AI compute

AI compute is the processing capacity used to train and run artificial intelligence models, together with the chips, memory, networks, data centers, electricity and capital that supply it. It is measured in several units: floating-point operations per second (FLOP/s) for the speed of a chip or cluster, total floating-point operations (FLOP) for the size of a training run, GPU-hours for rented time, accelerator counts for the size of a fleet, and megawatts (MW) or gigawatts (GW) for the electrical capacity of a data center. Demand has risen steeply: Epoch AI estimates that the compute used to train frontier language models has grown about fivefold a year since 2020, and NVIDIA's data-center revenue went from $47.5 billion in its fiscal 2024 to $193.7 billion in fiscal 2026.^[1][2][3] By 2026 compute had also become an object of finance and policy: the US Commodity Futures Trading Commission (CFTC) called it "a scarce, capital-intensive commodity" while weighing futures contracts on its price, and the US government both restricted exports of advanced AI chips to China and, in January 2026, imposed a 25% fee, levied as an import tariff, on some chips approved for sale there.^[4][5]

How is AI compute measured?

No single number captures compute. Each common unit measures a different layer, from the arithmetic a chip performs to the power a data-center campus can draw.

Peak ratings overstate what software achieves. Meta reported that its 405-billion-parameter Llama 3 model was trained on up to 16,000 H100 GPUs at 38-43% model FLOPs utilization, meaning that less than half of the chips' peak BF16 throughput went into useful training arithmetic. The run's training compute was almost 50 times that of the largest Llama 2 model.^[7] Ratings also depend on the number format: NVIDIA's peak figure for the H100 doubles from BF16 to FP8 and doubles again if sparsity is assumed.^[6]

GPU-hours tie compute to money. DeepSeek priced the 2.788 million H800 GPU-hours of its DeepSeek-V3 run at $5.576 million by assuming a rental rate of $2 per GPU-hour, and noted that the total excluded earlier research and ablation experiments; the model was trained on a cluster of 2,048 H800s.^[8] Many of the largest purchases since 2025 have been specified in gigawatts rather than chips, such as the 10 GW OpenAI and NVIDIA letter of intent.^[9]

What is in the compute stack?

Accelerators

NVIDIA's H100, a Hopper-generation GPU with 80 billion transistors made on TSMC's 4N process, is the chip behind most of the prices quoted in this article.^[6] NVIDIA said in August 2026 that its Blackwell generation was still "the vast majority" of its revenue and that it was preparing to introduce Vera Rubin in the quarter ending in October 2026.^[10] AMD sells competing Instinct GPUs and Helios rack systems, which it said were being deployed by companies including Anthropic, Meta, Microsoft, OpenAI and Oracle.^[11]

Other AI accelerators are application-specific chips, many of them designed by large cloud and internet companies for their own data centers. Google's Tensor Processing Units (TPUs) are rented through Google Cloud and, from the second quarter of 2026, have also been delivered as systems to customers' own data centers.^[12] Amazon designs Trainium; Anthropic said in April 2026 that it used more than one million Trainium2 chips to train and serve its Claude models.^[13] Microsoft said in January 2026 that it had deployed Maia 200, an inference chip made on TSMC's 3 nm process with 216 GB of HBM3e and a 750 W power envelope, in its US Central region near Des Moines, Iowa.^[14] Meta's MTIA roadmap covers four generations, from the MTIA 300, built for ranking and recommendation models, to later chips for generative-AI inference.^[15] Broadcom builds custom AI accelerators for large buyers.^[16]

Memory

Accelerators are paired with high-bandwidth memory (HBM), stacked DRAM placed beside the processor. The H100 SXM5 carries 80 GB of HBM3 with more than 3 TB/s of bandwidth, and Maia 200 carries 216 GB of HBM3e at 7 TB/s.^[6][14] HBM supply is one of the limits on how many accelerators can be built (see "What limits the supply of compute?").

