AI Data Center Delays: What the 35 Percent Pipeline Really Means

AI data center delays are now front-page news. On October 18, the Financial Times ran an investigation titled “America’s $10tn AI infrastructure boom is starting to fray.” One line in it has been repeated everywhere since. In particular, Bernstein analysts estimate that only about 35% of the planned US data center pipeline is credible.^{1, 2}

I read that and my first reaction was that we have been saying this for a year. Our delay tracker has been counting stalled, cancelled and contested projects since August. As a result, I wanted to know if the number proves the point. The short answer is that it supports the point, although not the way most headlines use it.

The 35% is a credibility screen rather than a forecast that 65% of projects will be cancelled. Those are two different claims. That difference matters, in particular, to anyone putting money, land or a research program behind a promised opening date.

This article does three things. First, it traces what the 35% figure is and is not. Second, it walks through the physical and financial reasons projects slip, using primary records from grid operators, utilities, regulators and courts. Third, it lays out the method we use at SAVRN to grade a project one phase at a time, with a worked example.

Key figures

What the 35% number actually is

The attribution chain

Every public use of the 35% figure I could find traces back to the FT. For example, Techmeme indexed the FT piece on October 18 under the headline “RegLayer: 47+ large US data center projects were delayed, withdrawn, or blocked in 2026; Bernstein says ~35% of the planned US data center pipeline is credible.”² Similarly, Seeking Alpha’s summary repeats the same pair of numbers.³ However, I have not obtained Bernstein’s client report, and nobody quoting it in public seems to have either.

Because the report is private, that leaves open questions only it can answer:

• What was counted? Projects, planned IT megawatts, total facility megawatts, or a stage-weighted total.
• What universe? Announced only, or announced plus under construction plus existing.
• What horizon? Credible by 2028, by 2030, or ever.
• How were duplicates handled? For instance, one hyperscaler shopping three sites for one cluster can look like three projects.
• What does “not credible” mean? Not enough public evidence yet, likely to slip, likely to shrink, or likely to die.

One write-up put it well. The 35% is the share that could be executed credibly, rather than a prediction for each project, and the other 65% is not necessarily unbuildable.¹⁶

Why “65% are fake” is the wrong headline

If you turn 35% credible into 65% cancelled, you are adding a claim the source did not make. A project can fail a credibility screen today because its utility agreement is not signed yet. However, in six months it may have the agreement. Another project may pass the screen and still slip a year because a transformer is late.

As a result, I will not run a pie chart that says 35% built and 65% dead, because that chart would invent a measured outcome and a denominator that nobody has published.

The other counts do not add up, and should not

Several trackers now count trouble in the pipeline, although they measure different things.

One outlet that covered both RegLayer and Data Center Watch said it is “inappropriate to treat the quarterly counts as 120 unique failed developments.”¹⁶ I agree, because a project that was delayed in Q1 and approved in Q2 is not a failure. Similarly, a project that shows up in three trackers is still one project.

In particular, the Heatmap data is worth a closer look. Heatmap says about half of projects that met local backlash this year were ultimately cancelled, and that about $130 billion was cancelled in the three months ending September 30 alone.⁶ Its Q1 review found at least 20 projects cancelled after local pushback in the first quarter, worth more than $41.7 billion.⁶ That is a sharp rise across two quarters, although it is a measure of announced investment value rather than money spent.

Demand is not the problem

In fact, none of this says AI demand is fading. For example, Berkeley Lab estimated US data centers used 176 TWh in 2023 and modeled 325 to 580 TWh for 2028.¹⁴ Similarly, the IEA’s 2025 analysis said around one fifth of planned data center projects globally could face delays if grid bottlenecks are not addressed, and described four to eight years to build new transmission lines in advanced economies.¹⁷ Its 2026 follow-up keeps the central demand outlook close to the 2025 path while flagging stretched planning systems and financing sensitivity.¹⁸

Strong demand and pipeline attrition can both be true. When demand is strong, a finished building is worth more, while a crowded pipeline means more proposals fail. In addition, a cancelled site can be replaced by another site for the same customer, and the national total holds.

Therefore, the useful question is not “will AI data centers get built?” Instead, it is “will this phase get the right power, equipment, permits and money, in this location, by this date?”

Start with the megawatt being promised

Most confusion starts with the unit, since “a 1 GW campus” can mean seven different things.

Worked conversion: facility power vs IT power

Power usage effectiveness (PUE) is total facility energy divided by IT energy.

• For example, at a PUE of 1.20, 100 MW of IT load needs 120 MW of facility power.
• Conversely, 100 MW of total supply supports about 83.3 MW of IT load, before any design margin.

