Looking back at July to September 2025, the public record shows two prices for the same compute moving apart. Posted hourly rates for NVIDIA H100 GPUs drifted lower. In the same three months, cloud providers and AI developers signed multi-year capacity contracts worth billions of dollars.
This review covers rental prices, Blackwell supply, capacity contracts, financing, capital spending and power. Every figure comes from a public source listed in the full report.
Posted H100 rents drifted lower
Silicon Data's H100 rental index (SDH100RT) turns spot rental prices from more than 30 sources into one daily figure in dollars per GPU-hour [1]. It stood at $2.37 when it launched on 27 May [1]. In August it moved between 2.22 and 2.26, with volatility of about 1.7% [2]. In mid-September it fell as low as 2.13 and closed the month near 2.16 [2]. Silicon Data linked the softening to new clusters coming online in North America, Europe and the Middle East, and to buyers looking ahead to Blackwell [2].
The tier a buyer rents from changes the answer. Silicon Data's year-end analysis gave median H100 prices for July to December 2025 of $6.26 per GPU-hour at hyperscalers, $3.33 at neoclouds and $1.95 on marketplaces [3]. Hyperscaler prices had reset in June, when AWS cut on-demand prices for its P5 instances by up to 45%, a move Silicon Data described as an H100 price cut [4, 3].
Contract prices moved less. CoreWeave said that in the third quarter a customer holding a contract for more than 10,000 H100s re-contracted two quarters before expiry, at a price within 5% of the original agreement [5]. SemiAnalysis, which surveys contract prices, later placed the low for one-year H100 contracts at $1.70 per GPU-hour in October 2025 [6].
Blackwell reached volume
NVIDIA reported data center revenue of $41.1 billion for its quarter ended 27 July 2025, up 56% from a year earlier. It said Blackwell data center revenue grew 17% from the prior quarter and that production of Blackwell Ultra was "ramping at full speed" [7]. On 3 July, CoreWeave announced that it had deployed GB300 NVL72 systems for customers, built with Dell, Switch and Vertiv [9]. At the Stargate site in Abilene, Texas, Oracle had been delivering GB200 racks since June [10]. NVIDIA's following quarter, to 26 October, brought data center revenue of $51.2 billion [8].
Contracts grew larger and longer
The quarter's deals tied named buyers to specific sites, terms and delivery conditions.
| Date (2025) | Parties | Stated value | Term or scale |
|---|---|---|---|
| 7 July | CoreWeave to acquire Core Scientific, all stock | About $9.0 billion equity value | About 1.3 GW of gross power |
| July | OpenAI and Oracle | Over $300 billion | Up to 4.5 GW, five years |
| 14 and 18 August | TeraWulf and Fluidstack, Google backstop | $6.7 billion contracted | 10-year leases, New York |
| 8 September | Nebius and Microsoft | About $17.4 billion | Five years, New Jersey |
| 9 September | CoreWeave and NVIDIA | $6.3 billion initial value | NVIDIA buys unsold capacity to April 2032 |
| 22 September | NVIDIA and OpenAI, letter of intent | Up to $100 billion NVIDIA investment | At least 10 GW |
| 23 September | CoreWeave and OpenAI | Up to $6.5 billion | Total with OpenAI about $22.4 billion |
| 25 September | Cipher Mining and Fluidstack, Google backstop | About $3 billion | 168 MW, 10 years |
| 25 September | CoreWeave and Meta | Up to about $14.2 billion | Through December 2031 |
Two features stand out. First, parties outside the buyer stood behind payments. Google backstopped about $3.2 billion of Fluidstack's lease obligations at TeraWulf and $1.4 billion at Cipher, taking warrants for equity in each [15, 22]. NVIDIA agreed to buy any CoreWeave capacity under the order that customers leave unused [18]. Second, the headline values depend on delivery. The Nebius, Meta and NVIDIA agreements make payment subject to delivery and availability of service [16, 18, 23].
Not every deal closed. Core Scientific's shareholders voted down the CoreWeave acquisition on 30 October, and the agreement was terminated [13].
Oracle reported remaining performance obligations of $455 billion at the end of August, up 359% from a year earlier [11]. On 23 September, OpenAI, Oracle and SoftBank said Stargate had nearly 7 GW of planned capacity and over $400 billion of planned investment over three years [10].
Debt followed the contracts
CoreWeave sold $1.75 billion of 9.0% senior notes due 2031 in July, closed a $2.6 billion delayed draw term loan at SOFR plus 4%, and in late September added a $3.0 billion loan tranche at SOFR plus 4.25% [24, 37, 25]. Nebius raised about $4.2 billion in shares and convertible notes and said it would fund the Microsoft build partly with debt secured against the contract [17]. On 26 September, Oracle closed an $18 billion bond sale in six tranches, the longest due 2065 [26].
Capital spending kept rising
| Company | Quarter to June 2025 | Guidance in late July | Quarter to September 2025 |
|---|---|---|---|
| Alphabet | $22.4 billion | About $85 billion for 2025 | $24.0 billion |
| Microsoft | $24.2 billion | Over $30 billion for the next quarter | $34.9 billion |
| Meta | $17.0 billion | $66 to $72 billion for 2025 | $19.4 billion |
| Amazon | $31.4 billion | Second-quarter rate representative of the second half | $34.2 billion |
Each company measures capital spending its own way, and the full report defines each line [27, 28, 29, 30, 31, 32, 33, 34]. By late October, Alphabet had raised its 2025 guidance to $91 billion to $93 billion and Amazon expected about $125 billion [28, 34].
Power set the pace
On Amazon's July call, its chief executive said "the single biggest constraint is power" [33]. Microsoft's chief financial officer expected to stay capacity constrained through the first half of its fiscal year [31]. On 22 July, PJM's capacity auction for 2026/2027 cleared at its cap of $329.17 per MW-day. PJM put the rise in forecast peak load at more than 5,400 MW and attributed demand growth largely to data center expansion, electrification and economic growth [35].
What the quarter means
For buyers, an H100 price needs a tier and a term. Posted rates fell while at least one large contract renewed close to its original price. For builders, the quarter showed a contract-first pattern: capacity is financed after a creditworthy buyer signs, and payment depends on delivery. For lenders, credit support came from outside the borrower, in the form of Google's lease backstops and NVIDIA's commitment to buy unsold capacity.
The Rillor Compute Index methodology v1.0 frames these questions as four series: RCI Rental, RCI Availability, RCI Hardware and RCI Rent-to-Value. This review reports third-party figures only.
Read the full report
The full report, with the method, a table for each thread, the limits of the analysis and all 37 numbered sources, is available as a PDF.