Interconnect

Inside a server, GPUs exchange data over NVLink, which gives each H100 900 GB/s of bandwidth to the other GPUs.^[6] Between servers, clusters use InfiniBand or Ethernet. Meta trained Llama 3 405B on a 24,000-GPU cluster connected by RDMA over Converged Ethernet (RoCE) and trained smaller Llama 3 models on NVIDIA Quantum-2 InfiniBand; in both, each GPU had a 400 Gbps network link, and the eight GPUs in each server were joined by NVLink.^[7] Google said in July 2026 that its Virgo Network could join a million accelerators across several data-center sites into one system.^[12]

Clusters and data centers

Accelerators are installed by the tens or hundreds of thousands as GPU clusters in purpose-built AI data centers. Epoch AI listed Colossus 2 in Memphis, Tennessee, as the largest known AI data center, at about 1.1 million H100-equivalents, and expected Meta's Hyperion campus to reach 3.7 million H100-equivalents by January 2028.^[1] Operators also describe their fleets in power rather than chips: CoreWeave reported 1.5 GW of active power and about 3.7 GW of contracted power at the end of June 2026.^[17]

Power

Electricity is the top layer of the stack. The International Energy Agency (IEA) estimated that data centers used about 415 TWh in 2024, around 1.5% of the world's electricity, and projected that this would more than double to about 945 TWh by 2030, with AI "the most important driver of this growth." It expected data centers to account for nearly half of US electricity demand growth to 2030.^[18] The AI energy consumption article covers the wider debate.

Training compute and inference compute

Demand for compute divides into training, in which a model's parameters are learned from data, and inference, in which the trained model is run each time it answers a request.

Epoch AI estimates that the training compute of frontier language models has grown about 5 times a year since 2020 (90% confidence interval 4-6 times), a doubling roughly every 5.2 months, and that the total computing power of the world's stock of AI chips has grown about 3.4 times a year since 2022.^[1] Developers size these runs with scaling laws, empirical relationships between compute, model size, data and performance. Meta wrote that its scaling laws, extrapolated to a budget of 3.8 × 10^25 FLOP, suggested training a 402-billion-parameter model on 16.55 trillion tokens, close to the 405-billion-parameter model it built.^[7]

Inference was already a large share of chip demand before reasoning models appeared: NVIDIA estimated that about 40% of its data-center revenue in fiscal 2024 was for AI inference.^[2] Measured at a single developer, the split can look different; Epoch AI estimated that only 30% of OpenAI's compute spending in 2024 went to inference.^[1] In September 2024 OpenAI introduced o1, a model trained with reinforcement learning to reason through a chain of thought, and reported that its performance "consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute)."^[19] Spending more test-time compute on each answer moves demand from one-off training runs toward every query a model serves. Token volumes have grown quickly: Alphabet said in July 2026 that its model APIs were processing about 22 billion tokens a minute, up from 16 billion a quarter earlier, and that it remained "supply constrained."^[12]

Who supplies AI compute?

Chip designers

NVIDIA's data-center revenue was $47.5 billion in fiscal 2024 (the year to January 2024), 217% more than a year earlier, and $193.7 billion in fiscal 2026.^[2][3] In the quarter to July 26, 2026 it was $89.0 billion, up 117% from a year earlier; $48.7 billion came from hyperscale customers and $40.3 billion from AI clouds, industrial and enterprise customers. NVIDIA guided to total company revenue of $108.0 billion for the following quarter and assumed no data-center compute revenue from China.^[10]

AMD's data-center segment, which sells EPYC server processors as well as Instinct GPUs, had revenue of $6.7 billion in the second quarter of 2026, up 107%; its results release also listed a partnership with Anthropic to deploy up to 2 GW of MI450-series GPUs in Helios racks.^[11] Broadcom reported AI semiconductor revenue of $16.7 billion in its quarter to August 2, 2026, up 221% from a year earlier, and forecast $21.7 billion for the following quarter.^[16] Google began recognizing revenue from TPU system sales in the second quarter of 2026 and said most of the revenue from those agreements would arrive in 2027.^[12] In February 2026 OpenAI agreed to consume about 2 GW of Trainium capacity through AWS.^[20]

Foundry and memory

TSMC manufactures both the H100 and Maia 200.^[6][14] In October 2026 the company said it expected revenue to grow about 44% in US dollars for the year, and chief executive C.C. Wei told analysts that "the demand actually is still far above our capability today to supply."^[21]

