provide up to $105 billion of credit and compute for an OpenAI campus in Pike County, Ohio. The credit supports 4.25 GW initially plus a 3.75 GW option, built and managed by SB Energy under a 20-year lease to OpenAI, phased from 2028.⁹⁵ NVIDIA's own 8-K puts it in the language of the filing: residual value guaranties with SB Energy covering approximately 4.25 gigawatts of IT load, with OpenAI as tenant. NVIDIA's aggregate payment obligation is "cumulatively capped at $105 billion for its initial commitment under the Agreements".⁹⁶ Fortune reports the first 800 megawatts are expected in 2028.⁹⁷

Twenty-year terms and six-year terms, priced in gigawatts and megawatts. Not one of those contracts prices a token.

#What the filings say and what analysts say

I want to be precise here, because the word "take-or-pay" gets used loosely and I have used it loosely myself.

The take-or-pay characterization of the compute SPV lease comes from an analyst newsletter, not from any filing I opened. The filings use "contracted revenue" and "expected to be supported by an investment-grade credit".⁹⁸ The TeraWulf 8-K states that "Anthropic's payment obligations under the Justified Data Campus Lease are expected to be supported by an investment-grade credit".⁹⁰ The words take-or-pay and non-cancelable do not appear in it. KBW's Stephen Glagola called the lack of finalized investment-grade credit support "the key remaining overhang" on that lease.⁹⁹ One analyst tally puts Anthropic's announced compute commitments at roughly $450 billion. It frames the coming S-1 disclosure as the question of "how much is genuinely take-or-pay, how much is cancellable capacity, how much is lease liability, and how much remains contingent on future deployment".¹⁰⁰ That split is not public yet. I am not going to pretend it is. What is public is the unit. Every one of these agreements is denominated in capacity and time.

#The price of a finished rack

Divide the contracts and the layers separate cleanly. Finished NVIDIA compute in Anthropic's book clears at roughly $16.3 to $16.7 million per MW-year. Shell-and-power leases clear at about $2.4 million per MW-year, so silicon is roughly 85% of the delivered capacity price.⁸⁷

Lambda's capacity was not disclosed, so its per-MW-year figure is an implied estimate rather than a derived one.

Take the Nscale contract all the way down to the rack. In my August 28, 2026 article I priced it at 2,038 racks and 146,740 accelerators: "Divide it down and it is $7.5 billion a year, $1,358.70 per kilowatt-month, and $5.83 per accelerator-hour".¹⁰¹ That works out to $419.76 per rack-hour. A rack-hour does not move when the workload gets chattier.

#The objection worth answering

Here is the strongest version of the pushback. Roughly two-thirds of Anthropic's committed dollars are on Google TPUs and Amazon Trainium, chips nobody else can buy. Pointing at the NVIDIA contracts cherry-picks the part that suits me.

The non-NVIDIA commitments are real and they are large. Anthropic's April 20, 2026 release commits over $100 billion to AWS over ten years for up to 5 GW on Trainium2 through Trainium4. The release adds an option to purchase future generations of Amazon's custom silicon.¹⁰² Broadcom's April 6, 2026 8-K says Anthropic "will access through Broadcom approximately 3.5 gigawatts as part of the multiple gigawatts of next generation TPU-based AI compute capacity committed by Anthropic".¹⁰³ The same filing says consumption of that capacity is dependent on Anthropic's continued commercial success. Epoch AI describes a compute SPV raising roughly $35 billion of debt, precisely $34.5 billion, to buy more than 1 GW of Google TPU systems and lease them to Anthropic for five years.¹⁰⁴ No enterprise can order any of that.

Two things follow. The two-thirds figure is a tally of announced commitments: roughly $200 billion of TPUs including the SPV and more than $100 billion of Trainium, against about $155 billion on NVIDIA.¹⁰² The audited split is not disclosed. The NVIDIA share is the part an enterprise can replicate. It is roughly $80 billion across Nscale and Lambda, and it is the part growing fastest.

That is the useful piece of the whole comparison. The chips a lab buys that nobody else can buy are a hedge. The chips a lab buys that anyone can buy are a market, and the enterprise stands in the same line for them.

#Where the meter starts

Follow the chain in order. NVIDIA sells capacity, guaranteed against a 20-year lease. Clouds and neoclouds sell capacity, priced per megawatt-year. Labs buy capacity, in gigawatts, on six- and twenty-year terms. Then the enterprise arrives, and the unit changes.

What waits at the end of that chain is the price list from the cold open. Tokens sold by the million, a ceiling the documentation describes as a maximum rather than a minimum, and the one flat-capacity product marked no longer for sale.⁵

So the buyer in every one of those capacity contracts is a lab or a cloud. The enterprise is downstream of a meter.

---

#5. Who already buys capacity as an end user

The wholesale layer already runs flat. I take that as settled. The question I care about is narrower, and it is the one a customer asks me on a Tuesday afternoon. Does anybody who actually consumes the output buy the machine instead of the meter?

Output tokens per month 500 billion Price basis on the meter Claude Fable 5.1 list, $50 per million Claude Opus 5 list, $25 per million GPT-5.6 Sol list, $20 per million GPT-5.6 Luna list, $1.20 per million Anthropic realized agentic price, $0.99 per million GPT-5.6 Luna batch, $0.60 per million

On the meter$12.5Mper month

Racks required3at 28.4T output tokens per rack-year

Cost to hold them$0.56Mper month, at $2.25M per rack-year

Ratio22xmeter against machine

Rack cost from the InferenceX operator-ownership TCO of $3.57 per GPU-hour. Output per rack-year is this piece's mid case. List prices from the OpenAI and Anthropic pricing pages. This compares a metered bill against the cost of holding the machine. It excludes your staff, your building and your utilization risk, which are the whole subject of the sections that follow.

The answer is yes. The answer is also small. End-user enterprises are the fastest-growing buyers of flat compute capacity, and they are still the smallest group of buyers. I want both halves of that sentence on the table before anyone accuses me of showing one and hiding the other.

#What the equipment makers are booking

Dell booked a record $60.9 billion of AI-server orders in its second quarter of fiscal 2027 and exited with a $95 billion backlog.¹⁰⁵ AI-server revenue in the quarter was $16.4 billion, up 100 percent year over year, and Dell raised full-year AI-server guidance to $74.0 billion, up 200 percent. That is a company selling racks to people who intend to own them.

Dell's own slides put the AI-server customer base at more than 6,500 customers spanning neocloud, national-AI and enterprise buyers, and say demand is broadening beyond the initial hyperscale deployments.¹⁰⁶ No order split by segment is disclosed. So the customer count broadens while the dollar mix stays private.

HPE tells a cleaner version of the same story at a tenth of the size. It booked $1.8 billion of AI-systems orders in fiscal Q2 2026, took its cumulative total to $16.4 billion, and entered Q3 with a $5.9 billion backlog it describes as primarily enterprise and national-AI orders.¹⁰⁷ Antonio Neri gave the reason without being asked twice. "Enterprise has won the flexibility of choosing multiple AI models with the governance and control of on-premises," he said, adding that inferencing will be done where the customer feels confident about governance and data privacy.¹⁰⁷ Marie Myers added that enterprise and national-AI business is more profitable than service providers or model builders.

The same transcript carries the correction. Service-provider AI orders in that quarter exceeded the combined total of the prior four quarters.¹⁰⁷ HPE calls that the inherently lumpy nature of its large-scale AI deals. That is what a market looks like on its way somewhere, not one that has arrived.

#NVIDIA's second half

NVIDIA now reports its data center business in two pieces, and the smaller piece is growing faster. Second-quarter fiscal 2027 data center revenue was $89.0 billion. Hyperscale was $48.7 billion, up 102 percent. AI Clouds, Industrial and Enterprise was $40.3 billion, up 138 percent year over year and 25 percent sequentially.¹⁰⁸ My arithmetic on those two lines: the non-hyperscale half is 45 percent of data center and is compounding faster than the half everyone writes about.

Colette Kress said the non-hyperscaler segment, covering national-AI buyers, regional neoclouds, enterprise edge and air-gapped data centers, will represent roughly half of the data center business.³⁹ National-AI business running through regional neoclouds more than tripled year over year.

Jensen Huang has been framing this since May 2025, when the RTX PRO Server line and the Enterprise AI Factory validated design launched with Cadence, Foxconn and Lilly named as first builders. His line was that "every company will build or rent AI factories to run their businesses and power the intelligence of their products".¹⁰⁹ By August 2025 the named list had grown to Disney, Foxconn, Hitachi, Hyundai Motor Group, Lilly, SAP and TSMC deploying those servers on premises, with no server or GPU counts disclosed for any of them.¹¹⁰

#The people who own the machine

A handful of end users have put their names on owned capacity, and the counts are public.

Lilly's LillyPod is 1,016 Blackwell Ultra GPUs, announced October 28, 2025, and described by NVIDIA as the most powerful AI factory "wholly owned and operated by a pharmaceutical company".¹¹¹ A secondary trade report says it was inaugurated in Indianapolis after a four-month assembly. Texas A&M bought a $45 million SuperPOD with 760 Hopper GPUs through World Wide Technology, tripling its supercomputing capacity.¹¹² Denmark's Gefion, 1,528 H100 GPUs, is owned and operated by the Danish Centre for AI Innovation. The money was about DKK 600 million from the Novo Nordisk Foundation and DKK 100 million from EIFO, and it serves universities, startups and public agencies.¹¹³ Isambard-AI runs 5,448 GH200 superchips with access for academia, industry and start-ups.¹¹⁴

The national builds are bigger and they cut the other way. Saudi Arabia's HUMAIN plans up to 600,000 NVIDIA GPUs over three years. The first phase is an 18,000-GB300 supercomputer in a facility over 500MW, and its first customer is xAI. AWS will deploy up to 150,000 GPUs in a HUMAIN AI Zone in Riyadh.¹¹⁵ The national buyer's anchor tenants are a lab and a hyperscaler. Stargate UAE is the same shape. It is a 200MW first phase of a 1GW cluster inside the 5GW UAE-US AI Campus, reported on track for 2026 delivery as of October 16, 2025. G42's Khazna is developing it with OpenAI, Oracle, NVIDIA, Cisco and SoftBank.¹¹⁶ A government wrote the check. A lab fills the racks.

#The people who reserve the machine

Mistral announced European Compute Units on August 11, 2026. ASML, Amadeus, Capgemini, Caisse des Depots and CMA CGM made multi-year commitments that convert into access to Mistral-built capacity, about 200MW by end-2027 and up to 1GW by 2030.¹¹⁷ CTO Lacroix described the terms: "The goal is to have customers commit for around five years, or at least a long time." No euro value was disclosed. The 44MW Paris-area site is the anchor facility.

Akamai signed a four-year, $200 million agreement announced March 5, 2026, under which an unnamed major US tech company gets a dedicated multi-thousand RTX PRO 6000 Blackwell cluster.¹¹⁸ That is roughly $50 million a year, flat, for dedicated capacity from a host that is not a hyperscaler.

AT&T processes about 45 billion tokens a day.¹¹⁹ Per the Wall Street Journal on August 12, 2026, citing Chief Data and AI Officer Andy Markus, open models power about 25 percent of its AI usage. AT&T expects 70 to 80 percent over time. The switch has produced 80 to 90 percent savings in certain applications. The setup is hybrid, not owned. OTel 2.0 was post-trained on more than 1 trillion tokens of managed Azure compute, on AMD Instinct GPUs in Microsoft Foundry. Some employee requests still route to Anthropic and OpenAI APIs, and open models handle about 40 percent of employee requests as of August 2026.¹²⁰ Those two open-model shares are measured on different bases and both are current. The 25 percent covers total AI usage as reported to the Journal. The 40 percent counts employee requests only.

#The reserved-but-not-owned middle

Azure provisioned throughput is a flat product. You buy a unit, you pay per unit per hour, and the deployment "holds a fixed amount of processing capacity exclusively for your deployment's use, whether or not requests are being made".¹²¹ Amazon Bedrock Provisioned Throughput works the same way, with model units billed hourly on no-commitment, one-month or six-month terms.¹²²

Microsoft's own documentation names the limit of that arrangement. "Reservations don't guarantee capacity," it says, telling you to create the deployment first and then buy the reservation to lock in the rate.¹²¹ A reservation on somebody else's cloud is a discount on the meter. It is not the machine.

The volume is real regardless. Microsoft's commercial remaining performance obligation reached $678 billion at fiscal Q4 2026, up 84 percent, and Amy Hood said all sequential growth came from customers outside the frontier model companies. Excluding OpenAI, RPO grew 25 percent.¹²³ Nearly 90 percent of Microsoft Cloud revenue comes from customers outside frontier model companies, and the weighted-average duration including OpenAI is 2.3 years.¹²³

#The part that cuts against me

Now the other side, in full.

Go back to Amazon's $496 billion of unrecognized customer commitments and look at who signed them. The same filing names two lab commitments: OpenAI's existing $38.0 billion commitment expanded by $100.0 billion over 8 years, and Anthropic's existing multi-year commitment expanded by more than $100.0 billion over 10 years. If those named amounts sit fully inside the backlog, at least about $238 billion of it, roughly 48 percent, is two AI labs. That 48 percent is an estimate, not a disclosed floor, because Amazon does not say how much of the OpenAI figure remains unrecognized. The true lab share is probably higher, because Anthropic's pre-existing commitment amount is not disclosed. Half of the largest flat-capacity book in the world is two customers, and neither one is an enterprise.

NVIDIA's disclosed on-prem enterprise verticals are small. Kress sized automotive at $8 billion trailing twelve months and financial services, manufacturing and healthcare combined at $7 billion.³⁹ My arithmetic: about $15 billion trailing twelve months against a $356 billion annualized data-center run rate is roughly 4 percent.

Andy Jassy put it plainly. "Remember, enterprises are still very early in using inference at scale in their current production applications," he said. He described adoption as a barbell, with the labs and a few runaway apps consuming enormous compute at one end and enterprises taking cost avoidance and productivity at the other.¹²⁴ Gartner said the same thing from the outside. It forecasts $2.59 trillion of AI spending in 2026, and says spending so far has been driven by technology companies and hyperscalers: "Enterprises have yet to really flex their spending potential".¹²⁵

Then the operating record. Cast AI's 2026 report, drawn from tens of thousands of enterprise Kubernetes clusters on AWS, Azure and Google Cloud, found average GPU utilization of 5 percent.¹²⁶ The exclusion the report states is clusters from AI labs, the data was collected from organizations before they enabled Cast AI's automation, and Cast AI sells optimization software. The same Flexera survey that recorded 29 percent wasted cloud spend also found fewer than half of organizations use any single commitment-discount type with any provider.³⁰ Enterprises are not yet committing flat even where committing flat is easy.

Owning the rack does not by itself end the meter either. Meta owns its capacity, consumed 73.7 trillion tokens internally in a 30-day period, and is imposing budgets, usage limits and an AI Gateway dashboard because internal use is on track to cost billions in 2026.¹²⁷

#The objection

Someone will say the flat buyers that move the numbers are labs, not enterprises. That is correct today. I am not going to pretend otherwise when the Amazon filing says it in one line and NVIDIA's own call says its enterprise on-prem verticals are about 4 percent of the run rate.

The growth rates say which way the line runs. The non-hyperscale half of NVIDIA's data center business grew 138 percent against hyperscale's 102 percent. Dell's AI-server customer count crossed 6,500 across three segments. HPE's backlog is primarily enterprise and national-AI. Microsoft's RPO grew 25 percent excluding OpenAI. Uber and ServiceNow reportedly exhausted their entire annual budgets for Anthropic's tools within the first few months of 2026.¹²⁷ A Cloudian-commissioned survey of 203 enterprise IT decision-makers in February 2026 found 79 percent had already moved at least one AI workload out of public cloud and 73 percent planned more within two years.¹²⁸ IDC predicts 75 percent of enterprise AI workloads will run on hybrid infrastructure by 2028.¹²⁹

SAVRN builds and operates AI factories for the buyer in that second column, and bills for the machine rather than for the tokens the machine makes.

All of which puts the weight back on one number. If an enterprise is going to buy the machine, somebody has to say what the machine makes and what it costs to hold for a year.

---

#6. What a rack makes and what it costs

A Vera Rubin rack makes frontier-class tokens for ten to twenty cents a million when you run it for throughput. That number is not a property of the silicon. It is a property of the service level you sold. Move the service level and the same rack produces a price fifty times higher.

Tokens per second, per user

50 tok/s per user

Cost per million output tokens$0.26

Batch-shaped serving. This is the profile an agentic workload actually runs at: long inputs, heavy caching, no human waiting on each token.

Measured points: SemiAnalysis InferenceX and NVIDIA's inference page. The 50 and 100 tok/s points are DeepSeek R1 on GB300 NVL72; 116 is NVIDIA's own published figure; 350 is Vera Rubin, a service level Blackwell cannot reach at any price.

This is the part of the argument I have to show all the way down, including the parts that cut against me.

#The rack we are talking about

Two names describe one machine. NVIDIA's own blog carries an editor's note recording a branding change from Vera Rubin NVL144 to Vera Rubin NVL72, confirming the two names denote the same rack.¹³⁰ If you see NVL144 in an older deck, it is this rack.

It is shipping. NVIDIA's Q2 FY27 investor presentation states that NVIDIA commenced production shipments of Vera Rubin in early August 2026.¹³¹

Power is the first place people get it wrong. SemiAnalysis's Vera Rubin co-design analysis puts VR NVL72 rack power at 180-220 kW, "up to 220kW" for the 2,300 W Rubin TDP SKU, with a 1,800 W Max-Q variant.¹³² NVIDIA does not publish a rack power figure for Vera Rubin, and the widely circulated 190 kW Max-Q and 230 kW Max-P pair traces to a single originator, Ming-Chi Kuo.¹³³

The 600 kW number that circulates in board decks is a different product in a different year. The Register reports the Rubin Ultra NVL576 rack at 600 kW with 15 exaFLOPS FP4 inference, due late 2027, and gives no power figure for the 2026 Vera Rubin NVL144.¹³⁴ If you are planning a 2026 build, plan around roughly 200 kW a rack, not 600.

#What nobody has published yet

As of today, nobody outside NVIDIA and its direct customers has put a Rubin throughput number on the record. CoreWeave's blog reports Vera Rubin NVL72 at 10x tokens-per-second per megawatt versus GB200 NVL72 on DeepSeek R1 in NVFP4 with TensorRT-LLM and Dynamo, and discloses no absolute tok/s, MW, or tok/s per user figures.¹³⁵ SemiAnalysis then took that 10x apart. Its InferenceX analysis states CoreWeave's 10x is against a 2025 GB200 baseline with year-old software. Renormalized against July 2026 GB300, the Rubin advantage is near 2x through 100 tok/s per user, about 4x around 200, and 5.4x at 300.¹³⁶

There is still no independent agentic benchmark of the rack. InferenceX's blog index shows an Aug 24-25, 2026 AgentX v3 wave covering GB300 NVL72, B300, MI355X, GB200 and OpenAI Jalapeño, but no Vera Rubin NVL72 AgentX post.¹³⁷ And no cloud had published Vera Rubin NVL72 rental pricing as of June 2026. Spheron expects a 30-50% premium over Blackwell NVL72 at initial availability, and warns the translation to token cost is not 1:1.¹³⁸

So every Rubin token number in this section, mine included, is built on measured Blackwell Ultra plus a stated generational assumption. I would rather say that out loud than hide it in a footnote.

#What a rack makes

The measured anchors are MLPerf, verified by MLCommons. On a 72-GPU GB300 NVL72, MLPerf Inference v6.0 results show DeepSeek-R1 server at 575,580 tok/s and offline at 673,936 tok/s, and gpt-oss-120B at 1,096,770 server and 1,046,150 offline.¹³⁹ Nebius recorded all four as first place.¹³⁹

Apply a 2x generational assumption to those anchors and a Rubin rack lands between 1.15M and 2.2M tok/s. That is an assumption, not a measurement. NVIDIA's own claim is up to 10x tokens per megawatt on Kimi-K2, projected, at 32K in and 8K out.¹³⁹

Run that band across a year and you get this.

Those three lines reproduce exactly over 31.536M seconds a year.¹³⁹ Sixteen to fifty trillion tokens a year, from one rack. Not quadrillions. When somebody tells you a single rack can serve a country, ask them for the seconds in a year.

#The number that guts the pretty version

The 16 to 50 trillion figure is a throughput-mode number. It assumes batch serving and low per-user speed. Run the same rack at agentic interactivity and the output collapses.

InferenceX measures Rubin at 96,446 tok/s per megawatt at 300 tok/s per user and 70,703 tok/s per megawatt at 350 tok/s per user on DeepSeek R1 671B with 8k in and 1k out. At roughly 0.2 MW per rack, that is on the order of 14,000 tok/s per rack. Call it 0.2 to 0.5 trillion tokens a year at agentic interactivity, two orders of magnitude below the batch-mode table.¹⁴⁰

Context length does the same thing again. Epoch's calibrated table gives GB200 NVL72 throughput on Kimi K2.6 of 400k tokens/s at 8k:1k context, 110k at 25k:1k, and 9k at 128k:1k, a 44x drop from short to long context.¹⁴¹

Two dials, service level and context, each worth one to two orders of magnitude. Any capacity contract that does not name both is not a contract. It is a hope.

