Looking back at October to December 2025, the public record shows demand for AI compute running into physical limits. Cloud providers signed contracts that named GB300 systems, one of them covering hundreds of thousands of NVIDIA GPUs for a single customer. Operators reported that powered buildings and grid capacity set how fast that capacity could arrive, and memory prices rose sharply.
This review covers rental prices, accelerator supply, contracts, financing, power, memory and capital spending. Every figure comes from a public source listed in the full report.
H100 rents found a floor
SemiAnalysis's contract index, built from monthly surveys of more than 100 market participants, places the low for one-year H100 contracts at $1.70 per GPU-hour in October 2025 [1]. Posted rates turned up late in the quarter. Silicon Data's H100 rental index rose from $2.00 on 9 December to $2.20 on 6 January, a 10% move in four weeks, while its A100 and B200 indexes held steady [2].
Silicon Data tied the move to providers steering allocation toward reserved capacity and long-term contracts, to tight supply in Europe and Asia, and to year-end training deadlines and budget cycles [2]. Spot buyers and smaller buyers, it said, "bore the full pricing pressure" [2]. The spread across providers stayed wide: from about $1.90 per GPU-hour at some specialist providers to over $4 at hyperscalers [2].
GB300 entered contracts by name
On 9 October, Microsoft Azure announced a production cluster of 4,608 Blackwell Ultra GPUs in GB300 systems for OpenAI workloads, toward a goal of hundreds of thousands of Blackwell Ultra GPUs [4]. On 3 November, two contracts named GB300 directly. AWS agreed to supply OpenAI with GB200 and GB300 capacity under a $38 billion, seven-year agreement, with all capacity targeted before the end of 2026 [13]. IREN signed a five-year, $9.7 billion contract to give Microsoft access to GB300 GPUs in Texas, buying the equipment from Dell for about $5.8 billion [15].
NVIDIA reported data center revenue of $51.2 billion for its quarter ended 26 October, and its chief executive said "cloud GPUs are sold out" [5]. Its next quarter, ended 25 January 2026, brought data center revenue of $62.3 billion, up 75% from a year earlier [6].
Buyers also widened their chip suppliers. OpenAI agreed to deploy 6 GW of AMD Instinct GPUs, starting with 1 GW of MI450 in the second half of 2026, and received a warrant for up to 160 million AMD shares that vests with deployments and share-price targets [7]. It also announced 10 GW of OpenAI-designed accelerators built with Broadcom, starting in the second half of 2026 [8]. Anthropic said it would use up to one million Google TPUs, with well over a gigawatt of capacity coming online in 2026 [11].
Contracts and counterparties
| Date (2025) | Parties | Stated value | Scale or term |
|---|---|---|---|
| 21 October | Meta and Blue Owl, Hyperion joint venture | About $27 billion development | Blue Owl 80%, Meta 20% |
| 3 November | AWS and OpenAI | $38 billion | Seven years |
| 3 November | IREN and Microsoft | About $9.7 billion | Five years, 20% prepayment |
| 11 November | Nebius and Meta | About $3 billion | Five years |
| 12 November | Anthropic with Fluidstack | $50 billion investment | Sites in Texas and New York |
| 18 November | Anthropic with Microsoft and NVIDIA | $30 billion of Azure compute | Up to 1 GW more; NVIDIA to invest up to $10 billion, Microsoft up to $5 billion |
Oracle's remaining performance obligations rose by $68 billion in the quarter to November, to $523 billion, highlighted by "new commitments from Meta, NVIDIA, and others" [21].
Financing took new forms
The quarter's financing moved beyond corporate bonds. Meta placed its Hyperion campus in Louisiana in a joint venture with funds managed by Blue Owl, which contributed about $7 billion in cash and took 80%; debt was sold to PIMCO and other investors, and Meta signed leases and a residual value guarantee [9, 10]. Meta also sold $30 billion of bonds on 3 November, and Alphabet sold $17.5 billion and €6.5 billion on 6 November, with maturities to 2065 and 2075 [23, 24].
Neocloud financing changed shape too. CoreWeave sold about $2.59 billion of 1.75% convertible notes due 2031 in December [25]. On 31 December it amended a $2.6 billion term loan, temporarily lowering a minimum liquidity requirement and deferring its first debt service coverage test to October 2027, to match delivery timing it had discussed in November [26]. IREN's Microsoft contract carried a 20% prepayment, so the customer funds part of the build in advance [15].
Buildings and power set the pace
CoreWeave cut its 2025 capital spending guidance to $12 billion to $14 billion after "a third-party data center developer, who is behind schedule" delayed powered shells; it said the affected customer kept the full contract value [17]. In a podcast interview reported on 3 November, Microsoft's chief executive said: "you may actually have a bunch of chips sitting in inventory that I can't plug in" [36].
On 17 December, PJM's 2027/2028 capacity auction cleared at its cap of $333.44 per MW-day and fell short of the reliability requirement by 6,623 MW, the first time the whole region fell short [37]. PJM said nearly 5,100 MW of the 5,250 MW rise in forecast peak load came from data centers [37].
Memory joined the cost stack
TrendForce raised its forecast for fourth-quarter conventional DRAM contract prices in late October, to 18% to 23% from 8% to 13%, as cloud providers increased orders [38]. The final figure was 45% to 50% [39]. In January, Microsoft said rising memory prices would affect its capital expenditures [35].
Capital spending
| Company | Quarter to December 2025 | Full year 2025 |
|---|---|---|
| Alphabet | $27.9 billion | $91.4 billion |
| Microsoft | $37.5 billion | Fiscal year ends June |
| Meta | $22.1 billion | $72.2 billion |
| Amazon | $39.5 billion | $131.8 billion |
Measures differ by company, and the full report defines each [32, 33, 34, 35].
What the quarter means
For buyers, H100 contracts found a floor while GB300 capacity was committed under multi-year contracts and NVIDIA described cloud GPUs as sold out. For builders, schedule risk moved from chips to powered buildings, and delivery terms reached into financing covenants. For lenders, the quarter added structures that spread risk: joint ventures with residual value guarantees, convertible notes, customer prepayments and vendor equity.
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 41 numbered sources, is available as a PDF.