Three companies supply nearly all HBM. Counterpoint Research estimated that SK Hynix had 50% of HBM revenue in the second quarter of 2026, Samsung 33% and Micron 18%, after Samsung rose from 21% in the first quarter.^[22] SK Hynix began mass shipments of HBM4 in that quarter, said it had signed long-term agreements with about 10 customers, and described a market "where customer demand exceeds supply capabilities."^[23]

Clouds

The largest cloud providers, or hyperscalers, are the biggest buyers of AI hardware: hyperscale customers accounted for more than half of NVIDIA's data-center revenue in the quarter to July 2026.^[10] In mid-2026 all three of the largest US clouds said demand exceeded what they could supply. Microsoft said in July that "customer demand continues to exceed available capacity."^[24] Alphabet reported a Google Cloud backlog of $514 billion and said it would expand its use of third-party capacity as "a bridging strategy."^[12] Amazon chief executive Andy Jassy said the company would "still not have enough capacity to meet all the demand we have in 2026."^[25]

Revenue from Oracle Cloud Infrastructure rose 121% to $7.4 billion in the quarter to August 31, 2026, and Oracle's remaining performance obligations, a measure of contracted future revenue, reached $664 billion. Its spending has exceeded its operating cash flow: Oracle's capital expenditures of $55.7 billion in fiscal 2026 left free cash flow of minus $23.7 billion, it spent another $28.5 billion in the first quarter of fiscal 2027, and it sold $20 billion of stock in that quarter.^[26]

Specialist GPU clouds, known as neoclouds, include CoreWeave, Nebius, Lambda and Crusoe. CoreWeave reported revenue of $2.58 billion for the quarter to June 30, 2026 and a revenue backlog of about $104 billion, not counting more than $25 billion of commitments added early in the next quarter.^[17] Neoclouds also sell to hyperscalers: in September 2025 Nebius agreed to provide Microsoft dedicated GPU capacity from a new data center in Vineland, New Jersey, for about $17.4 billion through 2031.^[27] NVIDIA rents capacity as well; at the end of July 2026 it listed $29 billion of multi-year cloud service agreements and a further $36 billion of AI cloud agreements among its commitments.^[10]

How is compute bought and priced?

Buyers obtain compute in five main ways, which differ in commitment, price and risk.

Published prices are mostly list prices, and they are not like-for-like: a hyperscaler's price covers a whole virtual machine with its processors, memory and local storage, while some specialist clouds quote a price per GPU. Google's A3 machine in 2024, for example, bundled eight H100s with 208 virtual CPUs, 1,872 GB of memory and 6 TiB of local SSD.^[29] Contract prices are mostly private: the CFTC noted in 2026 that compute prices are formed mainly in bilateral deals, which limits public price data.^[4] Current quotes from many providers are collected in the cloud GPU pricing comparison.

H100 rental prices, 2023-2026

NVIDIA said in November 2022 that it had started shipping the H100.^[32] The following dated observations trace its rental price from 2023 to 2026; per-GPU figures for 8-GPU instances are the listed price divided by eight.

Sources: AWS price lists and announcement, archived Lambda and Google Cloud pricing pages, and SemiAnalysis.^{[29][30][33][34][35][36][37]}

Hyperscaler list prices and the prices paid by contract buyers moved separately. AWS listed its 8-GPU P5 instance at $98.32 an hour in August 2023, more than four times Lambda's on-demand rate per GPU.^[33][30] In June 2025 AWS cut on-demand P5 prices by 44%, noting that the growth in demand for GPU capacity had "outpaced industry-wide supply"; its August 2025 price list showed $55.04 an hour.^[34][35] Among contract buyers, SemiAnalysis found that one-year H100 rental prices fell to $1.70 per GPU-hour in October 2025 and rose almost 40% to $2.35 by March 2026, when it reported that on-demand capacity was "sold out across all GPU types."^[36] Lambda's on-demand price for an 8x H100 system was $3.99 per GPU-hour in May 2026, up from $2.59 in August 2023.^[37][30]

Capital spending and the largest deals

How much are hyperscalers spending?

