Compute market review: Q3 2025

Summary

Looking back at July to September 2025, the public record shows posted GPU rental prices and contracted compute moving in different directions. Hourly rates for NVIDIA H100 GPUs drifted lower, while cloud providers, AI developers and their financiers committed to multi-year capacity on a scale measured in gigawatts.

• Posted H100 rents drifted lower. Silicon Data's H100 rental index ranged between 2.22 and 2.26 in August and touched 2.13 in mid-September before closing the month near 2.16 [2]. Its year-end analysis gave July to December medians of $6.26 per GPU-hour at hyperscalers, $3.33 at neoclouds and $1.95 on marketplaces [3].
• Contract prices moved less. CoreWeave reported that a customer renewed a contract for more than 10,000 H100s two quarters early, at a price within 5% of the original agreement [5]. SemiAnalysis later placed the one-year H100 contract low at $1.70 per GPU-hour in October 2025 [6].
• Blackwell reached volume. NVIDIA's data center revenue was $41.1 billion for its quarter ended 27 July, with Blackwell data center revenue up 17% from the prior quarter [7]. CoreWeave announced customer deployments of GB300 NVL72 systems on 3 July [9].
• Contracts grew larger and longer. The quarter brought Nebius and Microsoft (about $17.4 billion), CoreWeave and Meta (up to about $14.2 billion), a further CoreWeave and OpenAI order (up to $6.5 billion) and a letter of intent for at least 10 GW of NVIDIA systems for OpenAI [16, 23, 20, 19].
• Credit support came from outside the borrower. Google backstopped about $4.6 billion (our sum) of Fluidstack lease obligations across TeraWulf and Cipher Mining, and NVIDIA agreed to buy CoreWeave capacity that customers leave unsold [15, 22, 18].
• Spending and power constraints rose together. Alphabet, Amazon, Meta and Microsoft each reported higher capital spending in the quarter to September than in the quarter to June, and Amazon's chief executive called power "the single biggest constraint" [27, 28, 29, 30, 31, 32, 33, 34]. PJM's 2026/2027 capacity auction cleared at its price cap [35].

Background

AI compute reaches users in two ways. Buyers rent GPU capacity by the hour or by contract from hyperscale clouds and from neoclouds, the cloud providers built around GPU capacity. Or they own GPU systems, financed with equity, debt or customer prepayments. The two are linked. What a GPU earns on rent supports what an owner can pay for it, and what a lender will advance against it.

In mid-2025 this market was between generations. NVIDIA's Hopper GPUs (H100 and H200) made up the installed base that most rental prices describe. Blackwell systems (B200, GB200 NVL72 and the Blackwell Ultra GB300 NVL72) were ramping into customer deployments [9]. Rental prices for the older generation had already fallen from their 2024 levels [3]. Silicon Data's index, launched in May 2025, opened at $2.37 per GPU-hour, and the launch coverage described typical rates of $2.40 to $2.60 in the months before January 2025 [1]. In June 2025, AWS cut on-demand prices for its P5 instances by up to 45% and for P5en instances by up to 26%, a move Silicon Data described as an H100 price cut [4, 3].

At the same time, model developers and large platforms were signing long-term commitments for capacity that had not been built, sited at specific campuses and tied to specific power supplies. Those commitments, and the debt raised against them, became the quarter's main source of public information about compute prices and terms.

Data and method

Period. 1 July to 30 September 2025. Results reported shortly after the quarter (for example, third-quarter earnings released in late October and November 2025) are used for what they state about July to September, and each is labeled with its fiscal period. Two later sources are used for specific facts: Silicon Data's December 2025 analysis of 2025 tier prices [3] and SemiAnalysis's April 2026 contract index [6].

Sources. Company press releases and SEC filings (Forms 8-K and 6-K with their exhibits); earnings call transcripts published by the companies or by transcript services; grid operator releases; and published commentary from two GPU price index providers, Silicon Data and SemiAnalysis. Every source appears in the numbered list at the end. We opened each source and took figures as stated.