#What a rack costs

Morgan Stanley estimates a single Vera Rubin NVL72 rack at $7,803,148, with $2,001,600 of that in memory, a 435% memory-cost increase over GB300, representing what cloud providers pay rather than NVIDIA's manufacturing cost.¹⁴²

Operating it is the other half. SemiAnalysis InferenceX puts operator-ownership TCO at $3.57 per GPU-hour for Vera Rubin NVL72, $1.84 for GB200 and $2.36 for GB300, drawn from its AI TCO model covering capex, opex and other expenses. It does not itemize the components publicly. By arithmetic that is $2.25M, $1.16M and $1.49M per 72-GPU rack-year.¹³⁶ Rubin costs more per GPU-hour than Blackwell. It also makes far more tokens per hour, which is the whole point.

#The reconciled table

Here is every basis I can source, on one page, with the tokens each one assumes. I built this because the same rack shows up in document after document at a different price each time, and the difference is always the denominator.

That reconciliation is ours (SAVRN canonical factory economics, internal model).

The contract line is worth walking through, because it is the only one paid in real money by a named buyer. The Anthropic-Nscale agreement divides down to $419.76 per rack-hour, as set out earlier. Divide that by the MLPerf throughputs and you get $0.2026, $0.1730 and $0.1063 per million tokens, on 18.15T, 21.25T and 34.59T tokens per rack-year, at $3.677M of compute per rack-year.¹⁰¹

Now the correction, and it matters. That $3.68M is a contract price paid to a provider. It necessarily includes Nscale's margin and its facility, power, network and financing costs, although the article itself does not itemize them. It sits 22% above the high case of $3.02M and 63% above the $2.25M InferenceX TCO. The per-token figures derived from it are a lab's wholesale purchase cost, not a rack's production cost.¹⁰¹

I will take that correction and keep the number anyway. Twenty cents is what one of the most sophisticated buyers of AI compute in the world actually pays, with somebody else's margin already inside it.

NVIDIA's own marketing lands in the same neighborhood on the previous generation. Its inference page states GB300 NVL72 delivers $0.123 per million tokens at 116 tokens per second per user on DeepSeek-R1 at 32K in and 8K out, running Dynamo and TensorRT-LLM. That is about 6,000 tok/s per GPU.¹⁴³ InferenceX's own measurement of DeepSeek R1 on GB300 NVL72 is $0.26 per million at 50 tok/s per user and $0.49 per million at 100 tok/s per user. That is 2,438 tok/s per GPU on a $2.31 per GPU-hour hyperscaler tier.¹⁴⁴ Push the Nscale rate through that measured per-GPU throughput instead of the MLPerf peak and a DeepSeek R1 output token costs about $0.66 per million, roughly 3.3x the $0.2026 figure.¹⁴⁴

Three cents to sixty-six cents, depending entirely on which throughput you put in the denominator. That spread is the real finding of this section.

#The dial

Here is the same rack, priced by service level rather than by hardware.

InferenceX reports that by 350 tok/s per user, neither GB200 nor GB300 can serve the workloads at all. That leaves only Rubin with a curve, delivering 96,446 tok/s per megawatt at 300 tok/s per user and 70,703 at 350.¹³⁶ At that top operating point Rubin costs $4.18 per million output tokens on DeepSeek R1 671B in NVFP4 at 8k in and 1k out.¹³⁶

$4.18 against the $0.079 central case is a 52.9x penalty for interactivity.¹³⁶ Same rack. Same model. Same week. The only thing that changed is how fast the tokens come back to one user.

That is why I keep saying the price is a property of the service level. If you buy capacity and never state the envelope, you have bought a number between three cents and four dollars and eighteen cents.

#Our own rack

I will state SAVRN's numbers as ours, on the same basis, so you can compare them line by line.

Our canonical rack costs $3.66M a year: depreciation of $1.90M ($9.5M over five years), power of $0.361M (412.5 kW at 8,760 hours at $0.10 per kWh), facility of $0.445M, and server maintenance of $0.95M. That is $0.049 per million tokens at 100% utilization and $0.076 at 65% (SAVRN canonical factory economics, internal model). Depreciation is a little over half of it. Everything else is the building and the hands.

Ours runs about $1.56M a year above the $2.10M mid case, for three reasons I can name. Rack capex is $9.5M all-in versus $7.8M ODM, which adds $0.34M a year over five years. Power is priced on 412.5 kW of facility load at $0.10 per kWh rather than 230 kW at 6.4 cents, which is 2.8x and adds $0.23M. And we carry a $0.95M server-maintenance line that the mid case does not carry separately. Facility cost is nearly equal, $0.445M against $0.414M (SAVRN canonical factory economics, internal model).

Fully loaded is a different number and I would rather publish it than be asked for it. With every operating cost in, our rack carries $14.27M of opex at full utilization, of which G&A at 35.5% of revenue is $11.89M. Add $1.9M of depreciation and it is $16.17M per rack-year, or $0.217 per million tokens. At our 65% planning utilization the revenue-linked lines shrink with revenue, giving about $11.5M per rack-year, or $0.24 per million tokens (SAVRN canonical factory economics, internal model).

That 65% figure can be read two ways and both appear in our own model, so here are both. Let the revenue-linked lines shrink with revenue and the answer is 24 cents a million. Carry the full $16.17M loaded cost against 65% of the output and the answer is 33 cents a million. The first assumes G&A tracks the smaller book. The second assumes it does not (SAVRN canonical factory economics, internal model).

Twenty-four to thirty-three cents a million with the sales team, the overhead and the empty capacity all inside it.

#What you are comparing it against

Claude Opus 5 output lists at $25.00 per million tokens (SAVRN canonical factory economics, internal model). Our fully loaded twenty-four cents is roughly 100x below that (SAVRN canonical factory economics, internal model). Our $0.45 per million blended price is 55x below the Opus 5 output list and 5.7x above the $0.079 central production cost. A flat capacity contract priced at our rate is therefore both a deep discount to the meter and a healthy markup on the rack.¹⁰¹

The list price is the easy comparison and the weakest one, because almost nobody pays it. So take the cheap end of the metered market instead. OpenAI's live pricing page lists Luna at $0.20 in and $1.20 out standard, and $0.10 in and $0.60 out batch, per million tokens. Terra is $2 and $12, and Sol is $4 and $20, with Sol's price labeled promotional through at least Nov 21, 2026.⁶³ Against $1.20 out, our fully loaded twenty-four cents is a factor of five. Against $0.60 batch it is a factor of two and a half.

Say which one you are arguing. At list, the gap between a metered frontier token and a rack-produced token is around 100x. Against the cheapest adequate metered token, it is closer to 5x, and against that token in batch mode it is 2.5x. The 100x number wins arguments. The 5x number is the one an enterprise procurement team will actually check.

One more anchor for what the rack does at these prices. Morgan Stanley's AI-factory profitability model depreciates hardware over 4 years and construction over 10, and assumes 70% utilization at an average token price of $0.20 per million. It still yields a 77.6% operating margin for an NVIDIA GB200 NVL72 factory, against AMD MI300X at negative 28.2%.¹⁴⁵ Twenty cents a million is not a loss leader. It is a 77.6% margin business at somebody else's model.

#The objection

Here it is in full, with nothing taken off it. Nobody sells batch-mode frontier serving at low per-user speeds. The measured throughput peaks that produce three-cent and eight-cent tokens sit at interactivity nobody would ship to a coding agent. InferenceX's GB200 NVL72 Kimi K2.5 run peaked at 12,587 tok/s per GPU at 23.2 tok/s per user.¹⁴⁶ Its Kimi K2.6 comparison shows GB300 NVL72 peaking at 9,986.1 tok/s per chip at about 64 tok/s per user.¹⁴⁷ Agents want 300. At 350, the price is $4.18. The cheap number is real and it is not the number your agent runs at.

My answer has two parts, and the first is a concession. If you buy capacity and run every workload at maximum interactivity, you will not see three cents. You will see something between one and four dollars, and the contract should say so.

The second part is that agentic work is not one service level. It is a mix, and the measured agentic runs come in far below $4.18. InferenceX's August 2026 overview shows GB300 NVL72 serving DeepSeek V4 Pro 1.6T on agentic traces at $0.011 per million tokens on August 14, 2026, and Kimi K3 2.8T at $0.080 on August 18, 2026. Both use SemiAnalysis July 2026 TCO pricing, and come in 52% and 24% cheaper than B200.¹⁴⁸ Those are agentic traces on the previous generation, priced in cents. NVIDIA's own claim for the new generation is up to 30x higher AI-factory throughput per megawatt than GB300 NVL72, at 160 tok/s per user on DeepSeek V4-Pro under the SemiAnalysis AgentX workload. The results are pending SemiAnalysis review.¹⁴⁹ Pending review means unverified, and I am counting it as unverified.

So the rule I would write into a capacity contract is simple. State the interactivity envelope, state the context profile, and price the machine, not the meter. The buyer who does that gets the whole curve. The buyer who signs a per-token rate gets whichever point on the curve his vendor picked.

None of that arithmetic matters if the cheap tokens are not good enough to run the business on. So the next question is what the frontier premium actually buys.

<!-- FIGURE: cost per million output tokens versus tokens per second per user, one rack, showing $0.03 to $4.18 across the curve -->

---

#7. The frontier premium is shrinking faster than the marketing

My instinct with a claim like this one is to go look at the boards. The claim I want to test is that the best closed model is worth what it costs on the work an enterprise actually runs. The boards say the frontier still leads. They also say the lead is small on most work and large on a narrow band of it. Both are true at once, and the price cuts of the last two months tell you which one the labs are worried about.

#Two readings of the gap

On the Artificial Analysis Intelligence Index v4.1.1, the top score is Claude Fable 5.1 at max effort with 66, and Claude Opus 5 at max scores 63.¹⁵⁰ The highest open-weight models on the same index are Kimi K3 at max and GLM-5.3 at max, tied at 60, with Qwen3.8 2.4T A95B at 58.¹⁵¹ That is a six-point gap between the best proprietary model and the best open one. The top five slots on the live index are all Anthropic models.¹⁵⁰

A year earlier the same gap was about 13 points. Artificial Analysis measured DeepSeek V3 0324 at 22 against Claude 3.7 Sonnet Reasoning at 35.¹⁵² Thirteen points to six in twelve months.

Epoch AI runs the same question on a different instrument and gets a similar shape. Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months on the Epoch Capabilities Index, an average gap of about 8 ECI points.¹⁵³ Epoch's earlier estimate, from October 2025, was a three-month lag with a 90% confidence interval of 1.1 to 5.3 months.¹⁵⁴ So the lag widened slightly between those two readings.

The Stanford AI Index makes the point in both directions. In January 2024 the leading closed model led the top open model by 8.04% on Chatbot Arena, and by February 2025 that gap had narrowed to 1.70%.¹⁵⁵ The 2026 edition reports the gap re-widened to 3.3% as of March 2026, up from 0.5% in August 2024, with six of the top ten Arena models closed.¹⁵⁶ The trend line is not monotone. It moves with each release.

On the Arena text leaderboard as of September 2, 2026, all top ten entries are proprietary. The highest open-weight entry is kimi-k3-max at #11 with 1489, which is 19 points and 1.3% below #1.¹⁵⁷

#The same benchmark, two answers

Coding is where this argument usually gets settled. On the Vals AI SWE-bench Verified run updated September 1, 2026, Claude Opus 5 leads at 97.00%, the open-weight DeepSeek V4 Pro 0813 scores 96.40%, and Kimi K3 scores 93.40%.¹⁵⁸ That is a gap of 0.6 points.

Run a different model on a different harness and you get a different picture. NVIDIA's Nemotron 3 Super scores 60.47 on SWE-bench under the OpenHands harness, well below the 80.8% Anthropic reports for Claude Opus 4.6 on SWE-bench Verified.¹⁵⁹ Both numbers are real. They are different models and different harnesses, and anyone quoting one of them at you without the other is selling something.

Harder boards widen the spread again. On Scale's SWE-bench Pro public leaderboard, the best open-weight entry is qwen3-coder-480b-a35b at 38.70. It runs against Muse Spark 1.1 at 61.50 and gpt-5.4 at 59.10, a 22.8-point gap. That board does not yet list DeepSeek V4 Pro, Kimi K3 or GLM-5.3.¹⁶⁰

Terminal work sits in between. Claude Fable 5.1 at max leads Terminal-Bench v2.1 at 91.4%.¹⁶¹ The open-weight field clusters about three points back: Kimi K3 at 88.3%, GLM-5.3 at 88.2%, DeepSeek V4 Pro 0813 at 87.9% vendor-reported, Qwen3.8 Max at 86.6%. The board combines vendor and leaderboard harnesses.¹⁶²

#Where the gap concentrates

The premium is real, and it lives in the hard tail. In April 2026, open models scored 34-36% on Humanity's Last Exam against 44% for GPT-5.5 and 45% for Gemini 3.1 Pro Preview. On CritPt research physics they scored 4-12% against 27%. On TerminalBench Hard they scored 43-46% against 61%.¹⁵² The same survey found DeepSeek V4 Pro scoring -10 on the hallucination-sensitive Omniscience metric against +20 to +33 for proprietary leaders.¹⁵²

Long-horizon agentic work shows the sharpest break. In the AgentFloor study, across a 30-task tool-use ladder, the strongest open-weight model matched GPT-5 in aggregate at roughly 15x lower cost per passed task on a Mac Studio. On long-horizon planning under persistent constraints, GPT-5 passed 10% and the best open model passed 0%.¹⁶³ METR's Time Horizon v1.1 results file contains no open-weight entries at all.¹⁶⁴

That is the shape of it. Near parity on single-turn quality. A real and wide gap on the long agentic runs that decide whether an automated task finishes.

#What the labs did about it

Watch the dates on the price moves already on the table: Luna down 80% three days after Kimi K3 went free, Sol down 20% on input and 33% on output the day after the NVIDIA-Poolside deal was reported, and Anthropic's scheduled September increase canceled.⁶⁵ Each cut landed within days of an open-weight release. TechRepublic reads the pattern the same way. It reports the Luna cut, and Anthropic marketing Opus 5 at half the price of Mythos as open-weight models gain ground, with DoorDash, Airbnb and Coinbase shifting some workloads to cheaper models.¹⁶⁵

Here is the list price picture, blended at three input tokens to one output token.

Anthropic list prices are from its pricing page,⁴ OpenAI's from its developer pricing page,⁶³ DeepSeek's from its API docs where off-peak rates are half the peak rates,¹⁶⁶ and the hosted open-weight rates from Together.¹⁶⁷ Note what the table does to the easy story. The best open model is not the cheap one. Kimi K3's official price is $3 input and $15 output, a blended $6, only 1.7x below Opus 5 and above Sonnet 5, and Artificial Analysis calls it particularly expensive among open models.¹⁶⁸

#What the routing data shows

LangChain ran the test I would have run. It ran 145 multi-step agentic tasks averaging 6.3 calls each. Routing between Nemotron 3.5 Lightning and Claude Opus 4.8 with NeMo Switchyard sent 7% of calls to the frontier model. That cut cost 74%, from $11.45 to $3.00 per run, and dropped accuracy from 86.0% to 80.0%. Nemotron alone scored 77.7%. The 30B model handled 93% of calls for 10.4% of spend, and Opus handled 7% of calls for 68.4% of it.¹⁶⁹ NVIDIA reports Cognition's staged routing on FrontierCode reaching 50.6% accuracy, within 2.8 percentage points of Opus 5, at about 28% lower mean cost.¹⁷⁰

AT&T is the production version, and its routing numbers sit behind the token counts already given. It moves about 40% of employee AI queries to open models through LiteLLM routers and reports a 56% cut in AI coding and advanced task cost with a 2% decline in quality.¹⁷¹ Its stated target appears two ways in the record: 60 to 70% in the PYMNTS account of the routing program, and 70 to 80% in the Journal's account of open-model share overall. It built its own customized open model, OTel, and says its AI initiatives delivered a fivefold return on investment this year.¹⁷² AT&T's Mark Austin puts the lag at six to ten months and says the gap is narrowing.¹⁷¹

Salesforce did the same surgery on its own product. Eighteen months ago Agentforce ran on a single rented model with a token bill growing linearly with traffic. Salesforce pulled jobs out and tuned open-source precision models from gpt-oss-20b for intent classification, prompt-injection defense, toxicity, evaluation and reranking, with its HyperClassifier running in 26ms against 1,446ms. Its own engineering post says a frontier foundation model still handles the core multi-step reasoning.¹⁷³ Cursor built Composer 2 by continued pretraining and large-scale RL on the open-weight Kimi K2.5 base, with only about a quarter of the final model's compute coming from that base.¹⁷⁴

#The counter, stated in full

Now the objection, and it is a strong one. Menlo Ventures measured $37 billion of enterprise generative AI spend in 2025. It found that Anthropic, OpenAI and Google hold 88% of enterprise LLM API usage, while open-source share fell to about 11% from 19% in 2024.¹⁷⁵ Menlo's mid-year update put open source at 13% of AI workloads and said builders consistently choose frontier models over cheaper, faster alternatives.¹⁷⁶ SemiAnalysis wrote in May 2026 that open-source models are still noticeably worse than their closed source counterparts for real knowledge work. It added that there is no reason to believe the gap will close any time soon.⁷⁸

There is more. a16z's January 2026 survey of 100 Global-2000 executives found preference for closed-source models has increased steadily since March 2024. The same survey found enterprises trending away from fine-tuning open models toward prompt engineering and routing, and CIOs reporting total cost of ownership between open and closed converging.¹⁷⁷ Ramp found only 6.1% of AI-using businesses touch open-model serving platforms at all.¹⁷⁸ ServiceNow retired its in-house Now LLM, which as of the July 2026 release is no longer the default model for any AI skill or agent.¹⁷⁹ Anthropic reported a $30 billion revenue run rate as of April 2026, with over 1,000 customers spending more than $1M a year.¹⁸⁰ That is not a market walking away from the frontier.

My answer is narrow. Menlo's shares are estimated enterprise API dollars spent, not usage or token share, and since open-weight tokens cost roughly an order of magnitude less, the token share would be materially higher.¹⁷⁵ A share measured in API dollars is the meter measuring itself.

Look at a venue that publishes both halves. On Vercel's AI Gateway in July 2026, open-weight models ran 36% of tokens at about a seventh of frontier rates. Open-weight spend more than doubled to nearly nine cents of every dollar. Anthropic collected 65.1% of all spend on 30% of volume, and the average price paid per token fell 13.6%.¹⁸¹ Three months earlier the open-weight token share was about 11%.¹⁸² Dollar share and token share are moving in opposite directions in the same dataset.

OpenRouter's weekly rankings through September 1, 2026 put at least six confirmed open-weight models in the top ten by tokens, with GPT-5.6 Luna the only US-lab proprietary model on the list. The site says plainly that these rankings measure adoption, not quality.¹⁸³ Ramp adds the qualifier that matters: among businesses using open-model serving platforms, 96.4% also pay OpenAI or Anthropic, and they spend a median $248 per employee on AI, 23x the median AI-spending business.¹⁸⁴ Nobody is swapping the frontier out. They are sizing it down to the calls that need it.

#The claim that survives

The frontier is not needed for most calls. It is still needed for the calls that decide whether a long agentic task succeeds, and enterprises will keep paying for that tail. In LangChain's run that tail was 7% of calls and 68.4% of the bill. The other 93% is a capacity workload. It is steady, predictable, and it runs fine on models an enterprise can hold in its own racks.

That is the split SAVRN is built around. We run the open stack on capacity the enterprise buys outright, and we do not charge more when the enterprise runs more of it.

---

#8. Open weights with the data attached

When an enterprise asks me what it will actually own, it is not asking about benchmark scores. It is asking what it can inspect, retrain, and keep. That question has a different answer in 2026 than it had two years ago, because a class of models now ships more than the weights. Some ship the recipe. A few ship the data too.

The largest company building in that class is NVIDIA. That is worth sitting with for a moment, because NVIDIA sells the hardware these models run on.

#What actually ships with a model

"Open" covers two very different products. One is a file of numbers you may use. The other is the file plus the code and the corpus that produced it.

The Llama license carries a condition most people never read. Above 700 million monthly active users you must request a separate license from Meta.¹⁹²

#What NVIDIA released, stated precisely

I want to be careful here, because the marketing overshoots. NVIDIA's blog says Nemotron 3.5 Lightning "is fully open" on weights, data, and recipes, so "you can adapt it to your workflows and deploy it anywhere".¹⁹³ The model cards say something narrower. The Nemotron 3 Ultra card says "Major portions of the pre-training corpus are released in the Nemotron-Pre-Training-Datasets collection".¹⁸⁵ Major portions is not all portions.