Now run a year. For instance, a 100 MW IT phase running at an average 80% of nameplate uses 700,800 MWh a year (100 × 0.8 × 8,760 hours). At PUE 1.20, the facility uses about 840,960 MWh, or 0.84 TWh.

As a result, two projects both marketed as “100 MW” can be a 20 MW difference in IT load depending on which meter the number came from. Any serious comparison names the meter boundary, the phase and the date.

Grid access is a sequence of permissions

ERCOT shows the gap between requests and capacity

On June 18, 2026, ERCOT said it was tracking more than 438,000 MW of large-load requests, with nearly 89% tied to data centers.⁷ Those are requests in a queue, rather than plants that will run.

ERCOT’s answer was “Batch Zero.” It groups qualified projects of 75 MW and larger into a single study so the grid operator can see total future demand and plan transmission once, instead of project by project.⁷ In addition, it added two paths: customers can bring onsite generation for some or all of their supply, or accept curtailment during local transmission constraints. ERCOT expected to tell Batch Zero applicants their classification in August 2026 and to finish a transmission plan in fall 2027, while Batch 1 applications are expected to open in summer 2027.⁷

Then the process tightened. First, an August 3 market notice pushed back the planned classification notice after the governor required verification.¹⁰ Second, a September 9 notice sent requests for information that asked for supporting documents and notarized attestations, generally due within ten business days, and made passing verification a condition of staying in the batch.⁸

That is the core lesson. A spot in a queue now has to be backed by evidence, since a June forecast is not an October approval.

PJM is building the same filter

Similarly, on October 9, 2026, FERC approved PJM’s new large-load registry, which tracks facilities of 50 MW or more: location, expected operating date, projected consumption, supply arrangement and backup generation.¹⁹ However, FERC rejected the rest of PJM’s package, including the Interim Resource Adequacy Service and the “Bring Your Own New Capacity” proposal, saying parts were outside its authority or not shown to be just and reasonable.¹⁹ Most of PJM’s proposed tariff changes were not meant to take effect until June 1, 2027, and PJM lists the registry tool as ready January 1, 2027.^{19, 11}

The two largest US grid regions are both building a registry of who is real. That is what a credibility screen looks like when a grid operator does it.

Questions to ask for each electrical phase

A nearby transmission line on a map is not available capacity, because voltage, substation configuration, competing loads, contingencies and protection all matter. Those need engineering evidence for the actual connection rather than a map radius.

Service quality belongs in the commercial model too, since 100 MW that can be curtailed on short notice is worth something different from 100 MW of firm service. Interruptible research jobs can live with it, while latency-sensitive inference and tightly synchronized training often cannot.

Utility tariffs move financial risk forward

For example, AEP Ohio’s data center tariff is one of the clearest public examples of how reserving power now creates obligations before revenue.⁹

• Study fee. $100,000 for requests of 100,000 kW and greater, while smaller requests pay $50,000 for 50 to 100 MW and $10,000 for 25 to 50 MW.
• Ramp. It can run no longer than four years, while minimum contract capacity steps through 50%, 65%, 80% and 90% in years one to four.
• Term. The ramp period plus eight years. As a result, the term can run up to 12 years.
• Minimum billing demand. Specifically, for contracts above 75,000 kW, at least 57,500 kW plus 100% of capacity above 75,000 kW, capped at 85% of contract capacity.
• Credit support. Customers without both an A-/A3 rating and audited cash above ten times the collateral requirement must post a guarantee or collateral equal to 50% of total minimum charges for the full term. Surety bonds are not accepted.

For example, using the published formula, a 100 MW contract has a minimum billing demand of 57.5 MW plus the 25 MW above 75 MW, or 82.5 MW. That is a capacity charge floor rather than a requirement to consume that power around the clock.

That matters because the party that books power can owe infrastructure costs before compute service starts. For instance, if the customer ramps slowly, the project carries land, shell and utility obligations against partial revenue. If a tenant can walk after a missed delivery date, the mismatch gets worse. A 12-year utility term may also outlast the first accelerator generation and a shorter customer contract, so refinancing or re-leasing stops being upside and becomes an assumption.

Collateral is not an expense on day one, although it ties up cash or guarantee capacity that could fund the next project. Instead, model it as a funding requirement with release conditions.

Procurement sets the real schedule

Electrical equipment is not one market, because each class has its own lead time.

These are contractor-reported planning ranges, not guaranteed dates. Instead, a real project uses its own specification and quote.