Company totals for capital spending include offices, warehouses and other assets as well as AI data centers. The figures below are each company's cash purchases of property and equipment; the outlook column uses each company's own definition, which in some cases adds finance leases. The AI capex article tracks the series in more detail.

Sources: company annual reports, earnings releases and earnings calls.^{[12][24][25][28][38][39][40][41]}

At the low ends of their ranges, and on the companies' differing definitions, the four latest 2026 outlooks add up to more than $700 billion. Some of the increases reflect component prices rather than more capacity. Amazon attributed most of its increase from about $200 billion to about $220 billion to higher memory costs, and Microsoft said its capital spending of $41 billion in the quarter to June 2026 included "the impact from higher component pricing."^[25][24] Microsoft also said its $175 billion estimate reflected an accounting change that moves future data-center leases from finance leases, which count as capital expenditure, to operating leases, which do not, and that its underlying investment expectations were unchanged.^[24]

Largest compute deals of 2025 and 2026

The largest deals of 2025 and 2026 were multi-year commitments by AI developers, chiefly OpenAI and Anthropic, to buy capacity from clouds, chipmakers and data-center developers. Their stated values are not comparable: some are binding contracts, some are statements of intent, some are investment pledges, and some come only from press reports. The table states the basis of each figure; the list of AI compute deals covers many more.

Several deals tie compute purchases to investment flowing the other way. NVIDIA's letter of intent with OpenAI paired at least 10 GW of systems with an NVIDIA investment of up to $100 billion, AMD's agreement gave OpenAI warrants that vest as deployments proceed, and in August 2026 NVIDIA agreed to guarantees, capped at $105 billion, that provide credit support for a campus in Ohio that SB Energy will lease to OpenAI.^[9][44][10] Anthropic also listed a $30 billion commitment to buy Azure capacity, made as part of a partnership with Microsoft and NVIDIA, among its compute arrangements.^[47]

Compute as a market and a policy object

Price indexes and derivatives

Because most compute is sold through private contracts, research firms and data providers began publishing rental-price benchmarks; SemiAnalysis, for example, launched a monthly index of one-year H100 rental prices in April 2026.^[36] Such indexes made it possible to write financial contracts that settle on the price of compute, which the CFTC called "an emerging class of derivatives contracts"; they are described in compute derivatives and GPU compute futures.^[4] In August 2026 the CFTC asked for public comment on listing such contracts, with comments due by October 20, 2026. It wrote that compute markets were fragmented, that price formation "primarily occurs in opaque bilateral transactions," and that compute "may not yet exhibit" the fungibility, standardization and liquidity of commodities that usually underlie futures markets.^[4]

Export controls

In November 2022 NVIDIA reported that the US government had announced new restrictions on exports of its A100 and H100-based products to China during the preceding quarter.^[32] Such controls are administered by the Commerce Department's Bureau of Industry and Security (BIS) under the Export Administration Regulations, and before 2026 BIS applied a "presumption-of-denial" review to exports of chips such as NVIDIA's H200 to China, Macau and other arms-embargoed countries.^[5] In December 2025 President Donald Trump said NVIDIA would be allowed to sell the H200 in China in exchange for the US government receiving 25% of an unspecified sum, and on January 15, 2026 BIS adopted case-by-case review for chips within certain technical limits, including "the Nvidia H200, AMD MI325X, and similar chips," bound for China and Macau, subject to testing by a third-party laboratory in the United States.^[5] NVIDIA's sales there were small in mid-2026: in the quarter to July 2026, shipments of its data-center Hopper products to China were less than 1% of its data-center revenue.^[10] The export controls article covers the history of these rules.

The Section 232 chip tariff

The 25% fee was collected as an import tariff. On January 14, 2026, Proclamation 11002, issued under Section 232 of the Trade Expansion Act of 1962 after a Commerce Department investigation opened in April 2025, imposed a 25% tariff on "a very narrow category of semiconductors," which the White House said included the H200 and MI325X. Chips imported for US data centers, research and development, startups, consumer applications and several other domestic uses were exempted, and the Congressional Research Service noted that it was unclear whether the tariff would apply to any chips other than those exported to China.^[5] The Section 232 chip tariff article describes its terms and the legal debate.