NVIDIA's own recipe repository is blunter. It warns that the published recipes "train exclusively on the open-sourced subset of training data" and that results "will differ from the tech report benchmarks, which used additional proprietary data".¹⁹⁴ The Nemotron 3 Nano technical report says the same thing. In the paper's own words: "Along with the model weights, we provide the recipe, code, and most of the data we used to train the model".¹⁹⁵

Here is the arithmetic. Nemotron-CC-v2 reports 6,585.8 billion tokens. Add Nemotron-CC-v2.1 at 2,544.8 billion, CC-Code-v1 at 427.9 billion and Pretraining-Code-v2 at roughly 340 billion, and you get about 9.9 trillion released pretraining tokens. Nemotron 3 training saw roughly 25 trillion. That is about 40% by token count, and it is an upper bound, because tokens repeat across epochs.¹⁹⁶ Two of the largest pieces are gated. Nemotron-CC-v2.1 and Nemotron-Pretraining-Code-v2 both require contact information and acceptance of the NVIDIA Data Access Agreement for Model Training. That agreement, the card says, "is designed to limit use to those uses protected as fair uses, such as model training".¹⁹⁷ The Nemotron 3 Super card lists 153 datasets used, categorized as public sources, NVIDIA-crawled data, synthetic datasets, and third-party private datasets.¹⁹⁸

So the accurate sentence is weights, recipes and most of the data. That is still a large release. The Nemotron Post-Training v3 collection holds 43 datasets used to post-train Nano, Super and Ultra.¹⁹⁹ One detail in there is worth knowing before you build on it. The teacher models behind Nemotron-Post-Training-Dataset-v2 are DeepSeek-R1-0528 at 5.7 million samples, Qwen2.5-14B-Instruct at 3.9 million, Qwen2.5-32B-Instruct-AWQ at 1.0 million, and Qwen3-30B-A3B and Qwen3-235B-A22B at 627,720 each.²⁰⁰ The American open-data flagship was post-trained largely on distillation from Chinese open-weight models.

NVIDIA's corporate blog names the reason it publishes any of this. It publishes as much of the training data and techniques "as licensing permits, which allows for traceability, auditing and training of other models".²⁰¹ Traceability and auditing are procurement words, not research words.

#The class that releases everything

Ai2's OLMo 3 is the reference implementation of the other model. Ai2 releases "Every stage, checkpoint, dataset, and dependency required to create and modify" the model. That includes Dolma 3 at about 9.3 trillion tokens, the Dolmino mid-training set, the Dolci post-training sets, and the training and eval code.¹⁸⁶ Dolma 3 is 9.31 trillion tokens across 9.97 billion documents under the Open Data Commons Attribution License.²⁰²

The cost of that openness is smaller than people assume. On Ai2's own card, Olmo 3.1 32B Think scores 96.2 on MATH against 95.4 for weights-only Qwen3-32B, 80.6 on AIME 2024 against 80.8. It scores 91.5 on HumanEvalPlus against 91.2, 86.4 on MMLU against 88.8, and 88.6 on BigBenchHard against 90.6.²⁰³ Ai2's Nathan Lambert wrote that the Olmo 3 Thinkers "get within 1-2 points overall with their respective Qwen3 8/32B models".²⁰³

Apertus is the version of this built for a regulator. EPFL, ETH Zurich and CSCS released weights, intermediate checkpoints, training source code and datasets. The model "was developed with due consideration to Swiss data protection laws, Swiss copyright laws, and the transparency obligations under the EU AI Act".¹⁸⁷ It respects machine-readable opt-outs retroactively and reaches business customers through Swisscom's Swiss platform.

#The license, and why it is not Apache

Two documents govern the Nemotron releases. The first is the Nemotron Open Model License already quoted: perpetual, royalty-free, no claim on outputs. Its only termination trigger is patent or copyright litigation against the licensor.³⁶

OpenMDW-1.1, published by the Linux Foundation, goes further. It defines Model Materials to include the model plus associated data, documentation and software, and grants rights under copyright, patent, database and trade-secret law. It also states that it "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs generated by using the Model Materials".²⁰⁴ NVIDIA adopted it across Cosmos, Isaac GR00T, Ising and Nemotron on May 28, 2026.²⁰⁵

The gap that closes matters to a lawyer more than to an engineer. Linux Foundation SVP Legal Mike Dolan explains that Apache and MIT rely on copyright. Weights are machine-generated numbers that may not be copyrightable, so a recipient of Apache-licensed weights may hold uncertain rights. OpenMDW affirmatively conveys "copyright, patent, database, trade secret rights that may have been embedded in there".²⁰⁶

#Why the data is the part that counts

Regulation already asks for it. EU AI Act Article 53(1)(d) obliges providers of general-purpose models to publish a training-content summary. The open-source exemption in Article 53(2) covers only points (a) and (b), so the copyright policy and the training-data summary stay mandatory for open models too.²⁰⁷ The Commission's template requires data source types and origins, individually named large datasets, licensing agreements, the "top 10% of domain names used", user-interaction data, and the source model for synthetic data. AI Office enforcement runs from August 2, 2026, with fines up to EUR 15 million or 3% of global turnover.²⁰⁸

Compliance so far is thin. A data-vendor tally found seven providers had published by the enforcement date. Anthropic, Mistral and xAI "replaced the boxes with narrative phrasing" describing proprietary blends, and no fines or probes existed as of August 7, 2026.²⁰⁹ The exposure the author points to is not the fine. It is procurement due diligence on data lineage, which is the conversation I have with regulated buyers every week.

The Open Source Initiative states the technical version of the same point. "Without training code or intermediate checkpoints, researchers and auditors cannot replicate the model's development process".²¹⁰ Open-weight releases usually do not clarify how the dataset was built or cleaned, so nobody outside can assess its quality or diversity.

#The NVIDIA mirror

NVIDIA's FY2026 Form 10-K puts the motive in one sentence. "More recently, in support of market development, we have accelerated the release cadence of our open AI model platforms including NVIDIA Nemotron for agentic AI and Cosmos for physical AI".⁵⁴ The same document carries the other edge, a few pages earlier in the management discussion: open-source AI "is dependent on developer adoption and if deployed on our competitors' platforms, it could reduce demand for our products and services".⁵⁴

The $26 billion open-model figure belongs to WIRED rather than to the SEC, as set out earlier. WIRED also quoted NVIDIA VP Bryan Catanzaro saying the company "is taking open model development much more seriously". It reported that Nemotron 3 Super shipped with weights, 10 trillion tokens of pretraining data, 40 million post-training samples, 15 reinforcement-learning environments and complete training recipes.²¹¹

The program keeps growing. The Nemotron Coalition was announced March 16, 2026 with Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab. It commits that its first base model will be open sourced and will underpin Nemotron 4, though the release does not promise the coalition's training data will be published.²¹² Reuters, citing The Information, reports Nemotron 4's largest model is expected to have "at least 1 trillion parameters", with final training incomplete and no release date set.²¹³

Jensen Huang wrote the customer version of the argument in July 2026. Open weight models "help provide that assurance by allowing organizations to control their own data, evaluate and adapt models to their own needs, and deploy them wherever their business requirements demand".²¹⁴ The letter carries 25 signatories in the published PDF, including Dell, Hugging Face, IBM, Meta, Microsoft, Mistral, Mozilla, Palantir and ServiceNow. Security Boulevard reports "Within a day the signatory count had doubled to 50" with OpenAI and Google joining, and that Anthropic and Amazon did not sign.²¹⁵

None of this makes NVIDIA an opponent of metering. Its own technical blog tells telco operators to move "from Compute-as-a-Service to Token-as-a-Service", billing by tokens, requests or workflows.⁵⁵ Free models and a paid meter downstream sit together fine when you sell the machine underneath both.

#The objection: open weights are a lower grade of the real thing

Say it in full, because it is partly true. Epoch AI finds that since January 2026 the most capable open-weight models have lagged frontier closed models by an average of four months and about 8 points on its capabilities index. The fully open class sits behind the open-weight leaders in turn.¹⁵³ Buyers noticed, and Menlo's falling open-source share is the number they noticed with.

The answer is the workload, not the leaderboard. For the job an enterprise runs at volume, a model it can retrain on its own data beats a frontier model it cannot touch, and the companies doing it are named. CrowdStrike, Harvey, CodeRabbit, Lila Sciences and Fastino Labs are customizing Nemotron 3.5 Lightning.²⁰¹ The Nemotron 3 launch named Accenture, Cadence, CrowdStrike, Cursor, Deloitte, EY, Oracle Cloud Infrastructure, Palantir, Perplexity, ServiceNow, Siemens, Synopsys and Zoom.²¹⁶ Huang states the routing logic plainly. Open weights "let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else".²¹⁴

The gap that matters to a buyer is not four months of benchmark. It is whether anyone outside the vendor can open the box.

A chip company does not spend billions publishing free models for the good of the field. It does it because of who its two biggest customers have become.

---

#9. The army

Two customers. That is what NVIDIA's fiscal 2026 concentration disclosure comes down to. One direct customer was 22% of total revenue and another was 14%.⁵⁴ A year earlier the top three were 12%, 11% and 11%. The same filing says some indirect customers each exceed 10% as well.

OpenAI, non-NVIDIA~16.75 GW

Broadcom custom 10 GW, AMD Instinct 6 GW, Cerebras 750 MW

OpenAI, NVIDIA~12 GW

Existing and planned commitments through 2030, per NVIDIA's Q2 FY2027 call

Anthropic, Trainiumup to 5 GW

More than $100B over ten years with AWS, over one million Trainium2 chips already serving Claude

Anthropic, Google TPU~3.5 GW

Next-generation TPU capacity through Broadcom, from 2027

Anthropic, NVIDIA~1 GW

Initial commitment under the November 2025 Microsoft and NVIDIA agreement

Both labs are also hiring silicon teams of their own. Anthropic confirmed one in August 2026. Read the bars as diversification rather than exit: in the same year it expanded Trainium and TPU, Anthropic signed roughly $80 billion of NVIDIA-based neocloud capacity.

That is a company that knows exactly where its risk sits. What NVIDIA did next is the response.

#The line in the MD&A

Where NVIDIA put that open-source sentence is the part worth knowing. It sits in Item 7, the management discussion, not in the risk factors.⁵⁴ The Item 1A language about open-source models is a different topic. That one concerns regulatory restrictions on supporting Chinese open models such as DeepSeek, Qwen and KIMMI.

The same 10-K says customers with in-house expertise can build their own replacements, and discloses $17.5B invested in private companies in fiscal 2026, including AI model makers that purchase NVIDIA products directly or through cloud providers.⁵⁴

So the customer concentration, the substitution risk and the investment program all sit in one document.

#What the two customers are building

OpenAI and Broadcom unveiled Jalapeno on June 24, 2026, OpenAI's first custom inference chip.²¹⁷ It went from design to production in nine months, with help from OpenAI's own models. It is built around ChatGPT, Codex, the API and future agentic workloads. Engineering samples run GPT-5.3-Codex-Spark. Initial deployment was targeted by the end of 2026, and the chip is the first of a multi-generation custom roadmap.²¹⁸

Nine months. That is the number I keep coming back to. A chip program on that clock changes what a supplier can count on.

Add the announced non-NVIDIA programs together and OpenAI has about 16.75 GW of accelerators on the board. Broadcom is at 10 GW. AMD is at 6 GW across multiple Instinct generations, with the first gigawatt starting in the second half of 2026. Cerebras is at 750 MW through 2028 for about $10B.²¹⁹ Against that, CFO Colette Kress told the Q2 FY2027 call that "OpenAI's existing and planned commitments represent approximately 12 gigawatts of NVIDIA compute" through 2030.³⁹

Hold that comparison loosely. The figures are announced, multi-year, and measured on different bases. AMD and Broadcom are quoted in GPU and accelerator gigawatts, Cerebras is 750 MW of compute, and NVIDIA's Portsmouth number is IT-load gigawatts. The comparison is directional only.

Anthropic went the same direction from a different starting point. The Trainium and TPU commitments are already on the table: more than $100 billion and up to 5 GW with AWS, about 3.5 GW of Google TPU capacity through Broadcom, and over a million Trainium2 chips already training and serving Claude.¹⁰² What came next is new. On August 5-6, 2026 Anthropic confirmed it is hiring an in-house silicon team to co-design custom chips for Claude, with roles paying $320K to $485K.²²⁰

For scale: Anthropic's first deep NVIDIA partnership only arrived in November 2025, at the $30B of Azure capacity and up to 1 GW already described, with NVIDIA investing up to $10B alongside it.²²¹ Five gigawatts of Trainium against one gigawatt of NVIDIA systems is not a rounding difference.

#The correction I have to make

"The minute they build chips they stop buying NVIDIA" is a clean line, and 2026 does not support it.

In the same year Anthropic expanded Trainium and TPUs, it signed the roughly $80B of NVIDIA-based neocloud capacity set out earlier, Nscale and Lambda together.²²² Add the $30B Azure commitment and the total across late 2025 into 2026 runs to about $110B. It also took the full Colossus 1 cluster from SpaceX in May 2026: more than 300 megawatts and over 220,000 NVIDIA GPUs, delivered within the month.⁶⁸

Anthropic frames its own silicon team as complementing, not replacing, hardware from AWS, Google, NVIDIA and AMD.²²⁰ It says it matches workloads to the chips best suited for them.²²³ The accurate claim is diversification and hedging on both sides, not exit. NVIDIA is still the fastest path to capacity, which is why a TPU-first lab took a cluster of over 220,000 NVIDIA GPUs inside a month.

The hedge runs the other way too. NVIDIA's 10-K discloses an intellectual property license with Groq that required significant, nonrefundable payments, $27B of multi-year cloud commitments, and $3.5B of land, power and shell guarantees to early-stage partners in exchange for warrants.⁵⁴

#Where the demand base gets rebuilt

NVIDIA's answer to concentration is not a better chip argument. It is a different customer.

Start with the models. The $26B open-weight program, the Nemotron Coalition and the released datasets are all described earlier. Add the money that does not show up in a model card: a $6B non-exclusive license with Poolside for its Model Factory, plus a $1B investment at a $12B pre-money valuation, with some outlets reporting the package as $7B.²²⁴

Then look at who signed the letter, and who did not. Counts diverge by source, from dozens to 150-plus depending on the count date, which is why the 25 initial signatories and the reported doubling to 50 within a day are both in the record.²²⁵ What does not diverge is the absences. OpenAI signed within 24 hours of launch. Anthropic and xAI had not signed as of July 28. Dario Amodei responded on July 27 that "Anthropic has never advocated for a ban on open-weights models", calling instead for chip export controls, distillation limits and safety testing.²²⁶

Now the revenue line. The ACIE segment is the $40.3B already broken out, growing 138% against hyperscale's 102%. Huang drew the contrast himself: these customers "don't buy custom chips, they don't buy chips one at a time. They really need an entire factory platform".³⁹

That is the whole strategy in one sentence from the CEO. A lab buys chips. An enterprise buys a factory.

On the same call, alongside the roughly $15B of on-prem vertical revenue already noted, Huang named Amdocs, Palantir, Cadence, Siemens and Perplexity as Nemotron adopters.³⁹ NVIDIA's DSX platform, announced May 31, 2026, is pitched as a complete playbook for infrastructure builders rather than a chip line, though its named adopters are neoclouds and OEMs.²²⁷ And at GTC 2026 Huang said "Every company in the world today needs to have an OpenClaw strategy, an agentic systems strategy", unveiling NemoClaw, an enterprise-secured agent stack that runs locally on enterprise infrastructure.²²⁸

Huang also says nearly all open models run on NVIDIA, because its footprint is the largest and its architecture the most fungible.²²⁹ The open model is not a threat to the chip. It is the thing that gets the chip into an enterprise that will never call a lab.

#Where NVIDIA lends credit, it buys exclusivity

The Ohio structure is the clearest statement of intent in the record, and it is in a filing, not a press release.

On August 17, 2026 NVIDIA disclosed residual value guaranties on SB Energy leases covering about 4.25 GW of IT load at the Portsmouth site. Its aggregate payment obligation is capped at $105B, triggered only by OpenAI insolvency-driven lease default or failure to pay. "OpenAI has agreed to reimburse and indemnify NVIDIA for any and all amounts actually paid by NVIDIA to the Lessor".⁹⁶ The 8-K states that OpenAI will use that capacity "to deploy NVIDIA's full-stack DSX AI factory platform, subject to limited exceptions", and the exceptions are not defined in the filing.⁹⁶

The 10-Q says why. "In exchange for the guarantees, the site will exclusively host NVIDIA AI infrastructure, subject to limited exceptions".²³⁰ The same filing says the guarantees become effective per lease across nine construction phases, with the first expected in fiscal 2029. None of the $105B was effective at July 26, 2026.

The press release does not hedge either: "NVIDIA will be the exclusive AI compute infrastructure provider at PORTS-Pike".²³¹ NVIDIA also holds a unilateral option to extend credit support to roughly 3.8 GW more at the same site, exercisable in its sole discretion.²³²

Set the $105B guarantee cap next to the $3.5B maximum gross exposure across the land, power and shell guarantees to early-stage partners.²³⁰ Of that $108.5B of guarantee exposure, $105B, about 97%, is the piece conditioned on NVIDIA-only hardware. That share can move materially: the cap is not yet effective, and the 3.8 GW option is still open.

Kress described the mechanism plainly. "The LPS commitment secures a long-lived AI factory site, while the NVIDIA compute within the data center can be upgraded repeatedly".²²⁹ The lock is on land, power and shell for twenty years. The chips refresh each generation inside it. And a residual value guarantee is not spent money, because NVIDIA delivers nothing today.²³³

#Where it buys equity, it gets none

Here is the part that cuts against the clean version of this story.

The $40B of equity that went into the two labs, $30B into OpenAI and up to $10B into Anthropic, carries no disclosed chip condition. The Anthropic announcement discloses no NVIDIA exclusivity and instead stresses Claude running on all three major clouds.⁸⁵ Four and a half months later Anthropic signed for multiple gigawatts of Google and Broadcom TPU capacity.²²³

The OpenAI story is sharper. The September 2025 letter of intent was lock-in by construction: up to $100B invested progressively as each gigawatt of NVIDIA systems was deployed, at least 10 GW, with NVIDIA named preferred strategic compute and networking partner.²³⁴ The first $10B was to land only "once the two sides reached a definitive agreement for OpenAI to purchase Nvidia systems". That agreement took much longer than expected, and the LOI was replaced by an unconditioned $30B stake.²³⁵ The chip-purchase condition moved from covenant to expectation.

A competitor did the conditioning NVIDIA's own equity did not. AMD's July 2026 investment of up to $5B in Anthropic is tied to deployment milestones for up to 2 GW of Instinct MI450.²³⁶

#The objection

Someone will say NVIDIA is the labs' biggest financier, not their enemy. That is correct, and both things are true at once. Kress told the same August call that demand from the AI labs NVIDIA backs with its balance sheet is expected to contribute toward roughly a quarter of its business next year. She addressed the criticism head on: "We know some will call this circular financing. We see it differently".²²⁹

The filings tell you which instrument is which. Equity went in unconditioned. Credit went in with an exclusivity term written into the 10-Q. As one reading of the reversal puts it, what Huang cast doubt on was buying more shares, and what NVIDIA is offering now is its credit.²³³

The control instinct has limits, and it found one in August when the revenue-share program was paused. Providers objected that client selection was their prerogative and NVIDIA staff warned of antitrust exposure.⁴⁹ Exclusivity written into a lease held. A share of somebody else's rent did not.

So the second objection, that the minute the labs build chips they stop buying, does not hold in 2026 either. What is happening is diversification on the labs' side and demand-base construction on NVIDIA's. The army being raised is the enterprise, and the enterprise does not buy chips one at a time.

It buys a factory, which means it buys everything that runs on top of the chips. That software layer turns out to matter more than the model in it.

---

#10. The harness is the product

When a vendor tells me a new model changed everything, my first question is what else changed. The model is one part. The harness is the rest: the tools the model can call, the sandbox it runs in, how context gets assembled, when it retries, when it stops.

Most people buy the model and inherit the harness. The research from 2026 says the wrapper is doing more of the work than the thing inside it.

#What the controlled studies actually found

A May 2026 paper made the argument directly. The authors write that the agent execution harness "is often a stronger determinant of agent performance than the model it wraps".²³⁷ They scope that claim to models of comparable frontier capability. I will come back to that.

Their own controlled run is three harnesses crossed with three models on a 100-task SWE-bench Verified subset. GPT-5.4 landed at 55, 58.5 and 63.5. Kimi K2.6 landed at 52, 59 and 60.5. GLM-5.1 landed at 52.5, 56.5 and 65.5. Aggregate harness variance came out 7.80 times model variance.²³⁷

The same paper cites a figure people quote a lot. Under the standardized SEAL scaffold on SWE-bench Pro, Claude Opus 4.5 reaches 45.9%. Under Claude Code, the same model reaches 55.4%. That is 9.5 points. Say what it is not. Those figures came from external sources, a leaderboard on one side and an Anthropic-reported result on the other, so it is a cross-source comparison and not a paired run.²³⁷

Two other groups ran the paired version. Peking University's Harness-Bench collected 5,194 trajectories and found a 23.8-point gap between the best and worst configurable harness on identical tasks and identical model backends, NanoBot at 76.2% against OpenClaw at 52.4%.²³⁸ A GAIA study found scaffold choice alone moves measured accuracy by as much as 28 percentage points inside one model, in the robust slice for Opus at Level 2. In the primary slice the same cell is 55.8% against 69.8%, a 14-point gap.²³⁹

HarnessX froze the models and rewrote only the harness across five benchmarks, for an average gain of 14.5% and a maximum of 44.0%.²⁴⁰ Stanford searched harness designs automatically. Adding environment bootstrapping to an existing agent took Claude Opus 4.6 to 76.4% on Terminal-Bench 2.0 across 89 tasks and five trials, with the model unchanged.²⁴¹

One more GAIA swing gets quoted often in this debate. It could not be traced to a primary source, so it is not in this piece.