The words “secured equipment” need unpacking. First, a quote is not a factory slot. Second, a slot is not a binding order. Third, a binding order is not delivered equipment. Finally, delivered equipment still needs installation and testing. GE Vernova’s July 22, 2026 results separate 53 GW of gas equipment backlog from 63 GW of slot reservations.²⁰ The combined 116 GW is not a shipment schedule, and not all of it is data center demand. That is why “we have turbines” means little to a lender without the contract terms.

Cooling works the same way, because liquid cooling needs compatible cold plates, coolant distribution, water chemistry, heat rejection, controls and service procedures. Because of that, faster delivery of one part does not finish the system. Linesight’s June 2026 Americas review called electrical equipment a major constraint, while it described the mechanical market as more moderate.²¹ So not every category is equally scarce, and a plan should say which one is the critical path.

Timelines follow dependencies, not press releases

Virginia’s legislative audit staff (JLARC) found that building an individual data center building takes roughly 12 to 18 months.²² However, that interval does not include land, power, permits or long-lead equipment. Compare it with the IEA’s four to eight years for new transmission lines, and you can see the sequencing problem.¹⁷ In fact, the shell is often the fast part.

Illustration: say the building finishes in month 18, the transformer arrives in month 25, utility work finishes in month 31, and integrated tests need two more months. First service is month 33. Speeding up the shell by three months earns nothing unless it changes another dependency.

A phased plan helps only if the utility will authorize a smaller first load and the electrical design supports commissioning that phase on its own. It hurts if shared infrastructure gets mobilized over and over, if the first phase cannot carry operating overhead, or if the customer needs the full cluster before acceptance.

A credible schedule is a version-controlled dependency model with owners and change dates. Publish a base date and a stress date, and name the event that moves them.

Santa Clara: the network date matters

For example, Santa Clara’s October 2026 council record describes more than $450 million of electrical projects underway, with individual projects coming online starting in 2028 and major improvements scheduled for completion by 2030.²³ The city also says its utility, Silicon Valley Power, has not accepted new data center applications since 2023, and that it has no single data center with an electric load of 100 MW or more.²³

Therefore, a local supply bottleneck is not only a gigawatt AI campus problem. A large investment in buildings does not settle the power question either, since a finished shell waiting on a 2030 network upgrade is a carrying cost, not a data center.

Financing follows the delivery structure

Capital is available, although it comes with terms tied to milestones.

• Abilene: Crusoe, Blue Owl and Primary Digital announced a second phase of a $15 billion joint venture in May 2025. In addition, transaction counsel described a 1.2 GW campus, eight buildings and 300 MW of onsite generation.^{24, 25} That confirms a funding structure and scope at that date, although it does not confirm the full campus is energized.
• CoreWeave: its 2025 annual report separates more than 850 MW of active power at year-end 2025 from about 3.1 GW of contracted power for future deployment.²⁶ Active and contracted are different columns. The gap between them is normal in a build-out, while it becomes a problem only if contracts and timing do not line up.
• Milestone clauses: the FT reporting, as summarized by Seeking Alpha, describes an Anthropic agreement with Nscale for a West Virginia development that lets Anthropic withdraw if financing or delivery milestones are missed, and says Nebius estimated about $250 billion for a planned 5 GW expansion, roughly one fifth financed as of September.³ I have not checked those against the companies’ own filings, so treat them as reported.

That last point is the heart of it, because customers are writing delivery dates into contracts with exit rights. As a result, a missed power date is no longer only a schedule problem; it can be a revenue problem.

Who carries which exposure

Backlog is not cash. Remaining performance obligations describe future contracted business under accounting rules. However, the terms that start billing, the customer’s remedies, and what a lender will advance against the contract decide how much money is actually available to finish the building.

The cost of a slip can be modeled

All inputs here are assumptions for illustration rather than a forecast for any company.

For instance, assume a 100 MW IT phase with $1.2 billion of facility capital (excluding GPUs), 60% debt-funded, at 8% interest, plus $1 million a month in preservation and overhead during delay.

• Debt is $720 million.
• Interest on a constant balance is $4.8 million a month.
• With overhead, carry is $5.8 million a month.

At 10% interest, twelve months of interest becomes $72 million and total carry $84 million. Each extra point on $720 million adds $7.2 million a year.

That excludes principal, utility minimum charges, extra collateral, hardware depreciation, penalties and lost operating income. In addition, if the finished phase would have produced $3 million a month of operating cash before financing, a twelve-month delay defers $36 million of that.

The point is not the exact figure. Rather, it is that the same network delay pushes revenue out, raises carry, and makes the lender less willing to advance funds, all at once. A model that tests one stress at a time misses that.