Measures. Six groups:

1. Rental prices: published H100 rental index levels and operator statements about contract pricing.
2. Accelerator supply: NVIDIA data center revenue and Blackwell deployment announcements.
3. Capacity contracts: announced agreements between compute providers, chip suppliers and buyers.
4. Financing: debt and equity raised by compute providers and large buyers, and third-party credit support.
5. Capital spending: reported and guided capital expenditures of Alphabet, Amazon, Meta and Microsoft.
6. Power: statements about power constraints and the PJM capacity auction.

Treatment of figures. We keep three kinds of price evidence apart: posted or spot rental rates (what providers publish), contract prices (what buyers commit to for a term) and reported deal values (the maximum a buyer may pay under stated conditions). Contract values are reported as their filings state them, including the words "up to" where they appear. Capital expenditure measures differ by company: Alphabet reports purchases of property and equipment; Meta reports capital expenditures including principal payments on finance leases; Microsoft reports capital expenditures including finance leases; Amazon's figures here are the cash capital expenditures stated on its earnings calls. Fiscal calendars differ too. NVIDIA's quarters end in late January, April, July and October; Oracle's fiscal first quarter ends on 31 August; Microsoft's fiscal year starts on 1 July.

Relation to the Rillor Compute Index. The Rillor Compute Index methodology v1.0 defines four series: RCI Rental, RCI Availability, RCI Hardware and RCI Rent-to-Value. This review keeps the same separation between rental, availability, hardware and financing questions. It reports third-party figures only and contains no RCI values.

Findings

1. Rental prices: posted rates drifted, contracts held

Silicon Data described August as a month of low volatility and September as a slip with wider swings. It named four drivers: a seasonal slowdown in training demand in August, new GPU clusters coming online across North America, Europe and the Middle East, budget pauses by some users, and buyers looking ahead to Blackwell systems [2]. It identified 2.13 to 2.15 as a possible short-term floor [2].

The tier figures show that "the H100 price" was three prices. Silicon Data's December analysis placed hyperscaler H100 rates at $6.20 to $6.64 per GPU-hour from July to December 2025, neocloud rates at $3.01 to $3.40 and marketplace rates at $1.92 to $2.00 [3]. The hyperscaler tier had fallen to $6.94 in June, from a January to May range of $8.73 to $9.30 [3]. Silicon Data linked the reset to AWS's June price cut and described the result as a flip "from shortage-driven pricing to oversupply-driven pricing" [3].

Contract evidence pointed the other way. On its third-quarter call, CoreWeave said: "in Q3, we saw our first 10,000-plus H100 contract approaching expiration. Two quarters in advance, the customer proactively re-contracted for the infrastructure at a price within 5% of the original agreement" [5]. It also said it signed "a number of deals for older generations of GPUs, adding new customers and recontracting existing capacity" in the quarter [5]. One renewal is one data point, and its original price is not public. It shows that at least one large Hopper commitment renewed close to its earlier price while posted rates were falling.

SemiAnalysis's contract index, built from monthly surveys of more than 100 market participants and checked against transaction data, places the one-year H100 contract low at $1.70 per GPU-hour in October 2025 [6]. The posted indexes and the contract index measure different things, so their levels should not be compared directly.

2. Blackwell supply reached volume

Sources: [7, 8].

For the quarter ended 27 July, NVIDIA said Blackwell data center revenue grew 17% sequentially, and its chief executive said "production of Blackwell Ultra is ramping at full speed, and demand is extraordinary" [7]. It reported no H20 sales to China-based customers in that quarter and assumed none in its outlook of $54.0 billion, plus or minus 2%, for the next quarter [7]. For the quarter ended 26 October, which overlaps August to October, NVIDIA reported data center revenue of $51.2 billion, up 25% from the prior quarter [8].