Compute thresholds in AI law

Because training compute can be counted, lawmakers have used it to define which models face extra obligations, part of the wider field of compute governance. California's Transparency in Frontier Artificial Intelligence Act (SB 53), signed on September 29, 2025 and in force from January 2026, defines a frontier model as one trained with more than 10^26 integer or floating-point operations. That threshold matches the one in the 2023 US executive order on AI and is ten times the European Union AI Act's threshold of 10^25 FLOP.^[48]

What limits the supply of compute?

Power and grid connections

New AI capacity is constrained by how fast data centers can obtain electricity. The IEA estimated in 2025 that, unless grid risks were addressed, around 20% of planned data-center projects could be delayed.^[18] In the United States, BloombergNEF estimated in 2026 that almost 20% of a 253 GW national pipeline of proposed data centers, nearly 50 GW, was at risk from a pause in Texas, where Governor Greg Abbott imposed a moratorium on new data-center grid approvals in August 2026 and in September ordered the state environmental regulator to halt data-center permits until the grid operator audited its queue of projects waiting to connect.^[49]

Memory and packaging

Chip output also depends on memory and on advanced packaging, the step that assembles processor dies and HBM stacks into one package. In October 2026 TSMC's Wei said the company had "a little bit shortage on the back end," its packaging operations, and that its CoWoS packaging remained customers' "number one choice."^[21] Memory supply tightened in 2026. NVIDIA's purchase commitments rose from $119 billion to $279 billion in the quarter to July 2026, "primarily related to the procurement of memory," and SK Hynix reported significant quarter-on-quarter increases in DRAM and NAND flash prices.^[10][23] Those prices fed into the hyperscalers' budgets, as Amazon's and Microsoft's comments on memory and component costs showed.^[25][24]

Depreciation

Accelerators lose value as faster generations arrive, and companies have to choose how long to depreciate them. Their choices have diverged. Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five from January 1, 2025, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning," which raised its 2025 depreciation expense by $1.4 billion.^[38] Meta did the opposite, lengthening the estimated useful lives of most of its servers and network assets to 5.5 years from the same date.^[28] Short-lived equipment is a large share of the spending: Microsoft said roughly two-thirds of its capital spending in the quarter to June 2026 went to short-lived assets, mainly CPUs and GPUs.^[24] The GPU depreciation article covers the accounting debate.

Local opposition and moratoria

Opposition from residents has become a constraint of its own. Data Center Watch, an advocacy group, said $68 billion of data-center development was delayed or cancelled in the second quarter of 2026 because of local pushback, and an NBC News poll found that 64% of Americans would be less likely to vote for a candidate who "supports building local data centers."^[49] Data centers became an issue in the 2026 midterm elections, including Abbott's own re-election race, and the governor's office said he would work with the legislature to eliminate financial incentives for data centers.^[49] State and local pauses are tracked in the AI data center moratorium article.