Permits, lawsuits and water set the operating rights

Permits are site-specific

Texas shows how fast this changes. For example, on September 21, 2026, the governor directed TCEQ to halt permits sought by data centers pending completion of the ERCOT audit.²⁷ That is evidence of the directive, although whether a given permit is held, exempt, or later cleared takes the agency record for that permit.

Lawsuits are allegations until a court rules

The April 14, 2026 complaint in NAACP v. xAI, case 3:26-cv-00074 in the Northern District of Mississippi, challenges the Southaven gas plant and seeks relief under the Clean Air Act, alleging unpermitted turbine operation.²⁸ Those are the plaintiffs’ allegations. The case shows that fast onsite generation can produce a permitting and community fight even after equipment is installed. However, it does not show that onsite generation is unlawful.

A notice of intent to sue, a filed complaint, an injunction request, an injunction and a final judgment are five different states. Our watchlist counts 96 active lawsuits.¹⁵ That is not 96 cancelled projects.

Water is local and seasonal

Withdrawal, consumption and discharge are different flows. Microsoft’s December 2024 design describes closed-loop cooling with no evaporative water use for cooling, while keeping administrative water use and accepting an energy tradeoff.²⁹ That is a design direction for new sites rather than proof the whole fleet uses no water.

For example, at an illustrative direct cooling WUE of 0.5 liters per IT kWh, a 100 MW IT phase at 80% average load uses about 350.4 million liters a year, or about 92.6 million US gallons. The annual average can still hide a hard summer peak. Dry cooling cuts direct water use while raising electric load in hot weather, which can push the facility peak past its electrical allocation. Therefore, test water and power together.

How we grade a project at SAVRN

This is the part I care about most. Trackers count trouble, although they do not tell you if a specific phase will deliver. For that you need a method that is applied the same way every time and can be checked.

Here is the method. It is ours rather than Bernstein’s. However, it has not been back-tested on a fixed cohort yet, and I will say that every time we publish a score.

Rule 1. Grade a dated phase, not an announcement

Before any status, define: assessment date, phase, IT MW, facility MW, target service date, tenant acceptance criteria, and maximum allowed curtailment. A campus can have an energized first phase and a speculative fourth phase, so grade them separately.

Rule 2. Six stages, each needing more evidence

Announced → Studied → Contracted → Built → Energized → Accepted.

For example, a press release gets you to Announced, while a completed utility study gets you to Studied. An executed service agreement and binding equipment orders get you to Contracted. Energized means a defined electrical system passed commissioning. By contrast, Accepted means a customer took the service under its own tests. These are conceptual stages rather than a conversion rate.

Rule 3. Label evidence quality separately from status

• E0. Unsupported statement.
• E1. Dated public announcement.
• E2. Corroborated filing or study.
• E3. Executed site-specific agreement or final approval.
• E4. Witnessed completion or measured operation.

A project can have E4 land and E1 power. We report that gap instead of averaging it away. Ultimately, a critical unknown should stay visible.

Rule 4. Eight gates

Rule 5. Weighted screen, with fatal flaws overriding the math

Each dimension is rated 0 to 4. 0 is no evidence. 1 is an announcement. 2 is documented feasibility with open commitments. 3 is executed commitments with controlled dependencies. 4 is verified readiness or operation for that phase. Score equals the sum of weight × rating ÷ 4.

An 80-point threshold is a prioritization rule rather than an 80% chance of completion. In addition, any unresolved fatal constraint overrides the number.

Worked example: a 40 MW phase

Take an unnamed 40 MW IT proposal rated: power 1, finance 3, permits 3, procurement 2, technical 2, water and community 3.

• Power: 25 × 1 ÷ 4 = 6.25
• Finance: 20 × 3 ÷ 4 = 15.0
• Permits: 15 × 3 ÷ 4 = 11.25
• Procurement: 15 × 2 ÷ 4 = 7.5
• Technical: 15 × 2 ÷ 4 = 7.5
• Water and community: 10 × 3 ÷ 4 = 7.5
• Total: 55.0

Strong money and good permits do not fix an uncommitted power path. For instance, if an executed utility agreement moves power to 3, the score becomes 67.5. That earns a fresh review, although it still needs a reconciled procurement and acceptance schedule before it is ready.

This is what “credible” should mean in practice. Rather than a yes or no on a whole campus, it is a dated phase, a score, the evidence behind each rating, and the one thing that would change it.

Four decisions, not a score

The score feeds a decision:

• Proceed. The next defined commitment is supported, rather than every future phase.
• Proceed with conditions. Dependencies are named, budgeted and owned.
• Hold. A critical input is unresolved.
• Stop. The evidence shows an unacceptable constraint, or the sponsor rejects a workable fix.

