Looking back at January to March 2026, the public record shows scarcity returning to the GPU rental market. One-year H100 contract prices rose almost 40% from their October low, and on-demand capacity sold out across GPU generations. Memory prices nearly doubled in the quarter, and the large platforms raised spending plans while selling large amounts of debt.
This review covers rental prices, accelerator supply, memory, financing, capital spending and power. Every figure comes from a public source listed in the full report.
Rental capacity sold out
SemiAnalysis launched a one-year H100 contract price index in April, built from monthly surveys of more than 100 market participants and checked against transaction data [1]. It shows one-year rates breaking above $2 per GPU-hour by late January, rising 15% to 20% by mid-to-late February, and reaching $2.35 by March, almost 40% above the October 2025 low of $1.70 [1].
The same report described a market with little to sell. "On-Demand GPU rental capacity is sold out across all GPU types," it said, and by March it was "increasingly impossible" to find H100, H200 or B200 capacity for any term [1]. For 64-GPU requests, half of the providers it asked were completely sold out [1]. H100s were being renewed "at the exact same rate they were signed at 2-3 years ago", some for four years through 2028 [1]. Blackwell lead times stretched into June and July, and capacity coming online through August and September 2026 was already booked [1].
Posted indexes: B200 moved, hyperscaler H100 did not
| Silicon Data series | Reading | Source |
|---|---|---|
| B200 index, 1 January | 4.40 | [3] |
| B200 index, February range | 4.41 to 4.71 | [3] |
| B200 index, 25 March high | 6.11 | [3] |
| B200 index, 30 March | 5.48, up 24.4% from 1 January | [3] |
| H100 hyperscaler, March | $7.43 to $7.52 per GPU-hour | [3] |
| H100 neocloud, March | $2.43 to $2.63 per GPU-hour, about 8% higher | [3] |
Silicon Data's B200 index held near its January level until March, then rose sharply in the second half of the month [3]. Silicon Data linked the move to higher HBM3e memory costs passing through to cloud rates, to NVIDIA's GTC conference, to a run of frontier model releases that absorbed idle B200 capacity, and to agentic inference workloads that favor on-demand capacity [3]. Its hyperscaler H100 series barely moved: from 1 March to 20 April it stayed between $7.40 and $7.52 and was unchanged on 39 of 50 days, at about three times the neocloud tier [4].
Rubin entered production
On 5 January, NVIDIA said Rubin was "in full production" and that Rubin-based products would be available from partners in the second half of 2026 [5]. Its results for the quarter ended 25 January showed data center revenue of $62.3 billion, up 75% from a year earlier, a multiyear partnership with Meta covering "millions of NVIDIA Blackwell and Rubin GPUs", and an outlook of $78.0 billion for the next quarter [6]. At GTC in March, NVIDIA said members of its NVIDIA Cloud Partner program had deployed more than 1 million GPUs, representing more than 1.7 GW of AI capacity, up from 550 MW a year earlier [7].
Other suppliers signed volume too. On 24 February, AMD disclosed a warrant for up to 160 million shares to Meta, vesting as Meta buys up to 6 GW-equivalent of AMD Instinct GPUs [9]. Amazon said Trainium2 was fully subscribed with 1.4 million chips landed, and that nearly all Trainium3 supply was expected to be committed by mid-2026 [10].
Memory prices nearly doubled
On 2 February, TrendForce raised its forecast for first-quarter conventional DRAM contract prices to a rise of 90% to 95%, from 55% to 60% [27]. The outcome was 93% to 98%, and DRAM industry revenue reached $97 billion, up 81% from the prior quarter [28]. TrendForce linked the demand to cloud data centers adding general-purpose servers as AI work shifts toward inference [28]. SemiAnalysis reported that server repricing led some operators to "slow-roll or abandon deployments" [1].
Platforms borrowed, neoclouds refinanced
| Date (2026) | Issuer | Transaction |
|---|---|---|
| 26 January | NVIDIA and CoreWeave | $2 billion equity investment at $87.20 per share; plan for more than 5 GW of AI factories by 2030 |
| 2 to 5 February | Oracle | $25 billion of notes, an at-the-market program for up to $20 billion of shares, and mandatory convertible preferred stock |
| 13 February | Alphabet | $20 billion and £5.5 billion of notes, including a sterling tranche due 2126 |
| 27 February | Amazon and OpenAI | $50 billion Amazon investment in OpenAI; AWS agreement expanded by $100 billion over eight years, with about 2 GW of Trainium |
| 13 and 16 March | Amazon | About $37 billion and €14.5 billion of notes |
| 31 March | CoreWeave | $8.5 billion delayed draw term loan rated A3 by Moody's |
The CoreWeave loan stands out. It is non-recourse and secured by the assets of one subsidiary, and CoreWeave described it as financing secured by HPC infrastructure and an associated customer contract. Its floating tranche is priced at SOFR plus 2.25% [14]. CoreWeave's DDTL 3.0 loan, closed in the third quarter of 2025, had priced at SOFR plus 4% [31].
Spending plans reset higher
| Company | 2025 actual | 2026 guidance, January and February | Quarter to March 2026 |
|---|---|---|---|
| Alphabet | $91.4 billion | $175 to $185 billion | $35.7 billion |
| Amazon | $131.8 billion | About $200 billion | $44.2 billion |
| Meta | $72.2 billion | $115 to $135 billion | $19.8 billion |
| Microsoft | Fiscal year ends June | Lower in the March quarter than in December | $31.9 billion |
Measures differ by company, and the full report defines each [21, 22, 23, 24, 25, 26, 10, 11]. When reporting first-quarter results on 29 April, Meta raised its 2026 range to $125 billion to $145 billion, citing "higher component pricing", and Microsoft guided to roughly $190 billion for calendar 2026, including about $25 billion from higher component prices [25, 26].
Power stayed binding
On 16 January, PJM's board set out how it would connect large loads such as data centers, including letting new loads bring their own generation or accept earlier curtailment, and starting a backstop procurement of generation [30]. Alphabet's chief executive said he expected to go through 2026 "in a supply constrained way", naming power, land and supply chain [23]. In April, Microsoft said it expected to stay capacity constrained at least through 2026 [26].
What the quarter means
For buyers, the 2025 pattern reversed: older GPUs became harder to find, and renewals held at their original prices. Posted hyperscaler rates told buyers little about the market that cleared. For builders, memory became a budget line large enough to change capital plans. For lenders, the CoreWeave facility showed a non-recourse loan, secured by GPU infrastructure and a customer contract, earning investment-grade ratings.
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 32 numbered sources, is available as a PDF.