In September we argued that the GPU glut never arrived. October is the month providers wrote that reality into their rate cards.
On October 1, Nebius switched its public price list. An HGX H100 that rented for $3.85 per GPU-hour on September 30 cost $4.50 the next morning, and its H200, B200 and B300 rates rose between 19 and 21 percent on the same day. Four days later Jarvislabs lifted its H100 SXM rate by 30 percent. On October 22 DeepInfra raised every GPU tier it sells, including a 67 percent jump on the A100, a chip that launched in 2020.
In September we argued that the GPU glut everyone forecast for 2026 never arrived. October is the month providers stopped absorbing that reality and wrote it into their rate cards. The transaction indices agree. Silicon Data's neocloud H100 index read $2.88 per GPU-hour on October 24, about 44 percent above its December 2025 reading of $2.00, and the Ornn Compute Price Index settled the H100 SXM at $2.96 on October 28, up 15.6 percent in a month.
This report covers what GPU capacity costs at the end of October 2026, why the biggest buyers in the market are still short, what Vera Rubin's arrival does and does not change, and what to lock in before the next round of repricing.
Ask what an H100 costs in October 2026 and you will get at least three honest answers, each describing a different market.
The first is the trade price. Ornn builds its index only from executed rental deals between buyers and sellers, and it settled at $2.96 per GPU-hour on October 28. Silicon Data's neocloud series, which blends provider and marketplace observations, read $2.88 on October 24. This is roughly what a volume buyer with a term commitment pays a specialist provider today.
The second is the list price. Uvik Software's October index, which tracks published on-demand rates, put the median H100 SXM at $4.20 per GPU-hour across 16 providers on October 7, with a range from $3.20 at Hyperstack and Thunder Compute to $12.29 for the per-GPU equivalent of an Azure ND96isr H100 v5 instance. The list median sits about 42 percent above the trade index.
The third is the hyperscaler price. Silicon Data's hyperscaler H100 series stood at $7.17 on October 24, roughly 2.5 times its neocloud series on the same day. That premium has barely moved all year, because hyperscaler GPU pricing follows list prices and enterprise agreements rather than market conditions.
One caution on sources. Both index publishers have a commercial stake in their numbers being trusted: Silicon Data's indices are the settlement reference for the planned CME futures, and Ornn sells reserved neocloud capacity priced against its own index. Neither gains from a particular price level, and the two series track each other closely, which is the best evidence that the level is real. Read them as the floor a strong negotiator can reach, not the price your first quote will show.
The gap between those three prices is where procurement teams earn their salaries. A 64-GPU H100 cluster running all year costs about $1.66 million at the $2.96 trade index, $2.35 million at the $4.20 list median and $4.02 million at the $7.17 hyperscaler index. The difference between paying the index and paying the list median on that one cluster is about $695,000 a year.
| Price type | Source, date | Per GPU-hour | 64 GPUs, 1 year |
|---|---|---|---|
| Executed trades | Ornn, Oct 28 | $2.96 | $1,659,494 |
| Neocloud index | Silicon Data, Oct 24 | $2.88 | $1,614,643 |
| List median, 16 providers | Uvik, Oct 7 | $4.20 | $2,354,688 |
| Hyperscaler index | Silicon Data, Oct 24 | $7.17 | $4,019,789 |
Uvik's October 7 medians show how steeply prices climb with each generation: $1.62 for the L40S, $2.75 for the A100 80GB, $3.45 for the AMD MI300X, $4.20 for the H100, $5.70 for the H200, $8.19 for the B200 and $9.50 for the B300. The B300 median is effectively Nebius's new on-demand rate, which tells you how thin public Blackwell Ultra supply still is. Live provider-level rates for each of these chips are tracked daily in the DataStorage.com GPU Price Explorer.