Each decision names its owner, evidence links, open condition, deadline and next review. In fact, a clean Hold can save more money than an optimistic score.

Where the SAVRN platform fits

Here is how the public research we already publish maps onto the gates. Each page is a discovery tool, although none of them replaces the primary record for a specific parcel.

How a review runs, step by step

Take any campus announcement that lands in the news this week.

1. Define the phase. First, pull the sponsor’s own numbers. Write down IT MW versus facility MW and which meter it came from. If the release only says “1 GW,” the phase is not defined yet. As a result, every gate starts at 0 or 1.
2. Find the utility. Check the grid operator watchlist for the serving utility, its large-load tariff and the current queue process. In ERCOT, for example, ask whether the project passed Batch Zero verification. Similarly, in PJM, ask whether it is in the new registry. No named utility means Electrical stays at 1.
3. Check local rights. Search the moratorium tracker for the county, city and township. A pause in the jurisdiction does not kill the project. Instead, it goes on the schedule as a dated dependency.
4. Check disputes. Search the delay watchlist for the sponsor, the site and the county. Record each item by status: notice, complaint, injunction, ruling. In particular, do not count a lawsuit as a cancellation.
5. Check water. Look up the state in the water tracker, then ask the sponsor for the executed supply agreement and the peak-day balance. That matters because a state summary is not a supply right.
6. Check equipment. Ask which critical orders are binding, with release terms. Compare the claimed dates with the class-specific ranges in the supply chain report. For instance, a 2027 energization with no GSU order placed is a flag.
7. Check the money. Use the Capital Atlas to find who owns what. Then separate sponsor equity, funded debt, undrawn commitments, collateral and supplier deposits. Ask whether the phase stays funded through its stress date rather than just its base date.
8. Score, decide, and date it. Finally, apply the weights, record the E-levels, name the fatal flaw if there is one, and set the next review event.

That review produces a short, sourced record: here is the phase, here is what is proven, here is what is missing, here is the date it would change. It explains a project’s position instead of inventing a headline total.

What we will not claim

I want to be clear about the limits so the method stays credible.

• First, our tracker totals describe different populations. As a result, 390 projects, 1,449 restrictions, 289 utility actions and 96 lawsuits cannot be added up into a single failure rate.¹⁵
• Second, the 32,447-row national queue mixes generation, storage and load requests across all resource types. Only 46 rows are flagged data center, and 92 are large-load or data center studies.¹⁵ The 4,465 GW total is not a measure of AI demand.
• Third, we have checked our six headline figures against the live page. However, we have not done a full row-by-row audit of every record. “SAVRN reports” is the right attribution, while “an independent audit confirmed” is not.
• Finally, the scorecard above has not been back-tested. Before we publish scores as predictions, we will freeze a cohort, record the evidence available on each date, and compare predicted readiness with actual energization and customer acceptance. In addition, we will publish the unknown-data rate next to any score.

What a stronger delivery model looks like

We build data centers, so this is not only a research question for us. In fact, it is how we plan our own work.

For universities and research institutions, in particular, capacity should follow approved work and a real operating plan. Start with the actual research workloads, the data rules, the grant periods and the service level the institution can accept. Then size the first phase to that, with a documented power path and economics that work at realistic utilization. Finally, later phases stay conditional on demand and infrastructure evidence.

Modular construction, dedicated onsite power and low-water cooling are options to evaluate against that plan, because each brings its own fuel, permit, maintenance and commissioning obligations. None of them is an automatic exemption from the constraints above.

A practical first ninety days on a new institutional program looks like this:

Ninety days is how we organize the work rather than a promise that a grid connection, permit or factory slot can be secured in that time. Those can take much longer, and the work product should show it.

Any claim we make about faster delivery or lower cost should come from measured operating results on completed phases. Until then, what we sell is the evidence and the coordination.

Before the next big announcement

Ask for the evidence behind the first phase:

• Which electrical works are funded, and by whom?
• Which equipment orders are binding?
• Which permits apply, and are any challenged?
• When does the customer accept the service, and what can the customer do if it is late?
• How much cash is available if the date slips twelve months?

Ultimately, those answers tell you more about the next usable megawatt than the size of the announcement.

So, was the reading right? Mostly, because the physical and financial reasons behind the 35% figure are real and documented in public records. However, the figure is a credibility screen rather than a body count. The useful work, therefore, is grading each phase on its own evidence, and that is the work we have built SAVRN to do.

Chad Everett Harris is the founder of SAVRN. Figures checked October 19, 2026.