References

1. ^1 ^2 ^3 ^4 ^5"Machine Learning Trends." Epoch AI, 2026-02-05 (accessed 2026-10-15). epoch.ai/trends
2. ^1 ^2 ^3"Form 10-K: Annual Report for the Fiscal Year Ended January 28, 2024." NVIDIA Corporation (SEC filing), 2024-02-21. sec.gov/...nvda-20240128
3. ^1 ^2"NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026." NVIDIA, 2026-02-25. nvidianews.nvidia.com/...h-quarter-and-fiscal-2026
4. ^1 ^2 ^3 ^4"Request for Comment on the Listing of Compute Derivatives Contracts." Federal Register, Vol. 91, No. 161 (US Commodity Futures Trading Commission), 2026-08-21. govinfo.gov/...2026-17163.pdf
5. ^1 ^2 ^3 ^4"Legal Authority for Export Controls and Tariffs on Semiconductor Chips Sold to China." Congressional Research Service, Legal Sidebar LSB11409, 2026-03-23. congress.gov/...LSB11409
6. ^1 ^2 ^3 ^4 ^5 ^6"NVIDIA H100 Tensor Core GPU Architecture (whitepaper, V1.04)." NVIDIA, 2023-05-04. dam-cdn.nvd.orangelogic.com/...0117r17n8042d73.pdf
7. ^1 ^2 ^3 ^4Llama Team, AI at Meta. "The Llama 3 Herd of Models." arXiv:2407.21783, 2024-07-31. arxiv.org/...2407.21783
8. ^1 ^2DeepSeek-AI. "DeepSeek-V3 Technical Report." arXiv:2412.19437, 2024-12-27. arxiv.org/...2412.19437
9. ^1 ^2 ^3 ^4"OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems." NVIDIA, 2025-09-22. nvidianews.nvidia.com/...oy-10gw-of-nvidia-systems
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"CFO Commentary on Second Quarter Fiscal 2027 Results." NVIDIA (SEC filing, Form 8-K Exhibit 99.2), 2026-08-26. sec.gov/...q2fy27cfocommentary
11. ^1 ^2"AMD Reports Second Quarter 2026 Financial Results." AMD, 2026-08-04. ir.amd.com/...econd-quarter-2026-financial-results
12. ^1 ^2 ^3 ^4 ^5 ^6"Alphabet Second Quarter 2026 Earnings Conference Call: Transcript." Alphabet Investor Relations, 2026-07-22. s206.q4cdn.com/...2026_Q2_Earnings_Transcript.pdf
13. ^1 ^2"Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute." Anthropic, 2026-04-20. anthropic.com/...anthropic-amazon-compute
14. ^1 ^2 ^3Scott Guthrie. "Maia 200: The AI accelerator built for inference." Official Microsoft Blog, 2026-01-26. blogs.microsoft.com/...lerator-built-for-inference
15. ^Ryan Smith. "Meta's MTIA Custom AI Silicon at Hot Chips 2026." ServeTheHome, 2026-08-25. servethehome.com/...m-ai-silicon-at-hot-chips-2026
16. ^1 ^2"Broadcom Inc. Announces Third Quarter Fiscal Year 2026 Financial Results and Quarterly Dividend." Broadcom (SEC filing, Form 8-K Exhibit 99), 2026-09-02. sec.gov/...avgo-08022026x8kxex99
17. ^1 ^2"CoreWeave Reports Strong Second Quarter 2026 Results." CoreWeave, 2026-08-11. investors.coreweave.com/...default
18. ^1 ^2"Energy and AI: Executive Summary." International Energy Agency, 2025-04-10. iea.org/...executive-summary
19. ^"Learning to Reason with LLMs." OpenAI, 2024-09-12. openai.com/...learning-to-reason-with-llms
20. ^1 ^2"OpenAI and Amazon announce strategic partnership." OpenAI, 2026-02-27. openai.com/...amazon-partnership
21. ^1 ^2"TSMC (TSM) Q3 2026 Earnings Call Transcript." The Motley Fool, 2026-10-15. fool.com/...c-tsm-q3-2026-earnings-call-transcript
22. ^"Global DRAM and HBM Market Share: Quarterly." Counterpoint Research, 2026-09-02. counterpointresearch.com/...m-and-hbm-market-share
23. ^1 ^2"SK hynix Announces 2Q26 Financial Results." SK hynix Newsroom, 2026-07-29. news.skhynix.com/...q2-2026-business-results
24. ^1 ^2 ^3 ^4 ^5 ^6"Microsoft Fiscal Year 2026 Fourth Quarter Earnings Conference Call: Transcript." Microsoft Investor Relations, 2026-07-29. microsoft.com/...earnings-fy-2026-q4
25. ^1 ^2 ^3 ^4Shane Snider. "Amazon Lifts 2026 AI Capex to $220B, Still Capacity-Constrained." Data Center Knowledge, 2026-07-31. datacenterknowledge.com/...emand-outpaces-capacity