#The number that decides your bill

Accuracy is the part everyone quotes. Token consumption is the part that shows up on the invoice. Sentient Labs measured both. Harness choice moves tokens per solved task by up to 40x on Terminal-Bench Pro. Upgrading the model moves it 1.0 to 1.3x. Within a single model, pass-rate differences across harnesses ran 0 to 8 points.²⁴² The same paper points at two leaderboard entries for Claude Opus 4.5. One harness scored 52.1% and burned 256.9M input tokens. Another scored 57.8% and burned 3.9M. That is a 65x token difference for 5.7 points of accuracy.²⁴²

Those two sentences are a purchasing problem. Under per-token pricing, the harness sets your bill, and it can set it 65 times apart for nearly the same answer. NVIDIA says the same thing from the infrastructure side: token volume per user request in agentic systems is unpredictable and often orders of magnitude higher than single-shot inference, which makes average-based cost modeling unreliable.²⁴³

Under capacity pricing that variance is free. You bought the machine. A 40x swing in tokens per solved task costs you nothing but time.

#Where the harness argument stops

The effect is scoped, and I will not overstate it. Harness-Bench found weaker models show larger variance across harnesses. HarnessX found gains are largest where baselines are lowest. Harness engineering does the most for the model that needs the most help.

The model still moves things. Under Vals AI's fixed, minimal, bash-tool-only harness, Claude Opus 5 scores 97.00% on SWE-bench Verified against Opus 4.8 at 88.60%. That is 8.4 points from one model generation with the harness held constant.¹⁵⁸ And on long-horizon agentic work the gap can be far larger than any harness. Artificial Analysis runs Terminal-Bench v2.1 under a fixed harness and scores Claude Fable 5.1 at 91.4%. NVIDIA's own card puts Nemotron 3 Ultra at 56.4% on the same benchmark.¹⁶¹ Roughly 35 points, and no measured harness swap closes it.

Say what that 35 points is and is not, because the same board carries a much smaller number. On Terminal-Bench 2.1 the leading open-weight models, Kimi K3 and GLM-5.3, sit about three points behind Fable 5.1. The 35-point gap is specific to Nemotron 3 Ultra, a 550B model NVIDIA gives away and tunes for agent harnesses, not to the open-weight field. Both figures are correct. Which one applies depends on which open model you actually run.

So the claim I will defend is narrower than the headline. Between models of similar class, the harness decides more than the model does, and it decides the token bill by a wider margin than it decides the score.

#NVIDIA is giving the harness away

Here is what turns this from a research story into a business story. NVIDIA post-trained Nemotron 3 Ultra for other people's harnesses. The June 4, 2026 announcement names Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands and OpenCode.²⁴⁴ The flatness is deliberate. NVIDIA reports SWE-bench Verified between 65% and 70.4% across Pi, OpenHands, Hermes, OpenCode and Mini SWE Agent.¹⁸⁵ On the smaller model the spread is tighter still: Nemotron 3 Super scores 60.47% in OpenHands and 59.20% in OpenCode, 1.27 points apart, once NVIDIA tuned both harnesses for the model.¹⁵⁹

That is the whole trick. If the harness is the biggest lever, and NVIDIA tunes every major harness to its free model, the lever sits on the enterprise's side of the table.

LangChain showed it from the other direction. It tuned its Deep Agents harness for Nemotron 3 Ultra without retraining the model and reached an aggregate 0.86 on its own evaluation. It did that at $4.48 per run against $43.48 for the closest competitor, a 9.7x difference. LangChain's summary of the work: "Every gain came from engineering the environment around the model".²⁴⁵

The runtime is open too. OpenShell is a kernel-sandboxed, policy-enforced runtime in which OpenClaw, Anthropic's Claude Code and OpenAI's Codex run unmodified.²⁴⁶ Its privacy router keeps sensitive context on local open models and routes out to frontier models like Claude and GPT only when policy allows.²⁴⁷ NVIDIA did not displace the lab tooling. It wrapped it and put a policy engine around it.

Huang's line at GTC 2026 about every company needing an OpenClaw strategy was a statement about this layer, and he put OpenClaw alongside Linux, Kubernetes and HTML when he said it.²⁴⁸

#The stack inventory

Here is what is actually free and what is not.

The agent layer is free. The factory management layer is not. That is a fair trade when the factory is yours.

#What you still buy from a frontier lab

LangChain's Switchyard run, already on the table, gives the shape: 7% of calls to the frontier model, 68.4% of the cost, and a 74% cut in total spend for six points of accuracy. The authors caution that this is one workload where the small model happened to be close.¹⁶⁹

Other partners reported the same shape. NVIDIA cites Ramp at 58% lower cost and 33% faster runtime, and Cognition's Devin Desktop at 28% lower mean cost.²⁵⁶ NVIDIA's AI-Q release claims more than 50% off query cost for a hybrid design, frontier models for orchestration and Nemotron for research.²⁵⁷

Now the objection, in full. The router still buys the hard 7% from the labs, so you are back on the meter. It is worse than it sounds, because that 7% is 68.4% of the routed bill. You did not escape metered pricing. You concentrated it.

That is correct, and it is the design. Seven percent on a meter and ninety-three percent on a rack is the outcome I want. The 7% is the bounded part you can forecast. The 93% runs on hardware you already paid for, where a harness change that burns 65 times more tokens costs nothing extra, and where running the cluster harder makes every token cheaper.

Compare the alternative on offer. Anthropic Enterprise is $20 per seat with usage billed at API rates and no bundled token allowance.²⁵⁸ Azure's provisioned units are billed hourly on deployed capacity whether or not the deployment is handling requests.²⁵⁹ One charges you for every experiment. The other charges you for the hour whether you experiment or not.

SAVRN builds and runs the factory, and runs this stack on it. The customer contracts for capacity, we do not bill them per token, and when policy sends a call to a frontier model it goes out on the customer's own account.

---

#11. The end of the package

I have bought a lot of software in my career. I have also watched a market decide, in about six weeks, that it had been overpaying for it.

1. Feb 17, 2026

   Partnership deepens

   Figma and Anthropic jointly launch Code to Canvas. Anthropic's Chief Product Officer sits on Figma's board.

2. Apr 14, 2026

   The board seat ends

   Mike Krieger resigns from Figma's board, disclosed the same day in a Form 8-K. The Information reports that day that Anthropic's next model will carry design tools.

3. Apr 17, 2026

   Claude Design ships

   Built by Anthropic Labs, the team Krieger led. Three days after the resignation. The launch post names Canva and never mentions Figma. Figma falls 6.89%.

4. Apr 20, 2026

   The accusation

   Figma's chief executive says Anthropic "were not consistently candid in their communications."

5. May 28, 2026

   The demand

   Activist holder Findell Capital asks Figma's board to investigate whether Anthropic improperly used confidential information.

6. Aug 5, 2026

   The business, meanwhile

   Revenue up 48% to $370.1M, third straight quarter of acceleration, guidance raised twice. Gross margin down from 89% to 84% and a $117.3M operating loss, on AI inference costs paid to model providers.

The competing product cost Figma 6.89% in a session. The token bill cost it five points of gross margin and swung the company to a loss. Only one of those showed up in the headlines.

The labs are moving up the stack. They are doing the work that packaged applications used to own. The market has already priced that threat, and it priced it hard. What I want to do here is separate the part of the claim that is true from the part that is not, because the difference decides where an enterprise should put its money.

#Six weeks in early 2026

Anthropic launched Claude Cowork on January 12, 2026 as a research preview, described as "Claude Code for the rest of your work." It went first to Max subscribers on the macOS desktop app and reached Pro subscribers on January 16.²⁶⁰

On January 30 the company added plugins and open-sourced eleven of its own: Productivity, Enterprise search, Plugin Create/Customize, Sales, Finance, Data, Legal, Marketing, Customer support, Product management, and Biology research.²⁶¹ Those are not model releases. Those are job functions with vendors attached to them.

The market read it that way. The February 3, 2026 legal plugin sent Thomson Reuters and LegalZoom down more than 15% in a single day, with RELX and FactSet posting double-digit drops, and Salesforce and Workday falling earlier the same week.²⁶² Reuters carried a "nearly 18%" figure for Thomson Reuters that day, along with "biggest daily loss on record" and "lowest close since June 2021." Those were intraday "on track for" statements rather than confirmed closes.²⁶³ I keep the hedge because the hedge is the fact.

The index move was not a one-day event. The S&P Software and Services Index lost 25% between January 12 and February 23, 2026, and four days after the plugin release it dropped 4% and wiped out over $285B of market capitalization.²⁶⁴ Jefferies strategist Jeffrey Favuzza gave it a name, SaaSpocalypse.

Then the multiple went somewhere it had never been. Software forward price to earnings fell to 22.7x over January to March 2026, below the S&P 500 for the first time. IGV was down over 21% year to date and roughly 30% from its September 2025 peak, about $2 trillion of market value.²⁶⁵ Software had traded at a premium to the index for as long as anyone had been measuring. That premium is gone.

The people running the labs did not soften it. At Anthropic's May 6, 2026 financial services briefing, Dario Amodei said individual SaaS incumbents could lose value or go bust depending on their response, and that moats built on software complexity are going away.²⁶⁶ His words: "I think individual SaaS companies, it's very possible for them to lose market value, go bankrupt, completely, go bust, but it depends on the response."

OpenAI moved on the same ground. It launched Frontier on February 5, 2026, an end-to-end platform for enterprises to build and manage AI agents that connect to external data and applications, with pricing undisclosed.²⁶⁷

And the hiring told you where this was going before the products did. On July 15, 2025 Anthropic hired Paul Smith as its first Chief Commercial Officer, out of ServiceNow, and before that Salesforce and Microsoft.²⁶⁸ You do not hire that resume to sell tokens.

#The Figma sequence

One case has the whole mechanism in it, with dates, and it is documented in SEC filings rather than in commentary. I want to lay it out in order and then say what it does and does not prove.

Anthropic's Chief Product Officer, Mike Krieger, sat on Figma's board of directors. On February 17, 2026 the two companies jointly launched Code to Canvas, which pulls Claude Code output into Figma as editable layers.²⁶⁹ That is a partnership getting deeper, not thinner.

On April 14, 2026, Krieger resigned from Figma's board. Figma disclosed it that day in a Form 8-K.²⁷⁰ The same day, The Information reported that Anthropic's next model would carry design tools competing with Figma's core product.²⁷¹

Three days later, on April 17, Anthropic launched Claude Design.²⁷² It was built by Anthropic Labs, the team Krieger led. The launch post names Canva as a collaborator and quotes Canva's CEO. It does not mention Figma at all.

Figma fell 6.89% that session.

Then Figma's chief executive said something you do not often see one company say about another in public. Dylan Field said Anthropic "were not consistently candid in their communications".²⁷³ That is the phrase OpenAI's board used when it fired Sam Altman. On May 28, 2026 the activist holder Findell Capital demanded that Figma's board investigate whether Anthropic had improperly used Figma's confidential information before launching the competing product.²⁷⁴

Board seat, joint launch, resignation, competing product three days later, a public accusation, and a demand for an investigation. As a sequence it is exactly what people mean when they say a platform learned the business and then took it.

#What the Figma sequence does not prove

Now the part that cuts the other way, and it cuts hard enough that I will not write the sentence people expect me to write here.

Figma is down about 79% from its all-time-high close of $122.00, so "down at least half" is true measured from the peak. It is not true measured from the $33 IPO price, which is a 21.8% decline.²⁷⁵

The timing is the problem. Roughly 83% of that fall had already happened before Claude Design existed. Figma went from $122.00 on August 1, 2025 to $20.32 on April 16, 2026, the last close before the launch, on staggered lockup releases, insider selling and decelerating growth. Claude Design cost 6.89% in a session. Figma's worst single day of 2026 was not the Anthropic launch at all. It was August 6, when the stock fell 14.85% on AI cost inflation and softer guidance.²⁷⁶ And the stock now sits about 27% higher than it did the day before Claude Design shipped.

The business went the other way from the narrative entirely. Figma's revenue growth accelerated straight through the launch: 40%, then 46% to $333.4 million in the first quarter of 2026, then 48% to $370.1 million in the second.²⁷⁷ Net dollar retention hit 139%, its highest in over two years. Guidance was raised twice after Claude Design launched. Figma's own prepared remarks a month after the launch never mention Anthropic, Claude or Claude Design once.

Figma also still runs on Claude. Its own 10-Q says Claude is the model on which it built the AI features in its government offerings.²⁷⁸

#What actually transferred

Here is the thing I did not expect to find, and it is the reason this section belongs in this piece at all.

Something real did move from Figma to Anthropic in 2026. It was not Figma's customers and it was not Figma's growth rate. It was Figma's margin.

Figma's GAAP gross margin fell from 89% to 84% year over year, and the company swung from $2.1 million of operating income to a $117.3 million operating loss, driven by the AI inference costs it pays to model providers.²⁷⁸ The stock's worst day of the year was the market working that out.

Read those two facts together. The competing product took 6.89% in a day and the business shrugged it off. The token bill took five points of gross margin and swung the company to a loss. The application company that builds on a metered API does not usually lose its customers to the model provider. It hands over its margin, one invoice at a time, and the invoice grows with its own success.

That is the same structure as every other section of this piece, arriving from a direction I did not go looking in. Figma's risk factor now says it out loud: its model suppliers may at the same time offer competing products.²⁷⁸ Both halves of that sentence cost money. Only one of them showed up in the stock.

#The correction I have to make

The tempting story is that the labs eat the applications and the vendors die. Seven months of evidence say the opposite happened to the biggest one.

Anthropic and Salesforce started tightening in October 2025. On October 14, 2025 Claude became a preferred model in Agentforce 360 for regulated industries. Anthropic became the first LLM provider fully inside the Salesforce trust boundary, with all of Claude's traffic contained in the Salesforce virtual private cloud. Salesforce deployed Claude Code across its engineering organization.²⁷⁹

Ten months later it went further, not away. On August 26, 2026 the two announced Claudeforce. Claude is the default model for Slack AI, Slackbot, Salesforce in Claude, Headless 360 and Agentforce Coworker. Salesforce is named Anthropic's preferred CRM. Actions route through Salesforce so business rules are enforced.²⁸⁰ The Salesforce in Claude plugin ships with 37 prebuilt sales skills, in pilot at announcement with open beta expected in September 2026, and no financial terms were disclosed.²⁸⁰

The numbers that landed the same day were not the numbers of a company being eaten. Salesforce reported Q2 FY27 revenue of $11.3B, up 11%, and current remaining performance obligation of $33.5B, up 14% in constant currency. Agentforce ARR passed $1.5B, up over 240%, Agentforce plus Data 360 ARR landed near $3.9B, and the company raised FY27 guidance to $46.1B to $46.4B.²⁸¹ Run the arithmetic on that and Agentforce is about 3.2% of guided revenue, and Agentforce plus Data 360 about 8.4%.

The stock rose more than 12% after hours. Quarterly net income of $3.53B included a $2.6B investment gain tied in part to the company's Anthropic stake. On CNBC, Marc Benioff called the SaaSpocalypse narrative nonsense and said, "Frontier models depend on CRM. They don't replace it." ²⁸²

#Does the sequence generalize

Figma is one case. The interesting question is whether it is the first of a pattern or the exception in one, and the way to test that is the order of the dates.

For the companies that got hit hardest in February, the order runs backwards. Claude Cowork shipped January 12, 2026. The eleven job-function plugins shipped January 30. The legal plugin shipped February 3. The partnerships came after that, on February 24 and later: Thomson Reuters, LegalZoom, FactSet, DocuSign and Intuit.²⁸³ Product first, partnership second. Whatever those deals were, they were not a lab learning a business before entering it. They read more like a truce after the damage.

The companies whose jobs Claude covers most directly are also not the ones in trouble. Harvey competes with Claude's legal work as squarely as any company alive, and its annualized revenue went from $195 million at the end of 2025 to about $350 million by July 2026.²⁸⁴ Glean sits underneath the enterprise search plugin and grew about 44% in five months. Cursor competes head-on with Claude Code and its annualized revenue went from roughly $1 billion to an estimated $4 billion.

No partner has gone bankrupt. The only partner products that shut down were ones Anthropic bought outright.

So I will state the boundary of the claim plainly, in both directions.

The mechanism is real, it is documented, and in the Figma case it involved a board seat, a resignation three days before launch, a public accusation from the partner's chief executive and an activist demand for an investigation. Anyone building a company on top of a frontier lab should read those filings before they sign. What is not established is that this is a repeated strategy rather than one case, and what is contradicted is that it destroys the companies it touches. The stock moves were violent and the revenue lines did not follow them.

#The lab that said it out loud

There is one more thing in the record, and it is the part that should actually worry a software buyer. The clearest statement of intent to replace the application vendors does not come from Anthropic. It comes from OpenAI.

OpenAI executives told prospective investors that their AI products and agents would replace software from Salesforce, Workday, Adobe, Slack and Atlassian.²⁸⁵ That is a named list, delivered to the people funding the company.

There is precedent for what such assurances are worth in the other direction too. In May 2023, at a private London meeting with developers and founders, Altman said OpenAI would not release more products beyond ChatGPT, reasoning that great platform companies have one killer app. The blog post recording that promise was later taken down.²⁸⁶

So the shape of the risk is right even where the name attached to it is wrong. If you are an enterprise deciding where to put the next decade of your software budget, the question is not which lab is friendly. It is what you own when the friendliness ends.

#What is actually dying

The seat is dying. That is a different thing from the system of record dying.

Workday prices mostly on customer headcount. On March 13, 2026 the stock sat at $133, down 38% year to date and 44% over twelve months. The argument came down to a single line: "AI agents do not have seats." ²⁸⁷

Salesforce has been rebuilding its meter around the same problem. Agentforce moved to consumption pricing with Flex Credits at 20 credits, ten cents, per action, sold as pre-purchase, pay as you go or pre-commit. In August 2026 Benioff said the company is "still trapped in some ways in old per user pricing models" and sees an enormous opportunity in more aggressive value pricing.²⁸⁸

Satya Nadella said the structural version of this in December 2024. Business applications are CRUD databases with business logic on top, and "the notion that business applications exist, that will probably collapse in the agent era" as the logic moves to the AI tier.²⁸⁹ The database survives that. The per-person license does not.

#Build against buy, with the evidence on both sides

Two notes on that table. The roughly 1,200 shut-down applications everyone quotes from Klarna is an internal estimate relayed by a vendor, not an audited count.²⁹⁵ And the MIT NANDA gap, read literally, is about 22% against 67%, a 3x gap; the widely repeated 33% figure is a secondary-source paraphrase. Either reading leaves purchased tools succeeding at least twice as often as internal builds.

Most enterprises are not choosing a side. In Anthropic's 2026 survey of more than 500 US technical leaders, 47% take a hybrid approach combining off-the-shelf agents with custom components. Another 21% rely entirely on pre-built agents, and 20% build their own with APIs, open-source models or developer toolkits.²⁹⁶

#The objection, in full

The objection is that Salesforce, Workday and ServiceNow are all growing double digits, so the package is not dead. That objection is correct on its facts.

Workday's Aneel Bhusri put the buy-side case bluntly. His words: "I have not met a single customer who is looking to replace Workday with something they are building internally or buying from a start up." Gartner analysts said in February 2026 that predictions of the death of SaaS were premature. In their reading, Cowork and its plugins are "potential disrupters for task-level knowledge work but are not a replacement for SaaS applications managing critical business operations".²⁹⁹ Thomson Reuters CEO Steve Hasker said the selloff "represents anxiety and not fundamentals".³⁰⁰

The second objection is that enterprises cannot build software, which is why SaaS exists in the first place. MIT NANDA is real and I am not going to argue with it. What changed is the cost of a build and the location of the data. The systems of record stay. The layer of work on top of them is what moves.

#The asset is the data

Watch what the vendors do about data, not what they say about AI.

Slack's API Terms of Service took effect on October 10, 2025. They forbid third-party apps from using API Data to train a large language model, from bulk-exporting message and file data, and from keeping persistent copies, archives or indexes.³⁰¹ Glean's read was that the change limits access to a query-by-query basis and locks the data inside Slack.³⁰² That is a company that understands exactly what it owns.

The buyers understand it too. BNP Paribas built an internal LLM as a Service platform hosted in the Group's own data centers on GPU infrastructure, running open-source models and Mistral AI models, with models trained on internal datasets to follow.³⁰³ JPMorgan Chase built LLM Suite in house and onboarded 200,000 users within eight months of its summer 2024 release.³⁰⁴

Neither of those banks built a frontier model. They built a place to put their data and a way to run work against it.

The survey data points the same direction. Kong's 2025 enterprise report found 44% of enterprises cite data privacy and security as the top barrier to LLM adoption. Another 31% name security and privacy compliance as the top factor when choosing an LLM provider.³⁰⁵

To be fair to the labs, the default terms are already reasonable. Anthropic does not use inputs or outputs from its commercial products for training unless a customer submits feedback or opts in.³⁰⁶ A contractual promise and a machine you own are different kinds of protection. Enterprises are learning to tell them apart.

SAVRN builds the second kind. We give an enterprise a rack of capacity it controls, running open-weight models and NVIDIA's stack, so the work sits next to the data instead of leaving with it.

Which brings the argument back to the invoice, and to what every metered vendor did to it this year.

---

#12. A meter punishes use. A rack rewards it.

I keep a folder of pricing-page screenshots. It started as a sales habit. It turned into the clearest record I have of what actually happened since the summer of 2025.