Deployments of the newest systems began in the quarter. On 3 July, CoreWeave announced that it had deployed GB300 NVL72 systems for customers, in a deployment built with Dell, Switch and Vertiv, and that the systems joined a Blackwell fleet that already included HGX B200 and GB200 NVL72 [9]. At the flagship Stargate site in Abilene, Texas, the partners said Oracle "began delivering the first NVIDIA GB200 racks in June" and that the site was running on Oracle Cloud Infrastructure [10]. On its third-quarter call, CoreWeave described continuing to deliver "initial scaled deployments of the GB200s" during the quarter [5].

3. Capacity contracts grew larger and longer

Four patterns run through the table.

Commitments are conditional maximums. Nebius's filing makes the contract value of about $17.4 billion subject to the deployment and availability of the GPU services [16]. CoreWeave's Meta and NVIDIA filings use similar language [18, 23]. The NVIDIA and OpenAI agreement was a letter of intent, with details to be finalized [19]. Reported totals describe what buyers may pay if suppliers deliver on time.

Terms run five to ten years. The neocloud agreements with Microsoft and Meta run to 2031 [16, 23]. The Fluidstack leases at TeraWulf and Cipher Mining run ten years, with extension options [14, 22]. CoreWeave's NVIDIA order runs into 2032 [18].

Third parties stand behind payments. Google backstopped Fluidstack's lease obligations to support project debt, about $3.2 billion at TeraWulf and $1.4 billion at Cipher Mining, and received warrants for about 14% and 5.4% of those companies on a pro forma basis [15, 22]. NVIDIA, which supplies CoreWeave's GPUs, agreed to buy capacity that CoreWeave's customers leave unsold [18]. In both cases, a party with a stronger balance sheet than the borrower carries part of the demand risk.

Concentration fell as the book grew. CoreWeave's revenue backlog rose from $30.1 billion at 30 June to $55.6 billion at 30 September [36, 24]. On its third-quarter call, it said no single customer represented more than about 35% of backlog, down from about 50% the quarter before and about 85% at the start of the year, and that more than 60% of backlog was tied to investment-grade customers [5].

Not every transaction closed. Core Scientific's stockholders rejected the CoreWeave merger at a special meeting on 30 October 2025, with about 20.8 million votes for and 203.5 million against, and Core Scientific terminated the agreement the same day [13].

Oracle's results put a number on contracted demand at a hyperscale cloud. Its remaining performance obligations reached $455 billion at 31 August, up 359% from a year earlier, after it "signed four multi-billion-dollar contracts with three different customers" in the quarter [11]. Oracle forecast cloud infrastructure revenue of $18 billion for its fiscal 2026, rising to $32 billion, $73 billion, $114 billion and $144 billion over the following four years, and said most of that revenue was already booked in its reported obligations [11].

4. Financing followed the contracts

CoreWeave said its July notes priced 25 basis points below its inaugural high-yield offering in May, and that it had secured $14 billion of debt and equity in 2025 by the time of its third-quarter call [5]. Its second-quarter offering had been $2 billion of 9.25% notes due 2030 [36]. The September loan tranche finances equipment for "both investment grade and unrated strategic customers" [25].

Nebius stated that capital spending for the Microsoft agreement would be funded through "a combination of cash flows under the arrangement and the issuance of debt secured against the contract" [17]. This is the contract-first model in one sentence: a signed contract with a creditworthy buyer becomes the collateral for the capacity that serves it.

Oracle's $18 billion of notes, with maturities out to 2065, followed a fiscal quarter (to 31 August) in which its capital expenditures were $8.5 billion, against $2.3 billion a year earlier, and its trailing four-quarter free cash flow was negative $5.9 billion [26, 11].

5. Capital spending kept rising

Sources: Alphabet [27, 28]; Microsoft [31, 32]; Meta [29, 30]; Amazon [33, 34].

