# What AI Costs: GPU, Token and Data Center Prices, Tracked Source: https://savrn.com/ai-index/pricing Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- [SAVRN Index.](https://savrn.com/ai-index) Explore the IndexAI pricing AI pricing pagesPricing overview Tracking 796 prices · Checked 3 hours ago The price tracker ## Four markets, every price kept 796 prices from 55 providers, checked 3 hours ago against providers' price feeds and pages. Every change is recorded with its source. [GPU hourly pricesEvery GPU, by the hour$4.33median H100 list price per GPU-hour229 prices from 27 providers across 10 GPUsOpen →](https://savrn.com/ai-index/pricing/gpus) [Wholesale and landlordsWhat sellers get$3.96 to $3.97GB300 per GPU-hour on multi-year contractsLandlords: median $155.65 per kW-month across 20 shell leasesOpen →](https://savrn.com/ai-index/pricing/wholesale) [Frontier modelsWhat the labs charge$10.00median output price per million tokens, 73 models on the labs' own APIsFrom $0.40 to $600.00. Anthropic, OpenAI, Google, xAI, Microsoft Azure and AWS Bedrock, each in its own sectionOpen →](https://savrn.com/ai-index/pricing/models) [Open modelsThe same model, many hosts$0.17 to $0.75output per million tokens for gpt-oss-120b across 13 hosts118 open models priced by 17 hostsOpen →](https://savrn.com/ai-index/pricing/open-models) Who charges what ## Three businesses, one kilowatt Reduced to dollars per critical kilowatt-month. The dark part of each bar is what that business pays, and the copper part is what it keeps. Neocloud, as developerCore Scientific, Hut 8, TeraWulf, Cipher, Applied Digital, CleanSpark, Riot $128.00keeps $24 AI cloud operator, as tenantCoreWeave, Nebius, Lambda, Together $1,358.70keeps $363 Model companyAnthropic, OpenAI $2,717.85keeps $1,359 Latest changes ## What moved [All 296 changes →](https://savrn.com/ai-index/pricing/changes) | Date | Market | Seller | Item | Was | Now | Change | Source | | --- | --- | --- | --- | --- | --- | --- | --- | | Oct 21, 2026 | [GPU hourly](https://savrn.com/ai-index/pricing/gpus/h200) | Nebius | H200NVIDIA HGX H200, on-demand · per GPU-hour | $4.50 | $5.40 | +20% | [nebius.com ↗](https://nebius.com/prices) | | Oct 21, 2026 | [GPU hourly](https://savrn.com/ai-index/pricing/gpus/h100) | Nebius | H100NVIDIA HGX H100, on-demand · per GPU-hour | $3.85 | $4.50 | +16.9% | [nebius.com ↗](https://nebius.com/prices) | | Oct 21, 2026 | [GPU hourly](https://savrn.com/ai-index/pricing/gpus/b200) | Nebius | B200NVIDIA HGX B200, on-demand · per GPU-hour | $7.15 | $8.50 | +18.9% | [nebius.com ↗](https://nebius.com/prices) | | Oct 21, 2026 | [Open models](https://savrn.com/ai-index/pricing/open-models) | Baseten | Beam-1.0-501B-A23BHugging Face Inference Providers · output per million tokens | $4.30 | $1.20 | −72.1% | [huggingface.co ↗](https://huggingface.co/reflection/Beam-1.0-501B-A23B) | | Oct 21, 2026 | [Open models](https://savrn.com/ai-index/pricing/open-models) | Baseten | Beam-1.0-501B-A23BHugging Face Inference Providers · input per million tokens | $1.30 | $0.40 | −69.2% | [huggingface.co ↗](https://huggingface.co/reflection/Beam-1.0-501B-A23B) | | Oct 20, 2026 | [GPU hourly](https://savrn.com/ai-index/pricing/gpus/h200) | Verda | H200Serverless containers, 1x H200 SXM5 141GB, spot · per GPU-hour | $2.93 | $3.14 | +7.2% | [verda.com ↗](https://verda.com/pricing) | What is not in the price ## What they don't charge for, and what they don't say Rack-scale list prices 9 of 16 GB300 offers the Index tracks show no public price. [SAVRN Index price checks, Oct 24, 2026](https://savrn.com/ai-index/pricing/gpus/gb300) Providers with no public price Fluidstack, Lightning AI and TensorWave show no public price for any GPU the Index tracks. [SAVRN Index price checks, Oct 24, 2026](https://savrn.com/ai-index/pricing/gpus) Network, fiber and cross-connects Not disclosed in any of the 47 contracts in SAVRN's set, so no quoted rent includes them. [Twenty Cents to Twenty-Five Dollars](https://savrn.com/blog/twenty-cents-to-twenty-five-dollars) Throughput and utilization No contract discloses how many tokens a rack produces or how busy it runs. SAVRN measures it from MLPerf results instead. [Twenty Cents to Twenty-Five Dollars](https://savrn.com/blog/twenty-cents-to-twenty-five-dollars) Tenant improvements Tenants pay for fit-out, but no filing states the amount. SAVRN carries $16.43 per kW-month from market practice. [Twenty Cents to Twenty-Five Dollars](https://savrn.com/blog/twenty-cents-to-twenty-five-dollars) The discount off list An Opus 5 input token lists at $5.00 per million and bills at $0.25 once it is cached and batched. [Twenty Cents to Twenty-Five Dollars](https://savrn.com/blog/twenty-cents-to-twenty-five-dollars) Who the tenant is Five contracts withhold the tenant's name, including CoreWeave's build-to-suit and Riot's investment-grade counterparty. SAVRN contract workbook Redacted terms TeraWulf redacted the capacity and rent for Lake Mariner, and Cipher and CleanSpark redacted their escalator percentages. SAVRN contract workbook Rent inside full-stack deals Full-stack contracts such as Nscale's price the building and the compute together, with no separate rent. SAVRN contract workbook The SAVRN reference set ## Seven references, one system Capital, delay, policy, water, grid, permits and scarcity are one system seen from seven sides. Each reference is primary-sourced, refreshed on a schedule, and links into the others, so a project can be followed from the money committed to the docket that stopped it. Capital Atlas Who is funding AI data centers: capital providers, named sites, financing transactions, and the blocked-and-delayed record, state by state and company by company. [Open the atlas →](https://savrn.com/data-center-capital-atlas) Delay Watchlist Project-level slippage: named projects, jurisdictional actions, utility and federal dockets, SEC disclosures, and the interconnection queue. Open the watchlist → Moratorium Tracker The policy layer: every state measure and local moratorium mapped and cited, with one page per state. Open the tracker → Water Tracker Every federal, state, local and interstate water measure that touches a data center, plus a fifty-state water ledger. Open the water tracker → Grid Operator Watchlist ERCOT, PJM, MISO, CAISO, SPP, NYISO, ISO-NE and FERC: the large-load rules, filings and queue positions that decide whether a site can energize. Open the watchlist → Permits and Power Development Data center and substation permits, large loads in the interconnection queues, and the power plants planned to serve them, refreshed hourly from public records. Open the permits tracker → SAVRN Scarcity Tracker Eleven primary-sourced indicators of AI infrastructure scarcity, one composite index, read against the token index, SAVRN's own cost basis and the frontier ceiling. Open the scarcity tracker → Have a specific address? Site Intelligence reads all seven references against one location, county, utility and grid operator included, and cites every finding. [Check an address →](https://savrn.com/site-intelligence)