Every metered AI product in that folder moved in the same direction. Each move made the heaviest users pay more. Not one of them made heavy use cheaper.

#The moves, in order

Anthropic's public pricing page now lists Enterprise at $20 per seat with usage billed at API rates and no bundled token allowance.²⁵⁸ Its own service-tiers page says Priority Tier capacity commitments are no longer available for purchase.⁵ BCG put it flatly in July: the era of flat-fee pricing models for enterprise users of AI compute has ended.⁸⁴

The Copilot change is the one I would point a CFO at. In GitHub's own community thread, users reported 54% of a monthly quota consumed by a single request and 8% consumed in two hours. One reported four agents consuming 53% of quota at "at minimum 50x prior consumption".³¹² One developer wrote that to get the same level of coding as the week before, he needed "around 10 times the price of the entire monthly subscription." Nothing about his work changed. The invoice changed.

#Why the seat had to break

The seat economics explain the whole table. Citing SemiAnalysis data, one analysis puts a Claude Max seat at roughly 90% gross margin at 1% utilization and about negative 900% at full exhaustion.³¹³ A flat seat sold on top of a metered cost base is a bet that the customer will not use it. When agents arrived, the customer started using it around the clock, and the bet went bad.

So the vendors did what the math told them to do. They repriced the power user by gating new capability rather than raising the headline price. Anthropic never published token quotas for any plan.³¹⁴ Overage on Pro, Max, Team and Enterprise is billed at standard API rates with a $2,000 daily auto-reload limit.³¹⁵

Odlyzko wrote down the consequence twenty-five years ago, from inside AT&T Labs, and it is the finding this whole section runs into: tiny charges based on utilization decrease usage substantially, and in a rapidly growing market the provider's interest is to encourage usage.¹² AOL's usage tripled when the meter came off. Every product in the table above moved the other way.

#What the meter does to the people using it

Anthropic's own Claude Code documentation puts enterprise deployments at about $13 per developer per active day and $150 to $250 per developer per month, with 90% of users below $30 per active day.³¹⁶ Those are the users who stay inside the fence. The ones who do the most work do not.

CloudZero, citing Forbes, puts Uber's per-engineer Claude Code bills at $500 to $2,000 a month across roughly 5,000 developers.³¹⁷ That is the spend behind the exhausted annual budgets at Uber and ServiceNow noted earlier. Epoch AI reports that some of Apple's software engineers are permitted to use up to $300 in tokens per day, about 5 million output tokens at Claude Opus 4.7 pricing.¹⁴¹

Meta ran the largest version of this experiment. Fortune reported that Meta's internal dashboard showed employee usage above 60 trillion tokens in a 30-day period across more than 85,000 employees, with the top individual averaging 281 billion tokens.³¹⁸ By June, per The Information, monthly usage had peaked at 73.7 trillion tokens and internal AI was projected to cost billions in 2026. Meta said it plans to curb employee AI usage by imposing token limits through an internal gateway.³¹⁹ The dashboard came down. The cap went on.

Now set that next to what the company selling the machines tells its own engineers. At GTC in March, Jensen Huang said every NVIDIA engineer will get a token budget worth about half their base pay, on top of salary. His words: "I'm going to give them probably half of that on top of it as tokens so that they could be amplified 10 times".³²⁰ Days later he said he would be "deeply alarmed" if a $500,000 engineer consumed less than $250,000 of tokens a year. Asked whether NVIDIA spends about $2 billion a year on tokens, he answered "We're trying to".³²¹ At GTC Taipei in June, holding up a new chip, he described a personal agent that "could run 24/7, meter free... it's running all the time, no meter anxiety, and it's yours".³²²

Two companies, two instructions to their own staff. One is capping consumption. The other is subsidizing it. The difference is which side of the invoice they sit on.

#Capacity products sold by metered vendors

The metered vendors do sell capacity. They just do not price it as a reward for volume. OpenAI's Scale Tier and Azure's provisioned units are both described earlier: a unit of throughput per day, or per hour, bought in advance. Two details of the Azure product belong here rather than there. A deployment cannot be paused, and billing stops only on deletion. Reservations discount the hourly rate at $260 per Global PTU for one month and $2,652 for one year ([Azure Retail Prices API - Foundry Provisioned Throughput Reservation](<https://prices.azure.com/api/retail/prices?$filter=contains(productName,%27Provisioned%27)%20and%20contains(productName,%27Foundry%27)%20and%20armRegionName%20eq%20%27eastus2%27>)). When the units run out, the API returns 429 with optional spillover back to pay-per-token.¹²¹

That structure does reward use. Once you have paid for the hour, running it flat out costs nothing extra, so unit cost falls as load rises. The catch is where the price starts. Both products are priced at or above list, which makes them a premium for predictability rather than a discount for volume.²⁵⁹ A capacity charge sold by a company whose margin comes from the meter behind it will never go far below the meter.

#On a rack the marginal token is free

Buy the machine and the arithmetic inverts. NVIDIA states it on its own inference-economics page: "A cluster running at 40% utilization produces twice the effective cost per token compared to the same cluster at 80% utilization, regardless of the hardware platform".³²³ Double the work on fixed capacity and each token costs half as much. The customer keeps that.

Software moves the same curve without touching the hardware. NVIDIA reports that TensorRT-LLM and Dynamo cut cost per million tokens from $0.11 at launch to $0.02 within two months on fixed hardware.¹⁴³ DSX MaxLPS targets maximum token performance per megawatt and lets operators run up to 40% more GPUs at their most efficient point.²⁵³ Huang's framing at Taipei was "throughput per watt is revenue".³²² The slogan people actually quote him on is "the more you buy, the more you save".³²⁴

#The objection: 5% utilization

Here is the case against everything above, stated in full. Cast AI's 5% average GPU utilization, with its exclusions and its vendor interest already stated, is the number that kills a capacity contract.³²⁵ A customer who buys a rack and runs it at 5% has paid twenty times the price per token he was quoted, and he does not renew.

That is the seller's problem, and I treat it as ours. It is the reason a rack has to ship with the harness already on it and with someone accountable for load factor. NVIDIA gives most of that away. NemoClaw runs OpenClaw, Hermes and LangChain Deep Agents inside OpenShell with managed inference, under Apache 2.0.²⁵⁰ NeMo Switchyard routes traffic across providers and translates between OpenAI and Anthropic APIs.²⁵¹ DSX OS handles scheduling, health automation and multi-tenant operation.²⁵³

The second half of the answer is that the median enterprise is not the buyer. The agent-heavy enterprise is. Agentic work burns about 15x more tokens per request than ordinary human usage, and agentic tokens passed human-usage tokens on OpenRouter around February 1, 2026.³²⁶ Anthropic's own docs note that agent teams use roughly 7x the tokens of standard sessions.³¹⁶ A company running agents around the clock does not have a utilization problem. It has a bill.

#Our numbers

We at SAVRN carry a rack at $3.66 million a year, which is about five cents a million tokens at full load and twenty-four to thirty-three cents a million at 65% utilization once sales and overhead are carried in it. Those are our own model's figures, derived earlier, not a published benchmark.

Set that against the Uber range above. At $2,000 per engineer per month across roughly 5,000 developers, twelve months of metered spend multiplies out to about $120 million. At the $500 end of the same range it is about $30 million. Ramp's April card data put blended realized rates at $1.00 per million for Claude Opus 4.6 and $0.62 for Sonnet 4.6, after caching and volume discounts. Median business AI spend was $46 per employee per month.³²⁷ The realized meter is far below list and still an order of magnitude above what the machine costs to run.

What SAVRN sells is the machine and the load factor. That means NVIDIA's latest GPUs at time of delivery, the open harness already running on them, and a contract where the buyer stops paying per token and starts paying for the thing that makes them. Our incentive after signature is to get the racks busier. That is the whole reason I stopped selling the meter.

---

#13. Who owns the residual

I have signed for hardware that a newer model made look old before it was racked. Every buyer of a multi-year contract carries that fear. What happens when the next one lands?

NVIDIA answered that question a long time ago. At Computex on June 2, 2024 Jensen Huang described the cadence himself. "Our company has a one-year rhythm. Our basic philosophy is very simple: build the entire data center scale, disaggregate and sell to you parts on a one-year rhythm".³²⁸ The Rubin launch release of January 5, 2026 claims the platform delivers up to 10x reduction in inference token cost compared with Blackwell.³²⁹

A new generation every year, and a tenfold cut in cost per token claimed for the newest one. Set that against a five-year contract and the deal looks foolish. The rental data says otherwise.

#The chip does not die at year five

The A100 shipped in 2020. On its August 11, 2026 second-quarter call CoreWeave said it had recently signed an A100 contract that extends into 2029 at an attractive price. On the same call it said it had implemented a 25% price increase across SKUs in July.³³⁰ Two days later Huang posted that "The mighty A100 fleet are mission-capable from 2020 through 2029," and gave the credit to CUDA. Silicon Data had posted on August 11 that the A100 stopped depreciating in late 2025 as rental prices rose.³³¹ On the November 19, 2025 call Colette Kress said the A100 GPUs shipped six years earlier were still running at full utilization on a vastly improved software stack.³³² Silicon Data's live index read the A100 at $1.29 per GPU-hour as of the September 2, 2026 fetch.³³³

The H100 tells the same story one generation later. SemiAnalysis reported on April 2, 2026 that "H100 1-year GPU rental contract pricing has shot up almost 40% from a low of $1.70/hr/GPU in October 2025 to $2.35/hr/GPU by March 2026." The same piece said some H100 contracts were being renewed for four years through 2028. It also said half the providers it asked were completely sold out.³³⁴ Four-year renewals signed in 2026 on a prior-generation part.

Here is the part that cuts the other way, and it matters. The price recovery is a neocloud story. Silicon Data's history shows the segments moving apart.³³⁵

Silicon Data reported on July 31, 2026 that hyperscaler H100 rates fell 26 to 54% by region since the first quarter of 2025. Non-hyperscaler rates bottomed in mid-2025 and recovered, which compressed the hyperscaler premium from roughly 250% to about 120%.³³⁶ So the H100 residual is holding in one channel and falling by half in another. Both readings are real.

That segmentation is also why the H100 price appears in this piece in more than one form and every one of them is current. SemiAnalysis's $1.70 to $2.35 is a one-year contract rate. Silicon Data's $6.26 and $3.33 are channel medians. Quote the series, not just the number.

The people who own the fleets are not confused about the direction. Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five effective January 1, 2025. That raised 2025 depreciation by $1.4 billion and cut net income by $1.0 billion. The stated reason: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning".³³⁷

Amazon paid $1.0 billion of net income to move its book life from six years to five. CoreWeave is signing A100 rentals that run to 2029.

#The building is the stranding risk

Rack power is where the real obsolescence lives. ModulEdge lays out the curve. Hopper sits at roughly 30 to 40 kW per rack, GB200 and GB300 NVL72 at roughly 120 to 130 kW, and Vera Rubin NVL72 at roughly 190 to 230 kW. Kyber NVL576 arrives at roughly 600 kW in 2027, and the Feynman era lands in the 1 MW class.³³⁸ That piece states the conclusion in one line: "The asset at risk isn't the chip. It's the building."

NVIDIA's own engineering blog says it is leading the transition to 800 VDC data center power infrastructure to support 1 MW IT racks and beyond, starting in 2027. The same post says 54 VDC distribution hits physical limits above 200 kW.³³⁹ A published roadmap summary of June 8, 2026 puts Rubin Ultra Kyber racks in the second half of 2027 at 600 kW per rack with 800 V DC distribution.³⁴⁰

A hall built for 200 kW racks does not take a 600 kW rack. The chip keeps earning. The concrete around it does not.

Abilene is the case study. CNBC reported on March 9, 2026 that OpenAI stopped expanding its Abilene partnership with Oracle "because it wants clusters with newer generations of Nvidia graphics processing units, according to a person familiar with the matter." The same report said the site was expected to use Blackwell and that power was not projected to come online for a year.³⁴¹ OpenAI's own compute head, Sachin Katti, gave a different explanation. "We considered expanding it further, but ultimately chose to put that additional capacity in other locations".³⁴² Take the company at its word and the lesson does not change.

Whatever the reason, the position was Oracle's to hold. Oracle's own 10-K for the year ended May 31, 2026 says why that hurts. The terms and renewal options in its long-term data center leases "typically do not align with the duration and pricing of customer contracts." And "if customers do not renew their contracts, we may be unable to re-lease, repurpose or assign such capacity on acceptable terms, if at all".³⁴³ On July 9, 2026 S&P downgraded Oracle to BBB- with a stable outlook, citing AI infrastructure debt. S&P put OpenAI at about half of Oracle's $638 billion of remaining performance obligations.³⁴⁴

Ohio is the same shape, built deliberately. The $105 billion of guarantees over 4.25 gigawatts under a 20-year SB Energy lease is described earlier; NVIDIA is also investing $1.5 billion in SB Energy.⁹⁵ What matters here is the refresh arithmetic inside it. NVIDIA's CFO commentary says each generation deployed there could represent approximately 1.5 million GPUs, or approximately $150 billion to $200 billion in NVIDIA revenue. It also says that over 20 years the site can support multiple infrastructure upgrade cycles.⁴¹ A one-year chip rhythm inside a 20-year lease means the hardware turns over many times against a single turnover of the shell.

#Who is holding the bag, by name

The lab is not. Read the documents.

The $108.5 billion of maximum gross guarantee exposure has already been broken out. What has not is when NVIDIA actually pays. The 8-K says only on a Trigger Event, defined as OpenAI's insolvency causing a lease default or OpenAI's failure to make lease payments. Then NVIDIA "will pay an amount generally equal to any shortfall between the guaranteed minimum value of a lease (as set forth in the Agreements) and amounts recovered through a replacement lease or sale".⁹⁶ The same filing adds that the obligation ends on "OpenAI achieving a satisfactory credit rating," and that OpenAI indemnifies NVIDIA for anything it does pay. Guarantee amounts step up as each construction phase completes and decline as OpenAI pays its rent.⁵²

So the guarantor pays only if the tenant stops paying, and the tenant pays the guarantor back. That is not a subsidy. It is a credit wrapper on somebody else's lease.

NVIDIA does the same thing on the capacity side. Its 10-Q describes a new business model. In it, "AI clouds procure our data center infrastructure products and we commit to cloud service agreements, which the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates." Those commitments are typically six years in duration and totaled $36 billion as of July 26, 2026.⁵² It has done it in public since 2025. CoreWeave's 8-K for a September 9, 2025 order form with an initial value of $6.3 billion says "NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032".³⁴⁵ Alongside that sit equity investments of $99 billion and equity investment commitments of $25 billion.⁵²

NVIDIA says out loud why it is doing this. Its risk factor reads: "AI clouds and AI model makers have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity".⁵² The counterweight sits in the same paragraph. It is the sentence I would put on the wall: "We expect our large cloud service provider customers and investment grade enterprises to continue to secure land, power, and shell commitments independently".⁵² NVIDIA underwrites the parties that cannot underwrite themselves. It expects everyone else to stand on their own paper.

The lab-side financings are built the same way. Epoch AI describes Anthropic's structure as a five-year compute lease from a dedicated equipment company, with the racks as collateral: "Broadcom can take over the lease or arrange for the racks to be sold. The sale proceeds are used to repay investors," and the junior tranche relies more heavily on the value recovered from the racks.¹⁰⁴ The lab rents. The lab owns nothing.

On August 10, 2026 NVIDIA announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital. Huang said "In AI, compute is revenue," and described NVIDIA compute as fungible and transferable across customers and operators.³⁴⁶ That release does not contain the words residual value, guarantee or backstop. It says only that the machine is "continuously improved through CUDA software," which extends its useful life and improves its economics over time.³⁴⁶ Seven days later the 8-K used the words the press release avoided.

#The residual is about to get a public price

On August 11, 2026 CME Group and Silicon Data announced cash-settled H100 and B200 rental index futures. Each contract covers one month of rental for one GPU, standardized at 730 GPU-hours and listed on NYMEX, targeting an October 5, 2026 launch pending regulatory review.³⁴⁷ The release calls it "a public, tradable reference price for the resource every AI system runs on." The word token does not appear in it. The unit the market chose to hedge is the GPU-hour.

That is the whole argument in one product listing. When an asset gets a forward curve, the residual stops being a matter of opinion and becomes a number somebody will quote you.

Here is the rule I run by, and it is older than any of this. Never hold a long upstream obligation against a short downstream one. Oracle held a lease that outlived its customer's interest in the hardware. NVIDIA holds guarantees that only pay if the tenant stops paying rent, with the tenant on the hook to pay it back. SAVRN builds and operates the campus and sells the enterprise a fixed term on a rack, and the generation refresh sits inside our capital plan rather than the customer's.

Which leaves the objection sitting right where a careful buyer would put it. You are asking me for five years at a fixed price on a machine whose maker says the next one cuts cost per token by 10x.

---

#14. The instrument: a toll, not a PPA

People ask me what the contract is. Not the pitch. The paper. What does an enterprise sign when it buys a rack instead of a meter.

The answer is a tolling agreement. The buyer pays a capacity payment for dispatch rights over a machine for a fixed term. The owner warrants that the machine runs. The buyer decides what the machine makes. The envelope is written into the document so both parties know exactly what changed hands.

That is a well-worn instrument. It is not a power purchase agreement, and it is not "unlimited tokens." It is also not new to compute. The wholesale layer of this industry already signs a version of it.

#What the wholesale layer already signs

CoreWeave's IPO prospectus says its customers make large infrastructure commitments structured as take-or-pay contracts with terms typically two to five years in length.³⁴⁸ In both the first and second quarters of 2026, committed take-or-pay contracts were 98% of the company's revenue. Those contracts start either on a fixed date or when the contracted capacity is delivered.³⁴⁹ On the May 7, 2026 call, management said weighted average contract length for new capacity remains at approximately five years.³⁵⁰

The lenders write the same shape. SemiAnalysis puts it bluntly. Outside of the four corners of a 5-year hyperscale backstopped compute deal, the appetite to lend drops off almost entirely, and five years is the most common offtake period.³⁵¹

Then watch August. CoreWeave closed a $2.6 billion DDTL 5.5 facility with an approximate five-year maturity. That maturity extends beyond the roughly three-year average length of the underlying customer contracts. The facility is priced at SOFR plus 5.50% and rated Ba2 by Moody's and BB plus by Fitch.³⁵² Five-year money against three-year paper. The lender is underwriting the renewal, not just the term. A capacity market is forming, and it is forming above the enterprise, not at it.

NVIDIA sits underneath all of it with a price curve. SemiAnalysis describes NVIDIA's neocloud backstop as six years long. The illustrative curve runs from $3.68 per hour in year one to $1.88 in year six, averaging $2.36. NVIDIA takes a 40% share of revenue earned above the backstop, for an effective take rate of about 18% over six years.³⁵¹ SemiAnalysis is explicit that the curve is an illustrative example at the lower end of the backstop range. Take it as shape, not a quote sheet. The shape is a declining floor price per unit of machine, over six years, with upside shared above it.

#Why it is not a power purchase agreement

Mistral's European Compute Units are the cleanest end-user example on the record, and the terms are already set out: roughly five-year commitments from five named industrial anchors, with no early exit. The detail that matters for the instrument is what the units buy. They are spendable on inference, training, model adaptation or managed Kubernetes, at the customer's discretion.¹¹⁷ The buyer chooses the workload. The seller warrants the machine.

Luca Pedretti of Pexapark read the same deal and named the instrument. He argues the ECU is a tolling agreement rather than a PPA. His reason: "Rather than buying output, the offtaker pays for the right to direct an asset, dictating how that machine is dispatched." The offtaker absorbs the commercial risk of how it monetizes that capacity, and the owner warrants operational performance through uptime guarantees backed by liquidated damages.³⁵³

A PPA buys output at a strike price. A toll buys the machine's time and leaves the output to the buyer. Under a toll the buyer stops paying per token and starts paying for the machine that makes them.

#The farmer and the mill

The oldest description of this is still the best. Under a tolling agreement the toller has the right, but not the obligation, to call on the owner to convert fuel into electricity at a predefined heat rate. For those dispatch rights the toller pays a capacity payment. "In exchange, the toller pays a fee just as a farmer would pay a mill to grind his wheat into flour".²⁴ The same guidance notes that the purchaser keeps making capacity payments even when the power actually delivered falls short of the contracted amount.

The farmer owns the wheat. The farmer owns the flour. The mill owns the mill and gets paid to turn. Substitute data for wheat and tokens for flour and you have the enterprise AI contract I think the market is heading toward.

#Four prices, not one

A gas toll is not a single number. Puget Sound Energy's prototype tolling term sheet has the Buyer deliver fuel and receive all energy, ancillary services and capacity rights, with PSE assumed to be the exclusive toller and non-exclusive arrangements considered. Its Contract Price has four parts. There is a Monthly Capacity Payment in dollars per MW of Tested Capacity and a Variable O&M Charge in dollars per MWh. Then a Start-Up Charge in dollars per start, and a Heat Rate Adjustment measured against a Guaranteed Heat Rate.³⁵⁴ The filing labels the pricing structure illustrative only.

Four prices. The rack toll needs the same four, and they map almost one for one.

The capacity payment is the bankable piece. It is the fixed leg that turns a merchant asset into something a lender will size debt against. The operations charge covers what actually moves when the machine runs. The start-up charge stops a buyer from parking a partition cold for a quarter and demanding it hot in an hour. The performance adjustment is the buyer's remedy when the machine underdelivers. Every one of those four is a number a CFO can model.