Each company tied the increase to demand. Alphabet linked its July increase to "strong and growing demand for our Cloud products and services" [27]. Meta said in July that it expected "another year of similarly significant capital expenditures dollar growth in 2026", and in October that growth would be "notably larger in 2026 than 2025" [29, 30]. Microsoft said roughly half of its September-quarter spending went to short-lived assets, "primarily GPUs and CPUs", with $11.1 billion of finance leases "primarily for large datacenter sites" [32]. Amazon said in July that second-quarter spending would be "reasonably representative of our quarterly capital investment rate for the back half of this year", and in October that its full-year figure would be about $125 billion and would increase in 2026 [33, 34].

Backlogs grew alongside spending. Amazon reported an AWS backlog of $195 billion at 30 June and $200 billion by the end of September [33, 34]. Alphabet reported Google Cloud backlog of $155 billion at the end of September [28]. Microsoft said Azure surpassed $75 billion in annual revenue in its fiscal 2025 [31].

6. Power set the pace

Sources: [35, 31, 33, 32, 34].

PJM cleared 134,311 MW in its July auction at a price equal to its new cap, with cleared volume 139 MW above the projected reliability requirement [35]. It attributed demand growth "largely" to data center expansion, electrification and economic growth [35]. Amazon's chief executive said in July that resolving capacity shortages would take "several quarters" [33]. In October, Microsoft's chief financial officer described the constraint in physical terms: "we were short the space or the power, is the language we use, to put them in" [32].

Implications

For buyers of compute. A single H100 price does not exist. In the second half of 2025, hyperscaler, neocloud and marketplace medians differed by more than a factor of three [3], and at least one large contract renewal did not follow posted rates down [5]. A buyer comparing offers needs the tier, the term, the commitment size and the renewal terms. Buyers seeking new capacity in this period were competing with commitments that ran five years or more [16, 23].

For builders and operators. The quarter's deals show the contract-first sequence in practice: a buyer signs a conditional, multi-year commitment; the operator raises debt against it; capacity is then built at a named site [16, 17]. The conditions matter. Payments depend on delivery and availability, so construction delays move revenue. The Core Scientific vote also shows that acquiring power and sites can depend on parties outside the operator's control [13].

For lenders. Credit support came from outside the borrower. Google's backstops and NVIDIA's commitment to buy unsold capacity both put a stronger balance sheet behind part of the risk [15, 22, 18]. Contracted values are maximums with delivery conditions, so the cash flow behind a facility depends on construction schedules as well as on the buyer's credit. Customer concentration also matters: CoreWeave's disclosure that no customer exceeded about 35% of backlog compared with about 85% at the start of the year [5].

For researchers. Posted-price indexes and contract-price indexes answered different questions in this quarter and gave different signals. Studies of GPU rental prices should state which kind of price they use, which provider tier, and how they treat renewals.

Limits of this analysis

• This review uses public sources only. It contains no transaction-level price data and no data collected by Rillor.
• Index values are reported as their publishers stated them. Silicon Data's September commentary gives index levels without units; its launch coverage defines the index in dollars per GPU-hour [1, 2]. We did not recompute any index.
• The CoreWeave H100 renewal is a single reported example. The original contract price is not public.
• Contract values are maximums that depend on delivery and other conditions. The NVIDIA and OpenAI agreement was a letter of intent at announcement.
• Capital expenditure measures differ by company. Amazon's figures come from earnings call statements reproduced by transcript services, which can contain transcription errors [33, 34].
• Fiscal periods do not match the calendar quarter. NVIDIA's quarters, Oracle's fiscal first quarter and Microsoft's fiscal year are labeled where used.
• Two facts come from sources published after the quarter: Silicon Data's December tier analysis and SemiAnalysis's April 2026 contract index [3, 6]. They are used only for what they state about this period.
• This is a review of public information. It is not investment advice and makes no forecast of prices.

