Twelve months ago, the consensus on GPU cloud was a glut. It did not happen.
Twelve months ago, the consensus on GPU cloud was a glut. Blackwell was ramping, the H100 was a generation behind the frontier, and most forecasts had rental rates sliding toward $2 an hour. It did not happen. In the week of September 28, 2026, the median on-demand H100 rented for $3.42 per GPU-hour across 40 providers, 15 percent more than a year earlier, according to GetDeploying’s GPU price tracker. The B200 benchmark is up 27.6 percent year to date. And on October 5, CME Group plans to begin trading futures on GPU rental prices, pending regulatory review.
That last detail says more about the market than any price chart. Nobody builds a futures contract for a commodity that is getting reliably cheaper. Exchanges build them when buyers need protection from prices that move, and in 2026 GPU prices have mostly moved up.
This report covers where GPU cloud stands at the end of Q3 2026: what capacity actually costs, why record spending has not produced a surplus, how the neoclouds are doing beneath the headlines, and what infrastructure buyers should do with their Q4 and 2027 contracts.
GetDeploying records weekly on-demand pricing across 87 providers and 103 GPU models. Its September 28 update shows cloud GPU rental prices up 13 percent over 12 months. Of the 22 models priced at both ends of that window, 16 rose and 6 fell. Over the past four weeks alone, the median model moved up another 2.8 percent.
Independent benchmarks agree on direction. Silicon Data’s B200 rental index, the reference the new CME contract will settle against, entered September at $5.62 per GPU-hour, up 27.6 percent year to date, after an unusually calm August in which every daily reading sat between $5.58 and $5.69. Contract pricing tells the same story. SemiAnalysis reported in March 2026 that one-year H100 rental contracts rose almost 40 percent, from $1.70 per GPU-hour in October 2025 to $2.35.
The most important signal is at the bottom of the stack, not the top. Hardware is supposed to depreciate. Instead, SemiAnalysis found H100 contracts being renewed at the same rates they were signed at two to three years ago, some running through 2028, and half the providers it asked could not supply even 64 Hopper GPUs. Coverage of CoreWeave’s Q2 2026 earnings call reported management saying that older-generation systems now command prices at or above what they fetched years ago.
None of this surprises operators who own fleets. Sunny Smith, founder and CTO of Massed Compute, described the same pattern on Episode 7 of the DataStorage.com Podcast: GPUs have a commercial tail of five to seven years, and V100s still rent at close to full capacity. The practical implication for buyers is uncomfortable. Waiting for the H100 to get cheap because Blackwell exists is a strategy the 2026 market has not rewarded.
The spread between provider types is still wide enough to pay for a procurement team. GetDeploying’s late-September medians show AWS, Azure and Google Cloud charging 58 percent more than dedicated GPU clouds for an H100, 81 percent more for a B200 and 141 percent more for an H200. Silicon Data’s separate hyperscaler H100 index barely moved through spring 2026, holding near $7.48 per GPU-hour, which tells you hyperscaler GPU pricing is set by list price and enterprise agreement, not by market conditions. The comparison table below lays out the full gap.
| GPU | GPU clouds | Hyperscalers | Premium |
|---|---|---|---|
| NVIDIA A100 | $2.00 | $3.67 | +84% |
| NVIDIA H100 | $5.00 | $7.89 | +58% |
| NVIDIA H200 | $4.50 | $10.85 | +141% |
| NVIDIA B200 | $7.88 | $14.24 | +81% |
The supply side is spending at a scale without precedent. NVIDIA reported fiscal Q2 2027 revenue of $96.2 billion on August 26, up 106 percent year on year, with data center revenue of $89.0 billion, up 117 percent. It guided Q3 to $108.0 billion and said Vera Rubin is now in full production. Futurum Research’s breakdown puts $40.3 billion of that data center revenue with AI clouds, industrial and enterprise customers rather than hyperscalers, a measure of how much hardware the neocloud tier is absorbing.
On the buy side, Amazon, Microsoft, Alphabet and Meta were already tracking toward roughly $725 billion of combined 2026 capital expenditure after Q1 guidance, up 77 percent on 2025. Alphabet then raised again in July, to between $195 billion and $205 billion for the year, after spending $44.9 billion in Q2 alone.
Demand is still outrunning it. Google Cloud grew 82 percent year on year to $24.8 billion in Q2 and ended June with a $514 billion backlog. CFO Anat Ashkenazi said Alphabet would expand its use of third-party capacity in Q3 as a bridge while it builds its own. Read that twice: one of the largest compute owners on earth is renting from the same neocloud market you are buying from.
Component costs compound the squeeze. Microsoft attributed about $25 billion of its roughly $190 billion 2026 capex figure to higher component pricing, and Meta cited the same pressure when it lifted its range to $125 billion to $145 billion. TrendForce expects DRAM contract prices to rise another 13 to 18 percent quarter on quarter in Q3 2026, and NAND flash 10 to 15 percent. Those costs land in every new GPU server and, eventually, in rental rates. This is the core of our standing thesis: a cheaper token is not a cheaper bill, and a newer GPU is not a cheaper cluster.
The two largest listed neoclouds reported in mid-August, and both results cut two ways.
