GPU Capacity Block Prices Rose Again: What Leaders Should Do

If your GPU budget assumes reserved capacity has a fixed price, AWS just gave you a reason to reopen the spreadsheet. GPU capacity block prices on Amazon EC2 moved again this week, reportedly the third increase of 2026, and the new rates are live on the AWS Capacity Blocks pricing page. This post separates what AWS confirms from third-party arithmetic, then lists what a leader should do before the next reprice.

What changed in GPU capacity block prices

The pricing page now lists these rates per accelerator-hour: P5 (H100) $5.970, P5e (H200) $6.866, P5en (H200) $7.895, P6-B200 $14.208 and P6-B300 $16.146. The page says reservation prices are "updated regularly based on trends in supply and demand" and are scheduled to be updated next in January 2027. It also says a block is charged at the rate in effect when you buy it, even if the block starts after a price change.

AWS does not publish a changelog, so the October 7 date and the percentages come from a FastGPU post on dev.to. It calculates about 15% in US Regions and up to 26.5% elsewhere, because AWS now lists one rate across Regions (outside the US, P5 went from $4.720 to $5.970). FastGPU sells GPU price comparisons and the "before" rates come from its own log, so treat the percentages as third-party arithmetic. The same post says On-Demand and Savings Plans prices are unchanged.

The third increase in nine months

• January: about 15%. The Register reported p5e.48xlarge going from $34.61 to $39.80 per hour in most Regions. An Amazon spokesperson said the change reflects "the supply/demand patterns we expect this quarter."
• July 1: reported at roughly 20% by secondary outlets citing The Information. I could not confirm this one against an AWS source.
• October 7: about 15%, per the calculation above.

The Register also notes that AWS's earlier GPU price cuts of up to 45% covered On-Demand and Savings Plans, not Capacity Blocks. The product that guarantees you capacity is the one that keeps climbing.

What the increase costs on one small block

Take a four-node P5 block for 48 hours, Linux, reservation fee only. At the earlier US rate of $5.191 per accelerator-hour that is about $7,973. At $5.970 it is about $9,169, roughly $1,200 more. The fee is charged up front when you schedule the reservation, so the increase hits as cash, not as a slow drift.

Re-forecast reserved GPU spend

Price real offerings instead of copying a number from a slide. This command lists offerings and their upfront fee:

aws ec2 describe-capacity-block-offerings \
  --region us-east-2 \
  --instance-type p5.48xlarge \
  --instance-count 4 \
  --capacity-duration-hours 48 \
  --start-date-range 2026-10-20T00:00:00Z \
  --end-date-range 2026-11-10T00:00:00Z \
  --query 'CapacityBlockOfferings[].[CapacityBlockOfferingId,AvailabilityZone,StartDate,UpfrontFee,CurrencyCode]' \
  --output table

Only --capacity-duration-hours is required. Durations come in 1-day steps up to 14 days, then 7-day steps up to 182 days, and one block holds up to 64 instances, per the AWS CLI reference. Swap in your own Region, dates and counts.

Then put the price on a stepped scenario instead of a flat line:

# gpu_budget.py: reserved GPU spend if the rate steps at the next reprice
RATE_NOW = 5.970                       # P5, USD per accelerator-hour (AWS pricing page)
GPUS_PER_NODE = 8
NODE_HOURS_PER_QUARTER = 4 * 48 * 6    # 4 nodes, 48 h blocks, 6 blocks a quarter

scenarios = {'flat': 0.00, '+10% in Jan': 0.10, '+15% in Jan': 0.15, '+20% in Jan': 0.20}

for name, bump in scenarios.items():
    rate = RATE_NOW * (1 + bump)
    cost = rate * GPUS_PER_NODE * NODE_HOURS_PER_QUARTER
    print(f'{name:12s} ${rate:6.3f}/GPU-h  ${cost:,.0f} per quarter')

The bumps are assumptions, not predictions.

Compare reservations with the alternatives

Price the same workload three ways: Capacity Blocks, On-Demand or Savings Plans (which did not move, per the same coverage), and one or two other GPU providers. FastGPU's data shows the wider market drifting up too: of 176 on-demand GPU prices it compared in September, 15 rose and none fell.

Do not compare hourly numbers alone. A posted on-demand price carries no reservation, so weigh the start-time guarantee you need against the price.

What to do on Monday morning

1. Inventory planned blocks. Price each one with the command above. The rate locks at purchase, so a block with a firm date can be bought before the next update. Check cancellation terms first, because the fee is paid up front.
2. Give reserved GPU its own budget line with a stepped scenario like the script above, and re-plan it quarterly.
3. Cut idle hours. Tighter windows matter more when each reserved hour costs 15% more. Checkpoint long jobs so a shorter block is viable.
4. Get two outside quotes and ask your AWS account team what discount applies to Capacity Blocks in your agreement.
5. Add a repricing clause if you pass GPU costs to a customer. Flat quotes on variable capacity are how margins disappear.

For anyone deploying AI in production, reserved GPU capacity is a variable cost with a calendar, so plan for the next reprice instead of being surprised by it.