Sources

1. Samuel K. Moore, Price Index Could Clarify Opaque GPU Costs for AI. IEEE Spectrum, 28 May 2025. https://spectrum.ieee.org/gpu-prices
2. Carmen Li, H100 Rental Market Cools in September: Price Index Slips as Volatility Rises. Silicon Data, 20 October 2025. https://www.silicondata.com/blog/h100-rental-market-update-september-2025
3. Carmen Li, H100 Rental Price Over Time (2023-2025): A Complete Market Analysis. Silicon Data, 21 December 2025. https://www.silicondata.com/blog/h100-rental-price-over-time
4. Pricing and usage model updates for Amazon EC2 instances accelerated by NVIDIA GPUs. Amazon Web Services, 5 June 2025. https://aws.amazon.com/about-aws/whats-new/2025/06/pricing-usage-model-ec2-instances-nvidia-gpus/
5. CoreWeave, Inc. Q3 2025 Earnings Call, corrected transcript. CoreWeave, Inc., 10 November 2025. https://s205.q4cdn.com/133937190/files/doc_financials/2025/q3/CORRECTED-TRANSCRIPT_-CoreWeave-Inc-CRWV-US-Q3-2025-Earnings-Call-10-November-2025-5_00-PM-ET.pdf
6. Daniel Nishball, Jordan Nanos, Cheang Kang Wen and others, The Great GPU Shortage: Rental Capacity, Launching our H100 1 Year Rental Price Index. SemiAnalysis, 2 April 2026. https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
7. NVIDIA Announces Financial Results for Second Quarter Fiscal 2026. NVIDIA Corporation, 27 August 2025. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2026
8. NVIDIA Announces Financial Results for Third Quarter Fiscal 2026. NVIDIA Corporation, 19 November 2025. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2026
9. CoreWeave Becomes First Cloud Provider to Deploy NVIDIA GB300 NVL72 Platform. CoreWeave, Inc., 3 July 2025. https://www.coreweave.com/news/coreweave-becomes-first-cloud-provider-to-deploy-nvidia-gb300-nvl72-platform
10. OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites. SoftBank Group Corp., 24 September 2025 (dateline 23 September 2025). https://group.softbank/en/news/press/20250924
11. Oracle Announces Fiscal Year 2026 First Quarter Financial Results. Oracle Corporation, 9 September 2025. https://investor.oracle.com/investor-news/news-details/2025/Oracle-Announces-Fiscal-Year-2026-First-Quarter-Financial-Results/default.aspx
12. CoreWeave to Acquire Core Scientific (Exhibit 99.1 to Form 8-K). CoreWeave, Inc., 7 July 2025. https://www.sec.gov/Archives/edgar/data/1769628/000095010325008496/dp231293_ex9901.htm
13. Core Scientific shareholders reject merger with CoreWeave, agreement terminated. Investing.com, 31 October 2025. https://www.investing.com/news/sec-filings/core-scientific-shareholders-reject-merger-with-coreweave-agreement-terminated-93CH-4325472
14. TeraWulf Signs 200+ MW, 10-Year AI Hosting Agreements with Fluidstack (Exhibit 99.1 to Form 8-K). TeraWulf Inc., 14 August 2025. https://www.sec.gov/Archives/edgar/data/1083301/000110465925078084/tm2523008d2_ex99-1.htm
15. Press release on Fluidstack's exercise of its expansion option at Lake Mariner (Exhibit 99.1 to Form 8-K). TeraWulf Inc., 18 August 2025. https://www.sec.gov/Archives/edgar/data/1083301/000110465925079463/tm2523651d3_ex99-1.htm
16. Report of Foreign Private Issuer on Form 6-K (agreement with Microsoft). Nebius Group N.V., 8 September 2025. https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm
17. Press release on the closings of the public offering of Class A shares and the private offering of convertible notes (Exhibit 99.1 to Form 6-K). Nebius Group N.V., 15 September 2025. https://www.sec.gov/Archives/edgar/data/1513845/000110465925089969/tm2525580d4_ex99-1.htm
18. Current Report on Form 8-K (order form with NVIDIA, event of 9 September 2025). CoreWeave, Inc., 15 September 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm
19. OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems. NVIDIA Corporation, 22 September 2025. https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems
20. CoreWeave Expands Agreement with OpenAI by up to $6.5B. CoreWeave, Inc., 25 September 2025. https://investors.coreweave.com/news/news-details/2025/CoreWeave-Expands-Agreement-with-OpenAI-by-up-to-6-5B/default.aspx
21. Current Report on Form 8-K (order form with OpenAI, event of 23 September 2025). CoreWeave, Inc., 25 September 2025. https://www.sec.gov/Archives/edgar/data/1769628/000119312525216497/d17274d8k.htm
22. Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack (Exhibit 99.1 to Form 8-K). Cipher Mining Inc., 25 September 2025. https://www.sec.gov/Archives/edgar/data/1819989/000095010325012168/dp234624_ex9901.htm
23. Current Report on Form 8-K (order form with Meta, event of 25 September 2025). CoreWeave, Inc., 30 September 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000050/crwv-20250925.htm
24. CoreWeave Reports Strong Third Quarter 2025 Results (Exhibit 99.1 to Form 8-K). CoreWeave, Inc., 10 November 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000059/coreweave3q25earningspress.htm
25. Current Report on Form 8-K (incremental delayed draw term loan tranche, event of 29 September 2025). CoreWeave, Inc., 2 October 2025. https://www.sec.gov/Archives/edgar/data/1769628/000119312525227562/d910811d8k.htm
26. Current Report on Form 8-K (issuance of $18 billion aggregate principal amount of notes). Oracle Corporation, 26 September 2025. https://www.sec.gov/Archives/edgar/data/1341439/000119312525220445/d31817d8k.htm
27. Alphabet Announces Second Quarter 2025 Results (Exhibit 99.1 to Form 8-K). Alphabet Inc., 23 July 2025. https://www.sec.gov/Archives/edgar/data/1652044/000165204425000056/googexhibit991q22025.htm
28. Alphabet Announces Third Quarter 2025 Results (Exhibit 99.1 to Form 8-K). Alphabet Inc., 29 October 2025. https://www.sec.gov/Archives/edgar/data/1652044/000165204425000087/googexhibit991q32025.htm
29. Meta Reports Second Quarter 2025 Results. Meta Platforms, Inc., 30 July 2025. https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Reports-Second-Quarter-2025-Results/default.aspx
30. Meta Reports Third Quarter 2025 Results (Exhibit 99.1 to Form 8-K). Meta Platforms, Inc., 29 October 2025. https://www.sec.gov/Archives/edgar/data/1326801/000162828025047114/meta-09302025xexhibit991.htm
31. Microsoft Fiscal Year 2025 Fourth Quarter Earnings Conference Call. Microsoft Corporation, 30 July 2025. https://www.microsoft.com/en-us/investor/events/fy-2025/earnings-fy-2025-q4
32. Microsoft Fiscal Year 2026 First Quarter Earnings Conference Call. Microsoft Corporation, 29 October 2025. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q1
33. Amazon.com Q2 2025 earnings call transcript (call of 31 July 2025). VectorShift Research. https://vectorshift.ai/research/companies/amazon-com/earnings/2025-q2
34. Amazon (AMZN) Q3 2025 Earnings Call Transcript (call of 30 October 2025). The Motley Fool, 31 October 2025. https://www.fool.com/earnings/call-transcripts/2025/10/31/amazon-amzn-q3-2025-earnings-call-transcript/
35. PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal. PJM Interconnection, 22 July 2025. https://insidelines.pjm.com/pjm-auction-procures-134311-mw-of-generation-resources-supply-responds-to-price-signal/
36. CoreWeave Reports Strong Second Quarter 2025 Results (Exhibit 99.1 to Form 8-K). CoreWeave, Inc., 12 August 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000039/coreweave2q25earningspress.htm
37. Current Report on Form 8-K (DDTL 3.0 credit agreement, event of 28 July 2025). CoreWeave, Inc., 31 July 2025. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000033/crwv-20250728.htm