#The envelope

The word "unmetered" does not mean unbounded. It means the meter is not the pricing instrument. What replaces the meter is a written envelope, and it has four terms. The model class the rack is licensed and sized to run. The numerical precision it runs at. The tokens per second per concurrent user it will sustain. And the availability it is warranted to hold. Those are the specifications the owner guarantees. Everything inside them belongs to the buyer.

Compare that to what a metered platform warrants. Anthropic's rate-limit documentation says its numbers are maximum allowed usage, not guaranteed minimums.⁷ A ceiling that is not a floor is a rationing device, not a service level. A toll inverts that. The number in the contract is what the buyer is owed, not what the buyer is allowed.

Around the envelope go three commercial terms I would not sign a rack without. Stepped pricing, so the capacity payment declines across the term the way NVIDIA's backstop curve declines. Refresh options, so the buyer can move to the next generation of NVIDIA's latest GPUs at time of delivery on a priced option instead of a renegotiation. And a benchmark-linked adjustment, so the guaranteed throughput is measured against a stated harness and a stated model class rather than a marketing figure. SAVRN writes its enterprise capacity contracts this way, with the four prices and the four envelope terms on the face of the agreement. That is the whole product description.

#What the owner keeps, and what still floats

The owner keeps the residual. When the term ends, the machine is the owner's to redeploy, resell or refresh, and that residual is priced into the capacity payment from day one. The owner also keeps the utilization services: scheduling, batching, placement, the work of getting more out of the same silicon. Those are the two positions that make a toll worth writing from the seller's side.

The price of the machine's time is now hedgeable, at least in part. The October 5 CME listing described earlier turns a month of GPU rent into a contract a treasurer can trade, which is what CME's global head of energy meant in saying the contracts "will now turn compute into a standardized, tradable commodity".³⁴⁷

That is a rent hedge, not a token hedge, and the difference is the basis risk sitting inside every unmetered rack. A published analysis of token futures argues that GPU compute futures face obstacles from rapid hardware iteration and standardization. It concludes that token futures are more feasible, with an optimal launch window estimated at 2027 to 2028.³⁵⁵ A buyer hedged on GPU rent is not hedged against the token price it competes with downstream. Say that out loud in the deal. Do not pretend the futures market closes the gap it does not close yet.

#The objection

The strongest version is this: a toll moves utilization risk and price risk onto the buyer, and the buyer is the worst-positioned party to carry either.

The evidence for that objection is real and it points the other way from my thesis. BCG's declaration that the flat-fee era has ended, and its 5% utilization figure, both sit against a fixed capacity payment.⁸⁴ Five percent utilization against a capacity payment is a bad trade in any industry. And a five-year commitment with no early exit is exactly the exposure Mistral's own CTO described when he said there is no getting out.

My answer has two parts, and neither says the risk is imaginary.

First, the seller does not walk away from the asset. The seller holds the residual and sells the utilization services, which means the seller is paid to raise the buyer's utilization rather than to watch it sit at 5%. A toll where the owner keeps redeployment rights is an owner who has to care whether the machine earns.

Second, the direction of the price is opposite under the two instruments. Under a meter, heavy use raises the bill and the platform's remedy is a cap that tightens as you grow. Under a toll, heavy use lowers the effective price per unit of output, because the capacity payment is already spent. The buyer who runs a rack twice as hard pays half as much for the same token. That is the incentive an enterprise wants pointed at its own engineers, and it is the one a metered contract cannot give it.

---

#15. The objections, answered in order

I have argued one line for the length of this piece. Here is the case against it, stated the way the people who disagree with me state it. Where the objection is right, I say it is right and then say what follows from it.

#1. Basis risk

The objection: you are selling a five-year fixed price on a machine whose maker promises in writing to cut the cost of a token tenfold in the next generation.

That is right, and the promise is in NVIDIA's own launch release. The Rubin platform release of January 5, 2026 claims up to a 10x reduction in inference token cost compared with the Blackwell platform.³²⁹ A buyer who fixes a price against a cost curve like that is taking basis risk.

Two facts bound it. The machine has not been getting cheaper. One-year H100 contract pricing climbed almost 40%, from $1.70 an hour in October 2025 to $2.35 by March 2026. B200 capacity is described on September 1, 2026 as extremely limited, with pricing inflated by supply constraints.³⁵⁶ The token got cheaper. The machine got more expensive.

The second fact is about who keeps the savings. Efficiency gains land with whoever holds the meter, which is the whole content of NVIDIA's telco guidance and the subject of objection six below. SAVRN writes the refresh into the term, so the buyer sits on NVIDIA's latest GPUs at time of delivery rather than on the generation they signed against.

#2. The window

The objection: the cheapest adequate metered token is $1.20 out, $0.60 in batch. Your spread is 6x, not 100x, and it lost 80% of its width in a single afternoon in July.

I concede this one in full. It is the strongest objection here. OpenAI's pricing page lists GPT-5.6 Luna at $0.20/$1.20 standard and $0.10/$0.60 batch per million input and output tokens.⁶³ That price arrived on July 30, 2026 as an 80% cut from $1.00/$6.00, about three weeks after launch, which OpenAI attributed to internal efficiency gains.³⁵⁷ Three days before the cut, Moonshot AI released the full weights of the 2.8-trillion-parameter Kimi K3 for free public download.³⁵⁸ CloudZero states the consequence without softening it. At Luna-class token rates, some workloads that used to justify provisioned capacity may never reach break-even on PTUs at all, because the metered price fell out from under the reservation.³⁵⁹

What survives is narrow, so I will keep it narrow. The floor price carries an expiration date in the vendor's own footnote. OpenAI labels Sol's $4/$20 promotional through at least November 21, 2026.⁶³ DeepSeek's cheap number is the off-peak number, with V4 Pro at $1.32/$3.96 per million tokens at peak and $0.66/$1.98 off-peak.¹⁶⁶ A capacity contract is priced for five years. A promotional token price is priced through November.

#3. Utilization

The objection: enterprise clusters idle at 5%. That customer does not renew, which is why Anthropic, OpenAI and GitHub all killed their flat fees this year.

Right on both counts, and the record is already in front of you: 5% average utilization, the flat-fee era declared over, the seat-plus-meter move at Anthropic and the Codex move to token credits.⁸⁴ The Register adds one mechanic worth having. Legacy flat plans transition at renewal rather than immediately.³⁰⁹

Look at why the sellers moved. A Claude Max seat at 1% utilization yields roughly 90% gross margin and about -900% at full exhaustion. Providers repriced power users by gating new capability rather than raising headline prices, as GitHub Copilot did in June 2026 when it moved to usage credits.³¹³ A flat price that loses 900% on the customers who actually use it was mispriced. Withdrawing it is a repair, not a refutation.

The utilization number sets the real test. Together AI's reserved dedicated H100 endpoints run $2.55 to $2.99 an hour, which at $2.99 is $2,153 a month and competitive only above about 80% sustained utilization, while production AI workloads average under 50% sustained.³⁶⁰ Capacity pays when one machine carries the whole book of work. It does not pay when it carries one team's pilot.

#4. Occupied position

The objection: Together, Fireworks, Baseten, Lambda and Nebius already sell dedicated unmetered capacity today, and Mistral already sells the five-year version with no exit. The position is taken.

It is taken. That tells me the demand is real. Fireworks bills on-demand dedicated deployments per GPU-second at $8.00 an hour for H100 and H200 and $13.00 for B200 from September 1, 2026, up from $7.00 and $10.00 through August 31.³⁶¹ Baseten bills per minute, about $6.50 an hour for H100 and about $9.98 for B200, kept explicitly separate from its per-token model APIs.³⁶² Lambda lists B200 on-demand at $6.69 to $6.99 per GPU-hour and reserved 1-Click Clusters at $8.87 to $9.86. Nebius lists H100 at $3.85 and B200 at $7.15 on-demand, with up to 35% off for multi-month cluster reservations.³⁶³

Note what they are. They are hourly GPU rentals with terms running from a minute to a year, and their prices went up this year while the metered token fell. Mistral is the one selling the instrument this piece describes, on the terms already given: roughly five years, five named European anchors, no early exit.¹¹⁷

#5. The hard 7%

The objection: NVIDIA's own router still buys the hard cases from the labs.

True, and NVIDIA says so in its own documentation. OpenShell's privacy router keeps sensitive context on local open models and routes out to frontier models only when policy allows.²⁴⁷

The quality gap behind that routing decision is real and I will not talk around it. SemiAnalysis's judgment that open models remain noticeably worse for real knowledge work, and the 43% to 46% open-weight scores on TerminalBench Hard against 61%, are both stated in full earlier.¹⁵²

I am not arguing the meter goes to zero. The Switchyard result routes 7% of calls to a frontier model and cuts cost 74%.²⁰¹ Ninety-three percent of the calls run on the machine the enterprise already bought. Seven percent go out at frontier rates. That shape is a capacity base with a metered exception on top, which is the arrangement this piece has argued for from the first page.

#6. NVIDIA is agnostic about the unit

The objection: NVIDIA tells telcos an H100 earns about $157,680 a year on a meter and about $18,400 by the hour. Its own advice is to meter.

The figures are NVIDIA's and the advice is real. Its developer blog tells telcos to move from GPU-hourly rental to Token-as-a-Service, and recommends pricing per million tokens with per-tenant enforcement. It says revenue is no longer limited by how many hours a GPU can be rented but by how many tokens the stack can serve.⁵⁵

Ask who is being advised. A telco that owns GPUs is a reseller. NVIDIA is telling that reseller how to charge its own customers more per GPU. The advice is about the retail layer, and it concedes the argument rather than rebutting it. If the same machine bills many times higher on a meter, the party downstream of the meter is the one paying the difference.

Then look at how NVIDIA prices its own book. Its July 2026 AI Compute Partnership offered neoclouds a rental backstop, with NVIDIA renting unused GPU capacity itself, in exchange for a 50% cut of rental revenue above an agreed cost threshold.³⁶⁴ Its guarantees are written on land, power and shell for about 4.25 GW of IT load.⁵² Not one of those instruments settles on a token.

#7. Scarcity, not ideology

The objection: the labs buy flat because they are supply-constrained, not because flat is the better unit. Give them supply and they will still sell you a meter.

I think that is correct, and it is the spine of my argument rather than a hole in it. A meter is a rationing device. Anthropic is preparing to go public carrying roughly $450 billion of announced compute commitments against a run-rate revenue the company itself put at $47 billion in May 2026.³⁶⁵ A buyer in that position takes every megawatt on any terms available, which is why it signed the $1.25 billion a month contract described in the cold open.³⁶⁶

Where the evidence cuts against me, plainly: at retail the labs are moving the other way, to a $20 seat plus per-token API rates and to token metering in Codex.⁸⁴ I cannot point to a lab that moved a retail product from a meter to capacity this year. The claim I can defend is that the labs are not the only place an enterprise can buy compute, and that an enterprise which buys the machine is no longer inside anyone else's rationing scheme.

#8. The wholesale example cancels

The objection: your best flat contract, $1.25 billion a month, can be cancelled on 90 days' notice.

Correct. The contract runs through May 2029 at $1.25 billion a month, and either side may terminate on 90 days' notice.³⁶⁶ Priced properly, that is a monthly rental wearing a multi-year headline.

Two descriptions of the same deal circulate and both are sourced. The SEC filing is SpaceX's, and the June 3 amendment covers Colossus and Colossus II at approximately 325,000 GPUs. The trade press reports the counterparty as xAI and sizes the compute at the 300 MW and roughly 220,000 GPUs of Colossus 1 alone, which is what Anthropic's own May announcement described. The larger figure is the amended contract. The smaller one is the first cluster inside it.

The pattern runs into NVIDIA's own book too. Its 10-Q discloses $36 billion of AI cloud service commitments that the clouds can unilaterally stop providing to NVIDIA and sell to third parties.⁵²

So I will keep the wholesale examples where they belong. They are supporting evidence, not the case. The case rests on end users signing terms they cannot walk away from, and Mistral's European Compute Unit contract is the clean instance of that.

#9. Usage share

The objection: 88% of enterprise API usage goes to the big three, and open share fell from 19% to 11%.

Both figures are Menlo's and both run against me. Menlo measured $37 billion of enterprise generative AI spend in 2025. It found that Anthropic, OpenAI and Google together account for 88% of enterprise LLM API usage, with open-source share at about 11%, down from 19% in 2024.¹⁷⁵

That measurement was published on December 9, 2025. Everything that has moved since runs the other way: Kimi K3 given away free, Luna cut 80% three days later, and an intelligence-index gap that closed from 13 points to 6 in twelve months.¹⁵⁰

Share today is not my claim. Direction is. The buyer's own accounting is not settled either: 59% of organizations say wasted AI software spend increased in the last year, and only 31% report accurate visibility into AI spend.³⁶⁷

#10. The GPU-hour is back

The objection: the first compute futures settle on H100 rent per hour and never mention a token. The market is standardizing the unit you left.

Read the contract. Those futures track hourly H100 and B200 rental cost, and the word token does not appear in the release.³⁴⁷ The objection is right about what the market standardized.

A futures market forms on the unit people actually contract in. That unit is a month of a machine.

The token stays the unit of output of the factory. It is what the machine makes and what the enterprise measures its work in. A futures market on machine-months and a factory that makes tokens are the same picture seen from two sides. The buyer stops paying per token and starts paying for the machine that makes them.

#11. NVIDIA is the labs' biggest financier

The objection: NVIDIA is not building an army against the labs. It is their largest backer.

The money is real and it points at the labs. The $105 billion of Ohio guarantees is the largest number NVIDIA has ever put behind a customer.

Note the form of it. That is a guarantee on land, power and shell that pays out only if a tenant defaults, with the tenant indemnifying the guarantor, not a check written to a model lab. It sits against a concentration the same filing discloses: five direct customers accounted for 22%, 14%, 13%, 11% and 10% of accounts receivable as of July 26, 2026.⁵² Seventy percent of the receivable balance sits in five names.

Which is why the July program looked the way it did. NVIDIA offered neoclouds a rental backstop for half of rental revenue above an agreed cost threshold. It also insisted on vetting its partners' end-customers, and mandated that cloud capacity be distributed across multiple smaller AI firms rather than concentrated with single large buyers.³⁶⁴ Reporting says NVIDIA paused specific deals weeks later after cloud-provider objections and internal concern that the terms could draw regulatory scrutiny. NVIDIA's spokesperson says the model introduced in July "is still in place and continues to evolve due to high demand".³⁶⁸

Eleven objections in, several of them landed. Here is what survived them, and what I had to give up along the way.

---

#16. What I changed, and what changes next

I am a tech guy at heart. I did not start this piece with a theory about pricing. I started with a rack, a power bill, and a customer who wanted to know what a year of running it would cost.

That question turned out to be answerable. The token question was not. You can price a machine. Pricing a year of tokens means knowing how many tokens the work will burn, and nobody knows that before the work runs.

So I changed what I was arguing, and I want to be precise about how much changed. Eighteen months ago I stopped thinking about GPU hours and put the business on monetizing AI compute by the token. I still believe that. The token stays the unit of output of the factory. A machine that makes tokens is measured in tokens, and I am not going to pretend otherwise.

What moves is the invoice, not the meter. The buyer stops paying per token and starts paying for the machine that makes them. The token count keeps running underneath, the way a kilowatt-hour keeps running underneath a demand charge. It tells both parties what the work cost. It stops being the thing that sets the bill.

#What the evidence changed in me

I went in believing the wholesale layer of AI was my strongest proof. It is not proof. It is a signal, and I have to say so.

The invoice this piece opened with is a flat capacity contract at the largest scale anybody has filed with the SEC.¹ It is also a contract between a model lab and its compute supplier.

The buyer in that deal is not an enterprise. The buyer is a company that will then sell you a meter. The parties who understand compute best buy it flat and resell it by the unit. That is a fact about them. It is not yet a fact about the enterprise, and the case for the enterprise has to stand on the enterprise's own evidence.

The second thing that changed is that I stopped treating the meter as a mistake. Altman described the future as intelligence sold as a utility bought from OpenAI on a meter.⁵⁷ His words: "We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for." That is a coherent business. A meter rations supply that is short, and it collects the training bill while the supply stays short. Anyone in that position would price the same way.

The third thing is that NVIDIA does not price the way the labs price. Its whole software price list is written in GPUs and years, five years for the price of four, and there is not a token anywhere on it.³² The longer you commit, the less each year costs. That is the opposite direction from every metered product in this piece.

#The number a buyer can hold

Capacity pricing only works if somebody can put a defensible cost on a year of the machine. That number exists in public.

The InferenceX operator-ownership figures already reconciled in this piece are the ones I would hand a CFO. A Vera Rubin NVL72 rack costs about $2.25 million to hold for a year, a GB300 about $1.49 million, a GB200 about $1.16 million.¹³⁶

I am not presenting that as my cost or as yours. It is a third-party estimate, and the model behind it is not open. What matters is the shape of it. A buyer can sign a multi-year commitment against a number of that kind. No buyer can sign against a token forecast, because the token forecast depends on how the work is written, which nobody knows on the day the contract is signed.

#What SAVRN does

SAVRN builds the factory, runs NVIDIA's open stack on it, and sells the capacity, so the customer keeps the spread between what the machine costs to hold and what the work on it is worth. We deploy NVIDIA's latest GPUs at time of delivery, and we do not bill the enterprise again for using the machine it already paid for.

That is the whole of it. The sources page carries every row behind every number in this piece, and I would rather you argue with the ledger than with me.

#What changes next

The next generation ships, then the one after that. That cadence is the reason capacity pricing becomes possible, and it is also the trap in it.

Note the direction those numbers run. The newer rack costs more per GPU-hour, not less, and it makes that back on throughput rather than on price.¹³⁶ Anyone writing a multi-year capacity deal has to price that hour as it is, not as they wish it were, and has to say what happens when the successor lands mid-term. A capacity contract that assumes the hour gets cheaper on its own is a bad contract.

#The charge in front of the meter

In 1892 British engineer John Hopkinson differentiated fixed and variable costs in electricity rates, and this two-part rate was promoted in the U.S. by Samuel Insull.⁸

Insull did not take the meter off the wall. He put a capacity charge in front of it and drove the unit price down until almost nobody thought about the unit. The meter kept spinning. It stopped being the thing anyone bought.

---

#Frequently asked questions

What is an unmetered rack?

It is a capacity contract. The enterprise pays for a machine for a fixed term and decides what the machine makes, instead of paying per token produced. The unit is the one NVIDIA itself uses: AI Enterprise is licensed per GPU at $4,500 a year, and the annual software bill on a 72-GPU rack does not move whether it serves one token or a trillion.³² The legal shape is a tolling agreement, in which the buyer pays a capacity payment for dispatch rights over the asset.²⁴

Is it really unlimited tokens?

No, and any seller who says so is selling you a number they cannot hold. Unmetered means the meter is not the pricing instrument. What replaces it is a written envelope: model class, numerical precision, tokens per second per concurrent user, and warranted availability. The envelope matters because the same rack produces $0.079 per million tokens in throughput mode and $4.18 per million at 350 tokens per second per user.¹³⁶ Context length moves it again, by up to 44x from short to long context.¹⁴¹

What does a Vera Rubin rack cost to run for a year?

Three answers, all sourced, and they differ by what is inside them. SemiAnalysis InferenceX puts operator-ownership total cost of ownership at $3.57 per GPU-hour, which is about $2.25 million per 72-GPU rack-year.¹³⁶ Morgan Stanley estimates the hardware alone at $7,803,148 per rack, with $2,001,600 of that in memory.¹⁴² The Anthropic-Nscale contract prices a rack at about $3.68 million a year, which includes the provider's margin.¹⁰¹

How many tokens does it make?

Between about 16 and 50 trillion output tokens a year in throughput mode, depending on load factor. The measured anchors are MLPerf Inference v6.0 on a 72-GPU GB300 NVL72: 575,580 tokens per second on DeepSeek-R1 server and 1,096,770 on gpt-oss-120B.¹³⁹ Nobody outside NVIDIA and its direct customers has published a Vera Rubin throughput number, so the generational step is an assumption. Run the same rack at agentic interactivity and output falls to roughly 0.2 to 0.5 trillion tokens a year.¹⁴⁰

Why do OpenAI and Anthropic charge per token?

Because they are short, and because the meter collects the training bill. Sam Altman said in March 2026 that every model provider's business "is going to look like selling tokens" bought "on a meter".⁵⁷ Anthropic's own rate-limit page says its numbers "represent maximum allowed usage, not guaranteed minimums".⁷ The economics explain the rest: at most about 30 cents of every Anthropic revenue dollar actually serves a token somebody paid for, with the balance going to training and free users.⁸⁰

Does NVIDIA charge per token?

Only where it runs the model on its own machines for you. The hosted NIM catalog at build.nvidia.com carries per-token prices and NVIDIA describes that catalog as a trial experience.⁴⁴ The moment the machine is yours, the price reverts to per GPU per year. NVIDIA does tell telco operators to move from GPU hours to token metering, with a worked example putting an H100 at about $157,680 a year on a meter.⁵⁵ That advice is addressed to the factory owner, not to NVIDIA's own price list.

Are open-weight models good enough?