26. ^"Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues." Oracle (SEC filing, Form 8-K Exhibit 99.1), 2026-09-10. sec.gov/...orcl-ex99_1
27. ^1 ^2 ^3"Form 6-K: Commercial Agreement with Microsoft." Nebius Group N.V. (SEC filing), 2025-09-08. sec.gov/...tm2525580d1_6k
28. ^1 ^2 ^3"Form 10-K: Annual Report for the Fiscal Year Ended December 31, 2025." Meta Platforms, Inc. (SEC filing), 2026-01-28. sec.gov/...meta-20251231
29. ^1 ^2 ^3"VM Instance Pricing (archived capture)." Google Cloud, 2024-05-21. web.archive.org/...vm-instance-pricing
30. ^1 ^2 ^3 ^4"GPU Cloud: VMs for Deep Learning (archived capture)." Lambda, 2023-08-10. web.archive.org/...gpu-cloud
31. ^"SF Compute Raises $40M to Build Marketplace for AI Compute Capacity." eWeek, 2025-11-27. eweek.com/...sf-compute-ai-funding
32. ^1 ^2"CFO Commentary on Third Quarter Fiscal 2023 Results." NVIDIA, 2022-11-16. s201.q4cdn.com/...Q3FY23-CFO-Commentary.pdf
33. ^1 ^2"Amazon EC2 Price List, US East (N. Virginia), Offer Version 20230905140324." Amazon Web Services, 2023-09-05. pricing.us-east-1.amazonaws.com/...us-east-1
34. ^1 ^2Channy Yun. "Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances." AWS News Blog, 2025-06-05. aws.amazon.com/...nvidia-gpu-accelerated-instances
35. ^1 ^2"Amazon EC2 Price List, US East (N. Virginia), Offer Version 20250828212014." Amazon Web Services, 2025-08-28. pricing.us-east-1.amazonaws.com/...us-east-1
36. ^1 ^2 ^3"The Great GPU Shortage: Rental Capacity; Launching our H100 1 Year Rental Price Index." SemiAnalysis, 2026-04-02. newsletter.semianalysis.com/...age-rental-capacity
37. ^1 ^2"Pricing (archived capture)." Lambda, 2026-05-16. web.archive.org/...pricing
38. ^1 ^2"Form 10-K: Annual Report for the Fiscal Year Ended December 31, 2025." Amazon.com, Inc. (SEC filing), 2026-02-06. sec.gov/...amzn-20251231
39. ^"Microsoft Cloud and AI Strength Fuels Fourth Quarter Results." Microsoft Investor Relations, 2026-07-29. microsoft.com/...press-release-webcast
40. ^"Form 10-K: Annual Report for the Fiscal Year Ended December 31, 2025." Alphabet Inc. (SEC filing), 2026-02-05. sec.gov/...goog-20251231
41. ^"Meta Reports Second Quarter 2026 Results." Meta Platforms, Inc. (SEC filing, Form 8-K Exhibit 99.1), 2026-07-29. sec.gov/...meta-06302026xexhibit991
42. ^"Announcing The Stargate Project." OpenAI, 2025-01-21. openai.com/...announcing-the-stargate-project
43. ^Tobias Mann. "OpenAI reportedly on the hook for $300B Oracle Cloud bill." The Register, 2025-09-11. theregister.com/...openai_reportedly_on_the_hook
44. ^1 ^2"AMD and OpenAI Announce Partnership to Deploy 6 Gigawatts of AMD GPUs." AMD, 2025-10-06. amd.com/...nai-announce-strategic-partnership-to-d
45. ^"The next chapter of the Microsoft-OpenAI partnership." Official Microsoft Blog, 2025-10-28. blogs.microsoft.com/...icrosoft-openai-partnership
46. ^"Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute." Anthropic, 2026-04-06. anthropic.com/...google-broadcom-partnership-compute
47. ^1 ^2 ^3"Higher usage limits for Claude and a compute deal with SpaceX." Anthropic, 2026-05-06. anthropic.com/...higher-limits-spacex
48. ^Joshua A. Geltzer, Susan J. Hennessey, Hilary A. Hurd and Bardia Vaseghi. "Transparency in Frontier Artificial Intelligence Act (SB-53): California Requires New Standardized AI Safety Disclosures." WilmerHale, 2025-10-01. wilmerhale.com/...andardized-ai-safety-disclosures
49. ^1 ^2 ^3Isabel O'Brien. "Texas Gov. Abbott orders data center permit halt weeks after issuing moratorium." CNBC, 2026-09-21. cnbc.com/...governor-abbott-ai-data-center-permits