Three published increases define the month. Nebius raised its Hopper and Blackwell on-demand rates by 17 to 21 percent effective October 1, with the new column visible on its price list from late September, according to a capture by the pricing tracker UsagePricing. Jarvislabs raised H100 SXM from $2.69 to $3.49 and H200 SXM from $3.99 to $4.59 per GPU-hour from October 5. DeepInfra raised its dedicated GPU rates around October 22, again per UsagePricing's captures of its pricing page.
| Provider, date | GPU | Old to new rate | Change |
|---|---|---|---|
| Nebius, Oct 1 | H100 | $3.85 to $4.50 | +17% |
| Nebius, Oct 1 | H200 | $4.50 to $5.40 | +20% |
| Nebius, Oct 1 | B200 | $7.15 to $8.50 | +19% |
| Nebius, Oct 1 | B300 | $7.85 to $9.50 | +21% |
| Jarvislabs, Oct 5 | H100 SXM | $2.69 to $3.49 | +30% |
| Jarvislabs, Oct 5 | H200 SXM | $3.99 to $4.59 | +15% |
| DeepInfra, Oct 22 | A100 80GB | $0.89 to $1.49 | +67% |
| DeepInfra, Oct 22 | H100 | $2.20 to $2.90 | +32% |
| DeepInfra, Oct 22 | B200 | $3.69 to $4.89 | +33% |
Two details matter more than the headline percentages. First, the oldest silicon rose fastest. DeepInfra's A100 went from $0.89 to $1.49, a 67 percent increase, against 32 to 33 percent for its H100 and B200. That is the same pattern our September report flagged in contract renewals: hardware that should be depreciating is appreciating, because buyers who cannot get Blackwell fall back to whatever is free. Sunny Smith, founder and CTO of Massed Compute, made the same point on Episode 7 of the DataStorage.com Podcast, describing a five to seven year commercial tail in which V100s still rent near full capacity.
Second, the increases came with notice. Jarvislabs told customers that any reserved plan started before October 5 would keep the old rate for its full term. Nebius published its new column a week before it took effect. Announcement windows are the cheapest leverage a buyer gets in this market, and most teams miss them because nobody is watching provider changelogs.
The cost of missing one is easy to quantify. For a 64-GPU H100 deployment on Nebius on-demand, the October 1 change adds about $30,400 a month, or $364,416 a year. The same cluster on B200 absorbs $756,864 a year from a $1.35 per GPU-hour increase.
Late-October earnings removed any doubt about where demand stands. Microsoft reported on October 28 that Azure revenue grew 48 percent to $33.2 billion in its fiscal first quarter. CFO Amy Hood guided Azure to 49 to 50 percent constant-currency growth for the December quarter and told analysts that "some third-party capacity deliveries we expected in Q1 were delayed," according to CNBC's report of the call. Microsoft now expects about $170 billion of capital spending and finance leases in calendar 2027.
Alphabet reported the same day. Google Cloud revenue more than doubled to $30.7 billion, the cloud backlog reached $561 billion according to the earnings call, and quarterly capital expenditure was $51.4 billion. Alphabet held its 2026 capex range at $195 billion to $205 billion and said 2027 spending will rise significantly. Meta, which reported on October 21, spent a record $34.9 billion in the quarter, narrowed its 2026 range to $135 billion to $145 billion, and said on its call that it is supply constrained and is using third-party cloud capacity to add compute in 2026 and 2027.
Two conclusions follow for anyone renting GPUs. The first is contrarian. In July, the Bloomberg report on Meta Compute was read as the start of a supply release that would pressure neocloud prices. Three months later Meta has published no rate card, and on its own earnings call it described itself as a buyer of outside capacity. The seller the market feared in July spent October competing with you for the same racks.
The second is about allocation risk. When the largest software company in the world says third-party capacity deliveries slipped, it is telling you how the neocloud tier rations scarcity. Delayed capacity goes to the customer with the largest contract and the strongest remedies. If your agreement has neither a firm delivery date nor a penalty for missing it, yours is the allocation that moves. This is the practical reason to ask every provider whether it owns the GPUs it is selling: a broker with no fleet of its own cannot reprioritize in your favor.
NVIDIA's next generation reached paying customers in October. At GTC Berlin on October 21, Google Cloud said its first A5X instances, built on Vera Rubin NVL72 rack-scale systems, were running for an initial customer, with general availability planned for November and a stated plan to scale to as many as 960,000 Rubin GPUs across sites. Oracle made Vera Rubin NVL72 and Vera CPUs generally available on Oracle Cloud Infrastructure on October 27. CoreWeave has also claimed to be the first provider with a Rubin system running, according to DatacenterDynamics. NVIDIA claims Rubin delivers 5 times the inference performance and 3.5 times the training performance of Blackwell.
What October did not bring is a public hourly price. Neither the Google Cloud nor the Oracle announcement published a per-GPU-hour rate, and early Rubin capacity is being sold the way large Blackwell clusters were in 2025: as rack-scale commitments negotiated with frontier labs and hyperscale tenants. The unit of purchase is now an NVL72 rack of 72 GPUs, not a single card or an eight-GPU node.
For budget owners, the takeaway is to plan around the generation you can actually rent. Broad on-demand Rubin access through neoclouds is a 2027 story. Waiting for Rubin to pull H100 and B200 rates down has not worked yet, and October's repricing suggests it will not work in the first half of 2027 either.
CME Group and Silicon Data had targeted October 5 for the first H100 and B200 rental index futures on NYMEX. That date passed without a listing. The CFTC extended its review through November 9 after finding the filing raised novel or complex issues, and CME's revised notice says the effective date will be announced later, so November 9 is a decision deadline rather than a launch date. Competing products from ICE with Ornn, and from Nodal Exchange with Compute Desk, are also waiting on regulators. Each CME contract would cover 730 GPU-hours, roughly one month of a single GPU running continuously.
The practical effect is that nobody can hedge Q4 2026 renewals or early 2027 budgets on an exchange. The indices behind the contracts already exist, though, and they are the most useful thing the futures story produced. A public reference price is something you can put in a contract today: an index-linked rate with a cap and a floor gives both sides protection without a futures account.
The components inside every new GPU server are still getting more expensive. In its fourth-quarter forecast published September 30, TrendForce projected conventional DRAM contract prices rising 10 to 15 percent quarter on quarter, 15 to 20 percent once HBM is included, and NAND flash contract prices rising 15 to 20 percent. Smith said on our podcast that NVMe prices have roughly tripled and that scarcity should last until the end of 2027. Those costs reach rental floors with a lag, which is one more reason not to expect relief in early 2027.
Repricing also creates a reason to move workloads, and moving workloads means moving data. Take the 64-GPU H100 cluster again. Shifting it from Nebius's new $4.50 rate to a $3.20 on-demand provider saves about $60,700 a month. If the 500 TB training corpus behind that cluster sits in Amazon S3, pulling it out at AWS's published internet egress tiers ($0.09 per GB for the first 10 TB, $0.085 for the next 40 TB, $0.07 for the next 100 TB and $0.05 above 150 TB) costs about $29,500. That is roughly half a month of the savings, and it is charged again every time capacity forces another move. Egress remains the toll booth of the cloud, and a market where the cheapest GPU changes address every quarter makes you drive through it more often.
The architectural answer has not changed. Keep the system of record for training data and checkpoints on S3-compatible storage that charges little or nothing to read out, such as Backblaze B2, Wasabi or Cloudflare R2, and stage data onto the GPU provider's local NVMe only for the life of a job. Storage is the anchor of the AI stack, and an anchor in the wrong place is what turns a GPU price gap into a missed opportunity.
Every October increase was announced before it took effect, and at least one provider explicitly protected customers who reserved before the change. Assign someone to track the pricing pages of every provider you use. A week of notice on a 64-GPU H100 cluster at Nebius was worth $364,416 a year.
Bring the Ornn and Silicon Data readings to the table along with provider-level rates from the GPU Price Explorer. If a quote sits more than 40 percent above the trade index, ask what the premium buys: networking fabric, storage, support tier or guaranteed delivery. If the answer is none of those, it is margin.
Microsoft's delayed capacity is the warning. Require a firm delivery date, service credits for slippage, and the right to source replacement capacity elsewhere without penalty. Confirm who owns the hardware and who operates the data center before you sign.
Plan first-half 2027 inference and fine-tuning on H100, H200 and B200 at current rates or higher. Treat Rubin as upside for teams with rack-scale commitments, not as a cost reduction in the base case. For steady baseline load, one-year terms remain the sensible default; for bursts, on-demand or spot.
Move the canonical copy of datasets and checkpoints to low or zero-egress object storage now, while nothing is forcing a migration. The next price gap will not wait for a data transfer to finish.
In October the rate cards caught up with the index. Lock in what you can before the next card changes.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