For most calls, yes. For the hardest ones, no. On the Artificial Analysis Intelligence Index the best open model scores 60 against 66 for the best proprietary one, down from a 13-point gap a year earlier.¹⁵⁰ On the Vals SWE-bench Verified run, DeepSeek V4 Pro reaches 96.40% against Claude Opus 5 at 97.00%.¹⁵⁸ Long-horizon agentic work is where the frontier still wins outright. LangChain found that routing 7% of calls to a frontier model cut cost 74% and accuracy 6 points.¹⁶⁹

What is a tolling agreement?

An instrument from power markets. The buyer pays a capacity payment for the right, not the obligation, to have the owner run the asset. "The toller pays a fee just as a farmer would pay a mill to grind his wheat into flour," and keeps paying it even when delivered output falls short.²⁴ A gas toll has four prices: capacity payment, variable operations charge, start-up charge, and a heat-rate adjustment against a guarantee.³⁵⁴ A rack toll needs the same four.

What happens when the next GPU generation ships?

The chip keeps earning. CoreWeave signed an A100 contract extending into 2029 on hardware that shipped in 2020, and Jensen Huang says the A100 fleet is mission-capable through 2029.³³¹ One-year H100 contract pricing rose almost 40% between October 2025 and March 2026.³³⁴ The building is the stranding risk, because a hall built for 200 kW racks cannot take the 600 kW racks arriving in 2027.³³⁹

Who should not buy one?

Anyone whose machine will sit idle. Cast AI measured average GPU utilization of 5% across tens of thousands of enterprise Kubernetes clusters, with AI-lab clusters excluded.³²⁵ At that load factor a buyer pays twenty times the price per token he was quoted. Reserved dedicated endpoints only beat on-demand above roughly 80% sustained utilization, while production AI workloads average under 50%.³⁶⁰ Capacity pays when one machine carries the whole book of work. It does not pay for a single team's pilot.

#Sources

Every fact in this piece traces to one of the 368 sources below, across 180 organizations. Each was opened and read during the evidence program. Click any number in the text to jump here, and the arrow to jump back.

368 sourcesacross 180 organizations

1. SpaceX Form S-1 (accession 0001628280-26-036936)sec.gov↑↑↑
2. SpaceX Form S-1/A (accession 0001628280-26-040364)sec.gov↑↑↑
3. How long is Anthropic's lease with SpaceX? Opinions varytechcrunch.com↑
4. Pricing - Claude Platform Docsplatform.claude.com↑↑↑↑↑↑↑
5. Service tiers (Claude Developer Platform)platform.claude.com↑↑↑↑↑↑↑
6. Supported Region and models for Provisioned Throughput - Amazon Bedrockdocs.aws.amazon.com↑
7. Rate limits (Claude Developer Platform)platform.claude.com↑↑↑↑
8. Competing Perspectives on Demand Charges, Hledik and Faruqui, Public Utilities Fortnightly, September 2016ourenergypolicy.org↑↑↑↑↑
9. Samuel Insull, Wikipediaen.wikipedia.org↑
10. AOL's flat rate is a good deal offered in an insensitive way, Deseret News, December 2, 1996deseret.com↑↑↑
11. AOL to boost unlimited fee, cut 500 jobs, Baltimore Sun, February 10, 1998baltimoresun.com↑↑
12. Internet pricing and the history of communications, Andrew Odlyzko, Computer Networks 36 (2001)www-users.cse.umn.edu↑↑↑↑↑
13. AOL statistics compilation, Henning Schulzrinne, Columbia University (citing NYT 1/17/97 and 2/9/98)cs.columbia.edu↑
14. AT&amp;T WorldNet vs. AOL: How Flat-Fee Access Disrupted Dial-Up Internet, Tedium, April 9, 2020tedium.co↑↑
15. Netflix, Wikipediaen.wikipedia.org↑↑
16. Blockbuster Inc. Form 8-K Exhibit 99.1, December 2004sec.gov↑
17. Comcast Kills Its 250GB Data Cap, Is Testing More Flexible Data Plans, TechCrunch, May 17, 2012techcrunch.com↑
18. Comcast ends data caps, updates home Internet pricing, The Desk, June 26, 2025thedesk.net↑↑
19. Paying Not to Go to the Gym, DellaVigna and Malmendier, American Economic Review, 2006aeaweb.org↑
20. Reserved Instances for Amazon EC2 overview, AWS documentationdocs.aws.amazon.com↑
21. AWS Savings Plans, Amazon Web Servicesaws.amazon.com↑
22. Amazon.com, Inc. Form 10-Q for the quarter ended June 30, 2026sec.gov↑
23. Capacity Markets, PJM Learning Centerlearn.pjm.com↑
24. Output Contracts May Be 'Leases' of the Power Plant, Norton Rose Fulbright Project Finance, April 2002projectfinance.law↑↑↑↑↑
25. 2027/2028 Base Residual Auction Report, PJM Interconnection, December 17, 2025pjm.com↑
26. PJM Capacity Auction Procures 138,318 MW of Generation Resources, PJM news release, July 14, 2026pjm.com↑↑
27. PJM 2027/2028 Capacity Auction Results, Enel North Americaenelnorthamerica.com↑
28. Mainframe Pricing: Endgame, DataKinetics, August 26, 2019dkl.com↑↑
29. Spotify, Wikipediaen.wikipedia.org↑
30. Flexera 2026 State of the Cloud Report: The convergence of cloud and value, March 18, 2026flexera.com↑↑
31. NVIDIA AI Enterprise Licensing Guide: Licensingdocs.nvidia.com↑
32. NVIDIA AI Enterprise Licensing Guide: Pricingdocs.nvidia.com↑↑↑↑↑↑
33. NVIDIA Launches Next-Generation Rubin AI Compute Platform at CES 2026 (ServeTheHome)servethehome.com↑
34. ai-dynamo/dynamo (GitHub)github.com↑↑
35. NVIDIA Enters Production With Dynamoinvestor.nvidia.com↑
36. NVIDIA Nemotron Open Model Licensenvidia.com↑↑
37. NVIDIA Nemotron (product page FAQ)nvidia.com↑
38. NIM Offerings (NVIDIA NIM for LLMs docs)docs.nvidia.com↑
39. Nvidia (NVDA) Q2 2027 Earnings Call Transcript, The Motley Fool, Aug 31, 2026fool.com↑↑↑↑↑↑↑↑↑↑
40. NVIDIA Announces Financial Results for Second Quarter Fiscal 2027nvidianews.nvidia.com↑
41. NVIDIA CFO Commentary on Second Quarter Fiscal 2027 Resultssec.gov↑↑↑
42. NVIDIA NIM Free Plan (CostBench)costbench.com↑
43. Nvidia NIM API Pricing 2026: All Models &amp; Costs (APICents)apicents.com↑
44. NVIDIA NIM FAQforums.developer.nvidia.com↑↑
45. NVIDIA DGX Cloud Lepton documentation (current)docs.nvidia.com↑
46. NVIDIA DGX Cloud Lepton product pagenvidia.com↑
47. NVIDIA Announces DGX Cloud Lepton (NVIDIA Newsroom, May 18, 2025)nvidianews.nvidia.com↑
48. NVIDIA Unlocks AI Compute at Scale (NVIDIA Blog, July 1, 2026)blogs.nvidia.com↑
49. Yahoo Finance (citing WSJ): Nvidia pauses AI cloud revenue-sharing deals over antitrust concerns (Aug 28, 2026)finance.yahoo.com↑↑
50. Nvidia pauses revenue-sharing deals with AI cloud companies, WSJ reportswkzo.com↑↑
51. Nvidia pauses revenue-sharing deals with AI cloud companies, WSJ reports (Reuters via MarketScreener)marketscreener.com↑
52. NVIDIA Form 10-Q for the quarter ended July 26, 2026 (MD&amp;A)sec.gov↑↑↑↑↑↑↑↑↑↑
53. NVIDIA DSX Revenue Per Gigawatt Shapes New Financing Model (Futurum Group)futurumgroup.com↑
54. NVIDIA Form 10-K for fiscal year ended January 25, 2026s201.q4cdn.com↑↑↑↑↑↑↑
55. Building Token-Metered AI Services on Telco AI Factories (NVIDIA Technical Blog, May 21, 2026)developer.nvidia.com↑↑↑↑↑
56. Agents, inference and the new token economics: Nvidia pitches the AI future (RCR Wireless)rcrwireless.com↑
57. Business Insider: Sam Altman says AI will be a utility like electricity or waterbusinessinsider.com↑↑↑↑
58. Gizmodo: Sam Altman Says Intelligence Will Be a Utilitygizmodo.com↑
59. CNBC: OpenAI's data center pivot underscores Wall Street spending concerns ahead of IPOcnbc.com↑
60. Batch processing - Claude Platform Docsplatform.claude.com↑↑
61. Fast mode (research preview) - Claude Platform Docsplatform.claude.com↑↑
62. GPT-5.6 Sol model card | OpenAI APIdevelopers.openai.com↑
63. API Pricing | OpenAI Developersdevelopers.openai.com↑↑↑↑↑↑
64. Flex processing | OpenAI APIdevelopers.openai.com↑
65. OpenAI API pricing in 2026: every model after the July price cuts | CloudZerocloudzero.com↑↑
66. OpenAI cuts GPT-5.6 Sol API pricing 20-33%, Enterprise DNAenterprisedna.co↑
67. GPT-5.5 Pricing | Apidogapidog.com↑
68. Higher usage limits for Claude and a compute deal with SpaceX (Anthropic)anthropic.com↑↑
69. Anthropic admits Claude Code quotas running out too fast | The Registertheregister.com↑
70. Anthropic reinstates OpenClaw and third-party agent usage on Claude subscriptions, with a catch | VentureBeatventurebeat.com↑
71. Claude Code +50% Weekly Limits Promo | AI Catchupaicatchup.com↑
72. Anthropic is cutting Claude Code's current weekly limits by 17% | BleepingComputerbleepingcomputer.com↑↑
73. Use the Claude Agent SDK with your Claude plan | Anthropic Help Centersupport.claude.com↑↑
74. What is the Max plan? | Anthropic Help Centersupport.claude.com↑
75. Scale Tier for API Customers | OpenAIopenai.com↑↑↑
76. Reserved Tier for API Customers | OpenAIopenai.com↑
77. CNBC: OpenAI announces new Guaranteed Capacity offering for customers to secure computecnbc.com↑↑
78. SemiAnalysis, AI Value Capture: The Shift to Modelnewsletter.semianalysis.com↑↑↑
79. Anthropic Could Hit $1 Billion Quarterly Profit, API Model's 80%+ Gross Margin (BigGo, Jul 8, 2026, citing SemiAnalysis)finance.biggo.com↑
80. Anthropic's gross margin is the most important number in tech (PitchBook via Yahoo Finance)finance.yahoo.com↑↑↑
81. Can AI companies become profitable? (Epoch AI)epoch.ai↑
82. OpenAI's Gross Margin Is Going the Wrong Way (Control Plane, Jun 18, 2026)controlplane.news↑↑
83. OpenAI compute margins hit 70% (Fortune, Dec 21, 2025)fortune.com↑
84. Understanding the New Economics of AI Compute Markets (BCG)bcg.com↑↑↑↑↑↑↑
85. Microsoft Official Blog: Microsoft, NVIDIA and Anthropic announce strategic partnerships (Nov 18, 2025)blogs.microsoft.com↑↑
86. Microsoft Azure ND GB300 Sets New Records: 1.1 Million Tokens per Second for Inferencetest-news.aibase.com↑
87. Governor Morrisey's AI Strategy Positions West Virginia for Reported $45 Billion Agreementgovernor.wv.gov↑↑
88. Anthropic signs $35 billion Lambda cloud deal: WSJfinance.yahoo.com↑
89. Anthropic's $35B Lambda Deal Connects Nvidia to Hut 8's Texas AI Data Centertheenergymag.com↑
90. TeraWulf Form 8-K, July 6, 2026investors.terawulf.com↑↑
91. TeraWulf Announces Anthropic Lease at Justified Data Campusinvestors.terawulf.com↑
92. Riot Platforms Reports Second Quarter 2026 Financial Resultsriotplatforms.com↑
93. Riot Platforms strikes deal with Anthropic as bitcoin miners shift focus to AI infrastructurecnbc.com↑
94. Microsoft Form 10-Q (quarter ended Sep 30, 2025), Subsequent Eventssec.gov↑
95. CNBC: Nvidia backing $105 billion in financing for OpenAI data center in Ohiocnbc.com↑↑
96. NVIDIA Form 8-K, August 17, 2026 (Item 1.01 Residual Value Guaranties)sec.gov↑↑↑↑
97. Fortune: OpenAI data center deal with Nvidia comes in $145 billion lower than reportedfortune.com↑
98. What the Google/Broadcom/Anthropic financing tells us (View from the Peak)viewfromthepeak.substack.com↑
99. KBW keeps 'Outperform' on TeraWulffinance.yahoo.com↑
100. Anthropic's Compute Trap: From Compute Risk to Credit Riskdecodingdiscontinuity.com↑
101. Twenty Cents to Twenty-Five Dollars (SAVRN)savrn.com↑↑↑↑↑
102. Anthropic and Amazon expand collaboration for up to 5 gigawatts of new computeanthropic.com↑↑↑
103. Broadcom Inc. Form 8-K, Item 8.01, April 6, 2026sec.gov↑
104. Will financing bottleneck AI compute? (Epoch AI)epochai.substack.com↑↑
105. Dell Technologies Delivers Second Quarter Fiscal 2027 Financial Results (press release copy, StockTitan)stocktitan.net↑
106. Dell Q2 FY27 slides: AI server backlog hits $95B, revenue up 58% (Investing.com)investing.com↑
107. Hewlett Packard Enterprise Fiscal 2026 Second Quarter Earnings transcript (HPE IR)investors.hpe.com↑↑↑
108. NVIDIA CFO Commentary on Second Quarter Fiscal 2027 Resultss201.q4cdn.com↑
109. NVIDIA RTX PRO Servers Speed Trillion-Dollar Enterprise IT Industry Transition to AI Factories (NVIDIA Newsroom)nvidianews.nvidia.com↑
110. Industry Leaders Transform Enterprise Data Centers for the AI Era With NVIDIA RTX PRO Servers (NVIDIA Newsroom)nvidianews.nvidia.com↑
111. Lilly Deploys World's Largest, Most Powerful AI Factory for Drug Discovery Using NVIDIA Blackwell-Based DGX SuperPOD (NVIDIA Blog, Oct 28 2025)blogs.nvidia.com↑
112. Texas A&amp;M System Triples Supercomputing Capacity (Texas A&amp;M Stories)stories.tamu.edu↑
113. Denmark's first AI supercomputer is now operational (Novo Nordisk Foundation, Oct 23 2024)novonordiskfonden.dk↑
114. 2026: Inside Isambard-AI (University of Bristol)bristol.ac.uk↑
115. HUMAIN Expands Strategic Partnership with NVIDIA... (PR Newswire, Nov 19 2025)prnewswire.com↑
116. G42 Provides Update on Construction of Stargate UAE AI Infrastructure Cluster (PR Newswire)prnewswire.com↑
117. Mistral AI wants to build 1 gigawatt of European compute by 2030, and lock in customers now (VentureBeat)venturebeat.com↑↑↑
118. Akamai Discloses Technical Details on AI Cluster Deal (Akamai newsroom)akamai.com↑
119. AAI 2026: Open Telco AI Optimizes Models, Infrastructure and Token Costs (AMD Newsroom)newsroom.amd.com↑
120. AT&amp;T Turns to Open-Source AI to Cap Spending on Anthropic and OpenAI (BigGo Finance, Aug 21 2026)finance.biggo.com↑
121. Provisioned throughput for Foundry Models - Microsoft Foundry (Microsoft Learn, updated 2026-07-15)learn.microsoft.com↑↑↑
122. Increase model invocation capacity with Provisioned Throughput in Amazon Bedrock (AWS docs)docs.aws.amazon.com↑
123. Microsoft Fiscal Year 2026 Fourth Quarter Earnings Conference Call (transcript)microsoft.com↑↑
124. Amazon (AMZN) Q2 2026 Earnings Call Transcript (The Globe and Mail)theglobeandmail.com↑
125. Gartner Forecasts Global AI Spending to Hit $2.6 Trillion in 2026 (CIOL)ciol.com↑
126. AI Demand Surges as Billions in Compute Remain Locked (Data Center Knowledge, May 5 2026)datacenterknowledge.com↑
127. From Tokenmaxxing to Tokenminimizing: Meta Moves to Cap Employee AI Usage (The State of AI, June 12 2026)thestateofai.com↑↑
128. The Great Cloud Repatriation of 2026: Why 93% of Enterprises Are Pulling AI Workloads Home (Practical Logix)practicallogix.com↑
129. AI Infrastructure Spending Holds Near $90 Billion in Q1 2026... 2026 Forecast Raised to $497 Billion (IDC blog)idc.com↑
130. NVIDIA blog: Gigawatt AI factories, OCP, Vera Rubin (Oct 13, 2025)blogs.nvidia.com↑
131. NVIDIA Investor Presentation Q2 FY27 (August 2026)s201.q4cdn.com↑
132. SemiAnalysis: Vera Rubin Extreme Co-Design (Feb 25, 2026)newsletter.semianalysis.com↑
133. Spheron: NVIDIA Vera Rubin NVL72 specs guidespheron.network↑
134. The Register: Nvidia's Vera Rubin roadmap charts course for 600 kW rackstheregister.com↑
135. CoreWeave: First-Ever Measured Vera Rubin NVL72 Silicon Performance Statscoreweave.com↑
136. InferenceX: Vera Rubin NVL72 vs GB200 NVL72 Inference TCO &amp; Architecture Analysis (Jul 23, 2026)inferencex.semianalysis.com↑↑↑↑↑↑↑↑↑
137. InferenceX by SemiAnalysis blog indexinferencex.semianalysis.com↑
138. Vera Rubin NVL72 GPU cloud availability and cost per token (Spheron)spheron.network↑
139. Derived from Nebius MLPerf v6.0, NVIDIA inference page, and InferenceX Jul 23 analysisnebius.com↑↑↑↑↑
140. NVIDIA Vera Rubin NVL72 page (benchmark config) and InferenceX Jul 23 analysisnvidia.com↑↑
141. Is a compute crunch coming?, Epoch AIepoch.ai↑↑↑
142. Morgan Stanley estimate says single Vera Rubin rack costs $7.8M (Yahoo Finance)finance.yahoo.com↑↑
143. 35x Lower Token Cost with Blackwell - NVIDIA AI Inferencenvidia.com↑↑
144. DeepSeek R1 on GB300 NVL72 (SemiAnalysis InferenceX)inferencex.semianalysis.com↑↑
145. The profitability of AI factories: NVIDIA dominates with 77.6% margins (Cloud News)cloudnews.tech↑
146. InferenceX: GB200 NVL72 vs B200 on Kimi K2.5, 3.1x from Wide EP vLLM (Apr 23, 2026)inferencex.semianalysis.com↑
147. InferenceX compare: Kimi K2.5/K2.6/K2.7 GB200 NVL72 vs GB300 NVL72inferencex.semianalysis.com↑
148. InferenceX overview (SemiAnalysis)inferencex.semianalysis.com↑
149. NVIDIA Technical Blog: Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt (Aug 24, 2026)developer.nvidia.com↑
150. Comparison of AI Models across Intelligence, Performance, and Price | Artificial Analysisartificialanalysis.ai↑↑↑↑
151. Comparison of Open Source AI Models | Artificial Analysisartificialanalysis.ai↑
152. Recent open weights model launches | Artificial Analysisartificialanalysis.ai↑↑↑↑
153. Open models lag state-of-the-art closed models by 4 months | Epoch AIepoch.ai↑↑
154. Open-weight models lag state-of-the-art by around 3 months on average | Epoch AIepoch.ai↑
155. Technical Performance | The 2025 AI Index Report | Stanford HAIhai.stanford.edu↑
156. Technical Performance | The 2026 AI Index Report | Stanford HAIhai.stanford.edu↑
157. Text Arena Leaderboard | arena.aiarena.ai↑
158. SWE-bench Verified leaderboard (Vals AI), updated September 1, 2026vals.ai↑↑↑
159. NVIDIA-NeMo/Nemotron usage-cookbook, Nemotron-3-Super OpenScaffoldingResources READMEgithub.com↑↑
160. SWE-Bench Pro Leaderboard (Public Dataset) | Scalelabs.scale.com↑
161. Terminal-Bench v2.1 Benchmark Leaderboard (Artificial Analysis) and Nemotron 3 Ultra Hugging Face cardartificialanalysis.ai↑↑
162. Terminal-Bench 2.1 Leaderboard: GLM 5.3 &amp; New Models (2026) | CodingFleetcodingfleet.com↑
163. AgentFloor: How Far Up the tool use Ladder Can Small Open-Weight Models Go? (arXiv 2605.00334)arxiv.org↑
164. METR-Horizon-v1.1 benchmark results (raw YAML)metr.org↑
165. OpenAI and Anthropic Cut AI Costs as Open-Weight Models Gain Ground (TechRepublic)techrepublic.com↑
166. Models &amp; Pricing | DeepSeek API Docsapi-docs.deepseek.com↑↑
167. Pricing | Together AItogether.ai↑
168. Kimi K3 pricing | Kimi Open Platformplatform.kimi.ai↑
169. How many of your agent's calls actually need a frontier model? - LangChain Bloglangchain.com↑↑↑
170. Route AI Agents Across Models with NVIDIA NeMo Switchyard | NVIDIA Technical Blogdeveloper.nvidia.com↑
171. AT&amp;T Slashes AI Costs by Adopting Model Routers and Open Source | PYMNTSpymnts.com↑↑
172. AT&amp;T Says It's Not Scared of the Token Apocalypse (Insider Monkey)insidermonkey.com↑
173. How We Cut Inference Spend by Right-Sizing Our Models (Salesforce)salesforce.com↑
174. A technical report on Composer 2 (Cursor)cursor.com↑
175. 2025: The State of Generative AI in the Enterprise | Menlo Venturesmenlovc.com↑↑↑↑
176. 2025 Mid-Year LLM Market Update | Menlo Venturesmenlovc.com↑
177. Leaders, gainers and unexpected winners in the Enterprise AI arms race | Andreessen Horowitza16z.com↑
178. August 2026 Ramp AI Index: Cracks in the AI thesisramp.com↑
179. The End of NowLLM: Closing an In-House Era (ServiceNow Community developer blog)servicenow.com↑
180. Anthropic says it hit a $30 billion revenue run rate after 'crazy' 80x growth (VentureBeat)venturebeat.com↑
181. DeepSeek overtakes Google on volume, cost per token falls 13.6% (Vercel)vercel.com↑
182. Open-weight models surge to 29% of volume, price per token flattens (Vercel)vercel.com↑
183. LLM Rankings | OpenRouteropenrouter.ai↑
184. Ramp's latest data on China vs. the American AI Labsramp.com↑
185. nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 model card (Hugging Face) and NVIDIA Technical Blog, June 4, 2026huggingface.co↑↑↑↑
186. Olmo 3: Charting a path through the model flow to lead open-source AI (Ai2)allenai.org↑↑
187. ETH Zurich press release: Apertus, a fully open, transparent, multilingual language modelethz.ch↑↑
188. Introducing Marin: An Open Lab for Building Foundation Modelsmarin.community↑
189. LLM360/K2-V2 model cardhuggingface.co↑
190. HuggingFaceTB/SmolLM3-3B model cardhuggingface.co↑
191. openai/gpt-oss-120b model cardhuggingface.co↑↑
192. meta-llama/Llama-4-Scout-17B-16E-Instruct model cardhuggingface.co↑↑
193. NVIDIA Technical Blog: Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Executiondeveloper.nvidia.com↑
194. GitHub NVIDIA-NeMo/Nemotron: Developer Asset Hub READMEgithub.com↑
195. Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoningarxiv.org↑
196. nvidia/Nemotron-CC-v2 dataset cardhuggingface.co↑
197. nvidia/Nemotron-CC-v2.1 dataset cardhuggingface.co↑
198. nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 model cardhuggingface.co↑
199. Nemotron Post-Training v3 collectionhuggingface.co↑
200. nvidia/Nemotron-Post-Training-Dataset-v2 dataset cardhuggingface.co↑
201. NVIDIA Blog: Nemotron 3.5 Lightning and NeMo Switchyardblogs.nvidia.com↑↑↑
202. allenai/dolma3 dataset cardhuggingface.co↑
203. allenai/Olmo-3.1-32B-Think model cardhuggingface.co↑↑
204. OpenMDW License Agreement, version 1.1openmdw.ai↑
205. Linux Foundation Releases OpenMDW-1.1; NVIDIA Adopts OpenMDWlinuxfoundation.org↑
206. What OpenMDW 1.1 guarantees an enterprise that an Apache license on model weights can't (diginomica)diginomica.com↑
207. EU AI Act Article 53: Obligations for Providers of General-Purpose AI Modelsartificialintelligenceact.eu↑
208. WilmerHale: European Commission Releases Mandatory Template for Public Disclosure of AI Training Datawilmerhale.com↑
209. EU AI Act Article 53 Training-Data Summaries: Who Filled the Template (Pebblous)blog.pebblous.ai↑
210. Open Weights: not quite what you've been told (OSI)opensource.org↑
211. Nvidia Commits $26 Billion to Open-Weight AI, Ships Nemotron 3 Super (implicator.ai)implicator.ai↑
212. NVIDIA Launches Nemotron Coalition of Leading Global AI Labs to Advance Open Frontier Modelsnvidianews.nvidia.com↑
213. Nvidia building 1-trillion-parameter Nemotron 4 to rival open AI models, The Information reports (Reuters)finance.yahoo.com↑
214. Open Weights and American AI Leadership (Jensen Huang, NVIDIA)images.nvidia.com↑↑
215. Nvidia Rallied the Industry Behind Open Weights. Then OpenAI Joined Anyway. (Security Boulevard)securityboulevard.com↑
216. NVIDIA Debuts Nemotron 3 Family of Open Models (NVIDIA Newsroom)nvidianews.nvidia.com↑
217. OpenAI and Broadcom reveal Jalapeno, first AI chip in partnership (CNBC)cnbc.com↑
218. VentureBeat: OpenAI unveils first custom AI inference chip, Jalapeño, with Broadcomventurebeat.com↑
219. AMD and OpenAI announce strategic partnership to deploy 6 GW of AMD GPUsamd.com↑
220. Anthropic Enters The AI Chip Race With In-House Chip Team (Forbes)forbes.com↑↑
221. Microsoft, NVIDIA and Anthropic announce strategic partnerships (Anthropic)anthropic.com↑
222. Anthropic continues compute-gobbling streak in $45B deal with Nscale (TechCrunch)techcrunch.com↑
223. Anthropic: Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute (Apr 6, 2026)anthropic.com↑↑
224. Poolside Strikes $6 Billion Licensing Deal with Nvidia (Newcomer)newcomer.co↑
225. Jensen Huang's First X Post: 150+ Companies Sign Open-Weights Letter, Anthropic Doesn't (NYU Shanghai RITS)rits.shanghai.nyu.edu↑
226. Anthropic's Dario Amodei responds (TechCrunch)techcrunch.com↑
227. NVIDIA DSX Gives Infrastructure Builders the Playbook for AI Factoriesnvidianews.nvidia.com↑
228. Nvidia's version of OpenClaw could solve its biggest problem: security (TechCrunch)techcrunch.com↑
229. Nvidia (NVDA) Q2 2027 Earnings Call Transcript (Yahoo Finance)finance.yahoo.com↑↑↑
230. NVIDIA Form 10-Q for quarter ended July 26, 2026, Note 10 Commitments and Contingenciessec.gov↑↑
231. SB Energy / OpenAI / NVIDIA press release, exhibit to 8-Ksec.gov↑
232. NVIDIA Form 8-K, August 17, 2026sec.gov↑
233. The Motley Fool: Jensen Huang Said Nvidia's $30 Billion OpenAI Investment Might Be the Last. Nvidia Just Guaranteed Up to $105 Billion of OpenAI's Leases. (Aug 25, 2026)fool.com↑↑
234. NVIDIA Newsroom: OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems (Sep 22, 2025)nvidianews.nvidia.com↑
235. Reuters via Yahoo Finance: Nvidia close to investing $30 billion in OpenAI's mega funding round, source says (Feb 19, 2026)finance.yahoo.com↑
236. AMD IR: AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs (Jul 22, 2026)ir.amd.com↑
237. Stop Comparing LLM Agents Without Disclosing the Harness (Zhang, Wang, Ge, Xu, Hamm, Reddy; arXiv 2605.23950)arxiv.org↑↑↑
238. Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows (arXiv 2605.27922)arxiv.org↑
239. Scaffold Effects on GAIA: A Controlled Comparison (Jason Starace, arXiv 2606.08529)arxiv.org↑
240. HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry (arXiv 2606.14249)arxiv.org↑
241. Meta-Harness: 76.4% on Terminal-Bench 2.0 (Claude Opus 4.6), Stanford IRIS Lab artifact READMEgithub.com↑
242. The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable (Vats, Golev; arXiv 2607.22585)arxiv.org↑↑
243. Agentic AI Cost Model: NVIDIA Blackwell, Dynamo &amp; Tokens - NVIDIA Perspectivesperspectives.nvidia.com↑
244. Enterprise Software Leaders Build AI Agents With NVIDIA (NVIDIA Newsroom)nvidianews.nvidia.com↑
245. NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness (NVIDIA Blog) and LangChain NemoClaw Deep Agents Blueprint postblogs.nvidia.com↑
246. Run Autonomous, Self-Evolving Agents More Safely With NVIDIA OpenShell (NVIDIA Technical Blog)developer.nvidia.com↑
247. NVIDIA Developer Blog: Run autonomous self-evolving agents more safely with NVIDIA OpenShelldeveloper.nvidia.com↑↑
248. Nvidia's version of OpenClaw could solve its biggest problem: security (TechCrunch, Rebecca Szkutak)techcrunch.com↑
249. GitHubgithub.com↑
250. NVIDIA/NemoClaw - GitHubgithub.com↑↑
251. NVIDIA-NeMo/Switchyard - GitHubgithub.com↑↑
252. GitHubgithub.com↑
253. NVIDIA DSX OS Delivers Open, Modular Software for Operating AI Factories at Scale - NVIDIA Technical Blogdeveloper.nvidia.com↑↑↑
254. Hugging Facehuggingface.co↑
255. NVIDIA Docsdocs.nvidia.com↑
256. Nvidia's Switchyard router reshuffles AI models mid-task, cutting task costs to a third in its own tests (VentureBeat, Sean Michael Kerner)venturebeat.com↑
257. NVIDIA Ignites the Next Industrial Revolution in Knowledge Work With Open Agent Development Platform (NVIDIA Newsroom)nvidianews.nvidia.com↑
258. Pricing - Claudeclaude.com↑↑
259. Provisioned throughput billing and cost management - Microsoft Foundrylearn.microsoft.com↑↑
260. First impressions of Claude Cowork, Anthropic's general agent (Simon Willison)simonwillison.net↑
261. Customize Cowork with plugins (Claude by Anthropic)claude.com↑
262. Why a new AI tool hammered some software stocks this week (ABC News)abcnews.com↑
263. Anthropic's new AI tools deepen selloff in data analytics and software stocks, investors say (Reuters via Yahoo)finance.yahoo.com↑
264. Investors Panic Over Agentic AI: Claude Cowork plugins trigger a SaaS stock selloff (The Batch)deeplearning.ai↑
265. The SaaS Rout of 2026 Is Even Worse Than You Think (SaaStr)saastr.com↑
266. Anthropic CEO Dario Amodei warns some software companies will 'completely go bust' (Yahoo Finance)finance.yahoo.com↑
267. OpenAI launches a way for enterprises to build and manage AI agents (TechCrunch)techcrunch.com↑
268. Paul Smith to join Anthropic as Chief Commercial Officeranthropic.com↑
269. From Claude Code to Figma: Turning Production Code into Editable Figma Designsfigma.com↑
270. Figma, Inc. Form 8-K, Item 5.02, filed April 14 2026sec.gov↑
271. Anthropic CPO leaves Figma's board after reports he will offer a competing product (TechCrunch)techcrunch.com↑
272. Introducing Claude Design by Anthropic Labsanthropic.com↑
273. Scoop: How A Board Departure And Product Launch Sparked An Anthropic Accusation (Upstarts Media)upstartsmedia.com↑
274. Findell Issues Report and Letter to CEO and Board of Directors of Figmaprnewswire.com↑
275. Figma (FIG) quote, Yahoo Financefinance.yahoo.com↑
276. Why Figma Stock Just Sank (The Motley Fool)fool.com↑
277. Figma Announces First Quarter 2026 Financial Resultssec.gov↑
278. Figma, Inc. Form 10-Q for the quarter ended June 30 2026sec.gov↑↑↑
279. Anthropic and Salesforce Expand Strategic Partnership to Deliver Trusted AI for Regulated Industriessalesforce.com↑
280. Salesforce and Anthropic Announce Claudeforce: The #1 AI Meets the #1 AI CRMsalesforce.com↑↑
281. Salesforce Delivers Record Second Quarter Fiscal 2027 Results (press release)salesforce.com↑↑
282. Salesforce Q2 earnings beat; Claudeforce AI plugin launched with Anthropic (Yahoo Finance)finance.yahoo.com↑
283. Software stocks rebound as Anthropic partnerships ease AI disruption fearsinvesting.com↑
284. Harvey revenue, valuation and funding (Sacra)sacra.com↑
285. The world-leading AI company just threatened one of Australia's great tech successes, Atlassian (The Nightly)thenightly.com.au↑
286. OpenAI Privately Tells Developers Using Company's Tech That It Won't Compete With Them Beyond ChatGPT (Slashdot, relaying Humanloop)slashdot.org↑
287. AI Agents Don't Buy Seats, and That's Workday's $133 Stock Problem (24/7 Wall St.)247wallst.com↑
288. Salesforce's Outcome-Based Pricing: Strategic Masterstroke or Defensive Scramble Against AI Seat Compression? (CX Today)cxtoday.com↑
289. SaaS Applications 'Will Collapse' In The AI Agent Era: Microsoft CEO Satya Nadella (OfficeChai)officechai.com↑
290. The Build vs. Buy Shift: AI, Shadow IT, and the SaaS Replacement Era (Retool)retool.com↑
291. The Build-vs-Buy Shift: 32% of Enterprises Bet on Agentic Coding Tools (Yahoo Finance)finance.yahoo.com↑
292. MIT report: 95% of generative AI pilots at companies are failing (Fortune via Yahoo)finance.yahoo.com↑
293. Klarna-Salesforce-Workday Partnership Called Off Amidst Major Gen-AI Overhaul (Salesforce Ben)salesforceben.com↑
294. Klarna CEO doubts that other companies will replace Salesforce with AI (TechCrunch)techcrunch.com↑
295. Is traditional SaaS behind us? The Graph + GenAI revolution (Neo4j blog, Chandra Rangan)neo4j.com↑
296. The 2026 State of AI Agents Report (Anthropic with Material)resources.anthropic.com↑
297. Workday (WDAY) Q2 2027 Earnings Call Transcript (Motley Fool)fool.com↑
298. ServiceNow Reports Second Quarter 2026 Financial Results (SEC 8-K exhibit)sec.gov↑
299. Anthropic's Claude triggered a trillion-dollar selloff. A new upgrade could make things worse (Fortune)fortune.com↑
300. Thomson Reuters CEO says plunge in software stocks driven by anxiety, not fundamentals (Globe and Mail)theglobeandmail.com↑
301. Slack API Terms of Serviceslack.com↑
302. Salesforce Locks Down Slack Data: Time to Review Your Slack API Terms (Hunton)hunton.com↑
303. BNP Paribas Deploys Internal 'LLM as a Service' Platform (Markets Media)marketsmedia.com↑
304. LLM Suite named 2025 'Innovation of the Year' by American Banker (JPMorganChase)jpmorganchase.com↑
305. Enterprise LLM Adoption Accelerates Amid Security and Cost Concerns (theCUBE Research on Kong report)thecuberesearch.com↑
306. Is my data used for model training? (Anthropic Privacy Center)privacy.claude.com↑
307. Anthropic unveils new rate limits to curb Claude Code power users - TechCrunchtechcrunch.com↑
308. OpenAI Codex pricing in 2026: plans, token costs, and usage limits - CloudZerocloudzero.com↑
309. The Register, Anthropic ejects bundled tokens from enterprise seat dealtheregister.com↑↑
310. Claude subscriptions get separate budgets for programmatic use, billed at full API prices, The Decoderthe-decoder.com↑
311. GitHub Copilot is moving to usage-based billing - The GitHub Bloggithub.blog↑
312. GitHub Copilot is moving to usage-based billing - community discussion #192948github.com↑
313. Hitechies, The AI subscription subsidy: real cost 2026hitechies.com↑↑
314. Claude Code Usage Limits and Pricing, Explained (August 2026), CC for Everyoneccforeveryone.com↑
315. Extra usage for paid Claude plans, Claude Help Centersupport.claude.com↑
316. Manage costs effectively, Claude Code docscode.claude.com↑↑
317. Claude Code Pricing In 2026, CloudZerocloudzero.com↑
318. Meta killed the employee AI token dashboard, Fortunefortune.com↑
319. AI: Meta steps back from AI 'tokenmaxxing', Michael Parekhmichaelparekh.substack.com↑
320. Nvidia's Jensen Huang thinks $1 trillion won't be enough to meet AI demand - Fortunefortune.com↑
321. Nvidia CEO Jensen Huang says he'd be deeply alarmed if a $500K developer spent less than $250K on AI tokens - The Decoderthe-decoder.com↑
322. NVIDIA GTC Taipei at COMPUTEX: Live Updates on What's Next in AI - NVIDIA Blogblogs.nvidia.com↑↑
323. AI Inference Utilization: NVIDIA Blackwell Dynamo Cost Curve - NVIDIA Perspectivesperspectives.nvidia.com↑
324. Nvidia's CEO Math - The more you buy, the more you savekeithfitz-gerald.com↑
325. 2026 State of Kubernetes Optimization Report - Cast AIcast.ai↑↑
326. DeepSeek V4 adoption insights, OpenRouter Blogopenrouter.ai↑
327. How Much Do AI Tokens Cost Businesses? 2026 Spending Benchmarks, Rampramp.com↑
328. NVIDIA Computex 2024 keynote blogblogs.nvidia.com↑
329. NVIDIA Rubin Platform AI Supercomputer press releasenvidianews.nvidia.com↑↑
330. CoreWeave Q2 2026 earnings call transcript (Motley Fool)fool.com↑
331. Yahoo Finance: Jensen Huang says Nvidia's 2020 A100 GPUs can stay mission-capable through 2029finance.yahoo.com↑↑
332. NVIDIA Q3 FY2026 earnings call transcript (Motley Fool)fool.com↑
333. Silicon Data H100 Rental Price Indexsilicondata.com↑
334. SemiAnalysis: The Great GPU Shortage, Rental Capacitynewsletter.semianalysis.com↑↑
335. Silicon Data: H100 Rental Price Over Time (2023-2025)silicondata.com↑
336. Silicon Data: H100 Rental Prices by Region (Q3 2026)silicondata.com↑
337. Hudson Labs: Amazon Extended Server Life Twice, Then Reversed Ithudson-labs.com↑
338. ModulEdge: Nvidia Vera Rubin data center impactmoduledge.com↑
339. NVIDIA Developer Blog: 800 V HVDC architecturedeveloper.nvidia.com↑↑
340. VRLA Tech NVIDIA GPU roadmap 2026-2030vrlatech.com↑
341. CNBC: Oracle is building yesterday's data centers with tomorrow's debtcnbc.com↑
342. AP via Spectrum News: Microsoft takes over a Texas AI data center expansion after OpenAI backs awayspectrumlocalnews.com↑
343. Oracle Corporation Form 10-K, fiscal year ended May 31, 2026sec.gov↑
344. heise online: S&amp;P downgrades Oracle to BBB-, only one notch above junk levelheise.de↑
345. CoreWeave Form 8-K (event September 9, 2025; filed September 15, 2025)sec.gov↑
346. NVIDIA Newsroom: $500B AI compute infrastructure financing platformsnvidianews.nvidia.com↑↑
347. CME Group and Silicon Data to Launch Compute Futures on October 5 (PR Newswire copy of CME release)prnewswire.com↑↑↑
348. CoreWeave, Inc. Form 424B4 prospectussec.gov↑
349. CoreWeave, Inc. Form 10-Q for quarter ended June 30, 2026sec.gov↑
350. CoreWeave Q1 2026 earnings call transcripts205.q4cdn.com↑
351. Nvidia GPU Debt Backstop Unleashes the AI Project Trinity: Capital, Offtake and Datacentersnewsletter.semianalysis.com↑↑
352. CoreWeave Closes $2.6 Billion Loan Facilityinvestors.coreweave.com↑
353. Luca Pedretti on LinkedIn: Mistral is not selling an AI PPAlinkedin.com↑
354. Prototype Natural Gas Tolling Agreement Term Sheet (PSE, WA UTC Docket 091618)apiproxy.utc.wa.gov↑↑
355. AI Token Futures Market: Commoditization of Compute and Derivatives Contract Design (arXiv 2603.21690)arxiv.org↑
356. Akash: H100 Rental Price in August 2026; Thunder Compute: AI GPU Rental Market Trends (Sep 1, 2026)akash.network↑
357. OpenAI cuts GPT-5.6 Luna and Terra prices by up to 80% (Yahoo Finance/CNBC syndication)finance.yahoo.com↑
358. WION: OpenAI just cut prices by 80% 3 days after China gave its best model away freewionews.com↑
359. CloudZero: Azure OpenAI pricingcloudzero.com↑
360. CloudZero: Together AI pricing (May 11, 2026)cloudzero.com↑↑
361. Fireworks AI pricingfireworks.ai↑
362. Baseten pricingbaseten.co↑
363. Lambda GPU Cloud pricing; Nebius priceslambda.ai↑
364. Techstrong.ai: NVIDIA Pauses $36 Billion Revenue-Sharing Financing Model Amid Antitrust Scrutinytechstrong.ai↑↑
365. Anthropic: Series H announcementanthropic.com↑
366. TechCrunch: Anthropic will pay xAI $1.25 billion per month for computetechcrunch.com↑↑
367. Flexera: When AI budgets balloon: what enterprises are learning in 2026flexera.com↑
368. Reuters via Yahoo Finance: Nvidia pauses revenue-sharing deals with AI cloud companies, WSJ reportsfinance.yahoo.com↑