CoreWeave posted Q2 revenue of $2.575 billion, up from $1.212 billion a year earlier, and a revenue backlog of about $104 billion, before more than $25 billion of new commitments signed in early Q3. Active power grew by roughly 500 megawatts to about 1.5 gigawatts, with 3.7 gigawatts contracted. The same income statement shows $640 million of net interest expense against a $626 million net loss. The interest bill is, almost exactly, the loss.
Nebius grew revenue 454 percent to $582.3 million and exited the quarter at $3 billion of annualized run-rate revenue. It closed four deals averaging more than $1 billion each, with customer prepayments covering 50 to 60 percent of related capex on most new deals, and ran a capacity auction that cleared 15 percent above its previous Blackwell price highs. New deals carry $20 million to $25 million of annual contract value per megawatt, against roughly $12 million across its existing 2026 base.
Two conclusions follow for buyers. First, providers with owned, energized capacity have pricing power, and they are using it. Second, their growth rests on debt secured against GPUs and on a short list of very large customers. That concentration became visible on July 1, when Bloomberg reported that Meta is building Meta Compute to sell surplus AI capacity to outside customers. CoreWeave and Nebius, both Meta suppliers, fell by double digits in a single session.
Meta Compute, together with the SpaceX and xAI capacity deals reported in May, is the clearest sign that new sellers are entering the market. It has not changed prices yet. Meta has published no launch date and no rate card, and its 6 gigawatt AMD Instinct agreement only begins shipping in the second half of 2026. Treat it as negotiating leverage for 2027 renewals, not as supply you can buy this quarter.
The vetting rule has not changed either. Ask whether a provider owns the GPUs it is selling. In a tight market, brokers are the first to lose allocation when an owner finds a better-paying customer, and the support chain gets longer at exactly the moment you need it short.
CME Group and Silicon Data announced on August 11 that they will list two contracts on NYMEX from October 5, 2026, pending regulatory review: H100 and B200 rental index futures, each representing one month of rental for a single GPU and settling against Silicon Data’s daily indices.
For finance and procurement teams, this matters in three ways. First, there is now a public reference price to hold invoices against. Two companies buying equivalent capacity at very different rates is no longer invisible. Second, a team can protect a 2027 inference budget against further price rises without signing a multi-year provider commitment. Third, the forward curve becomes a readable forecast. Silicon Data noted in September that its B200 term curve sits below spot, meaning the market already prices some easing once Rubin volume arrives.
Two cautions. New contracts usually launch with thin liquidity, so early prices can mislead. And a futures hedge protects a budget, not an allocation. It will not put a single GPU in your rack if your provider runs out.
GPU scarcity has a direct storage consequence. When allocation shifts between providers, workloads follow the GPUs and data has to follow the workloads. Checkpoints, training sets and embedding stores now run to hundreds of terabytes, and every move out of a hyperscaler is metered.
Run the numbers. At AWS’s published first-tier internet egress rate of $0.09 per GB, moving 500 TB costs about $45,000 before volume tiers. A $0.50 per GPU-hour price advantage on a 64-GPU cluster saves about $23,360 a month. One data move eats roughly 1.9 months of the savings that justified switching. Egress is still the toll booth of the cloud, and a volatile GPU market means you pass through it more often.
The storage hardware itself is getting dearer too. Smith said on the same podcast episode that NVMe prices have roughly tripled and that scarcity should last until the end of 2027, and TrendForce’s Q3 NAND forecast points the same way. Storage adjacent to GPUs is no longer a casual add-on in any cluster quote.
The architectural answer is to keep the system of record on storage that does not tax exit. Backblaze B2, Wasabi and Cloudflare R2 are S3-compatible and charge little or nothing to move data out, which lets you chase GPU availability without paying a toll each time. This is less about any single vendor than about keeping compute portable in a market where the cheapest GPU changes location every quarter.
Five moves make sense before 2027 budgets lock.
Before signing, compare the rate against Silicon Data’s indices, GetDeploying’s medians and the DataStorage.com GPU Price Explorer. If a quote sits well above the neocloud median with no clear reason (networking, storage, support tier), ask for the reason in writing.
Commitment discounts are large. GetDeploying’s September medians put a 1-year reservation at 29 percent below on-demand and a 3-year term at 49 percent below. On a 64-GPU H100 cluster at the $3.42 median, a year of on-demand costs about $1,917,389. A 1-year reservation brings that to about $1,361,346, a saving of $556,043. A 3-year term gets the effective rate to $1.74 per GPU-hour, but locks you in through 2029 while Rubin ramps and the forward curve already points lower. For most teams the sensible default is 1-year terms for the steady baseline, on-demand or spot for bursts, and nothing longer without a price reset clause.
A 58 to 141 percent premium is justified when you need tight integration with managed services, existing enterprise discounts or strict compliance boundaries. It is not justified for batch training or self-managed inference that could run anywhere.
Ask who owns the hardware, who operates the data center and how the fleet is financed. A provider carrying heavy GPU-backed debt and a few giant customers can reprioritize your capacity if one of those customers changes course.
Keep the canonical copy of training data and checkpoints on low or zero-egress object storage, and replicate to the GPU provider’s local NVMe only for the duration of a job. That single decision determines whether you can act on the next price gap or watch it go by.
The GPU glut was a forecast. The price index is a fact. Build your 2027 compute budget around the fact.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds