AWS raises Capacity Blocks 15% as September GPU list prices rose 15 times with no cuts

September saw 15 rises and no cuts

A FastGPU analysis published on dev.to compared 176 posted on-demand GPU rates — each cloud's own published price for each card it rents — between the first three and the last three days of September 2026. Fifteen ended the month higher, none ended lower, and the remaining 161 did not move. Marketplace listings, where individual hosts pick their own prices, were left out of the count.

All fifteen increases came from two providers: ten from Verda and five from RunPod.

The cheapest offers moved in both directions

The floor price — the lowest posted on-demand rate outside the hyperscalers — is set by only one or two providers at any moment, so it swings. Measured as the geometric mean of the first three days against the last three:

• H100 SXM: $1.34 to $1.57 per hour, up 17.4% (two providers)
• H200: $2.60 to $2.16, down 17.1% (two providers)
• B200: $3.69 to $3.69, unchanged (one provider)
• A100 80GB: $0.64 to $0.53, down 16.7% (one provider)

The gap between sellers was also striking. On September 30, an NVIDIA V100 could be rented for $0.07 per hour at Vast.ai or $2.95 per hour at Oracle Cloud, across eight providers — a 44x spread on identical hardware, the widest recorded that month.

AWS adds a 15% Capacity Blocks increase

Effective October 7, 2026, AWS's pricing page lists these hourly rates per accelerator across all available regions for EC2 Capacity Blocks for ML: P6-B300 at $16.146 ($16.819 in GovCloud), P6-B200 at $14.208 ($14.801 in GovCloud), P5en at $7.895, P5e at $6.866, P5 at $5.970, P4de at $2.546 and P4d at $1.696.

Measured against AWS's earlier published rates, as logged by FastGPU, the change works out to roughly 15%. Because AWS now charges a single rate in every region, block types that previously cost less outside the United States — the P5 and P5en — absorb the largest increases, which FastGPU calculates at up to 26.5% for non-US buyers.

AWS states that Capacity Block reservation prices are dynamic and revised from time to time according to supply and demand, and that all other EC2 Capacity Block prices, along with fixed pricing models such as On-Demand and Savings Plans, remain unchanged.

What renters should note

A Capacity Block is reserved capacity: you reserve a start date and a length of time, and AWS commits the hardware for that window. That is a distinct product from an on-demand instance you can stop at will, so the two price tracks are not directly comparable — but the gap is still instructive. As of October 7, the cheapest posted H100 SXM rate was $1.79 per hour at Vast.ai, the September-end median across sixteen providers was $3.70 per hour, and a P5 Capacity Block now costs $5.970 per hour.

The practical takeaway depends on the workload: short jobs are priced at the floor, but anyone planning a quarter's spend should watch where list rates are heading — and in September that direction was upward.

A caveat on method: FastGPU tracks posted prices, not completed transactions. Its median counts only providers that published a price on all six days examined, so a provider joining or leaving the market cannot register as a price move, and the daily series is released as an open dataset under a CC BY 4.0 licence.

Why it matters

GPU cost is a direct line item in every AI project, from fine-tuning to inference at scale, and September's data shows the direction of travel at established clouds is upward — fifteen increases and zero cuts. The AWS change is the sharper signal: a hyperscaler lifting guaranteed-capacity prices by 15%, and effectively more for customers outside the US, suggests demand for reserved accelerators still outstrips supply. For teams on fixed budgets, small per-hour increases compound quickly across clusters and months, which makes price tracking and reserve-versus-rent arithmetic a routine part of infrastructure planning rather than a one-off exercise.