An AWS p5e.48xlarge Capacity Block reservation cost $34.61 an hour in December 2025. By July 1, 2026, the same reservation costs roughly $47.76 an hour. That's a 38 percent increase in six months, delivered in two separate hikes, on a product AWS customers reserve specifically to guarantee GPU availability.
An AWS p5e.48xlarge Capacity Block reservation, eight NVIDIA H200 GPUs bundled into a single instance, cost $34.61 an hour in December 2025. AWS raised that rate to $39.80 on January 4, 2026, a quiet, weekend announcement with no advance customer notice. Then, on July 1, 2026, AWS raised it again, this time to roughly $47.76 an hour. That's a 38 percent cumulative increase in six months, on a product enterprises specifically pay a premium for because it guarantees GPU capacity will be there when a training run needs to start.
Two price increases inside six months is not a one-time correction. It's a pattern, and it changes how any organization budgeting AWS GPU reservations needs to plan for the rest of 2026 and beyond.
| Metric | January 4, 2026 | July 1, 2026 |
|---|---|---|
| Increase | ~15% | ~20% |
| Families affected | P5e, P5en (H200-based) | P6-B300, P6-B200, P5, P5e, P5en, P4de |
| Example: p5e.48xlarge | $34.61 to $39.80/hr | $39.80 to ~$47.76/hr |
| Advance notice | None, posted over a weekend | Published on AWS documentation page |
The January hike affected P5e and P5en instances specifically, the H200-based families. The July hike was broader, covering P6-B300 and P6-B200 (AWS's Blackwell-generation instances), plus P5, P5e, P5en, and P4de, essentially AWS's entire GPU reservation lineup. Older instance families saw smaller dollar increases but similar percentage jumps, meaning this wasn't a single product line getting repriced, it was the whole Capacity Blocks catalog.
The new July 1 rate card, per accelerator-hour in US regions:
| Instance Family | New Rate (per accelerator-hour) |
|---|---|
| P6-B300 (Blackwell) | $14.04 |
| P6-B200 (Blackwell) | $12.355 |
| P5en (US) | $6.865 |
| P5e (US) | $5.97 |
| P5 (US) | $5.191 |
| P4de (US) | $2.214 |
It's worth being precise about what actually got more expensive here, because AWS's GPU pricing tells two very different stories depending on which product you're looking at.
Capacity Blocks for ML are AWS's reserve-ahead product: you commit to a fixed block of GPU accelerators for a defined future time window and pay upfront for the guarantee that hardware will be available when your window starts. This is what got the two hikes covered here.
Standard on-demand and Savings Plan pricing actually moved in the opposite direction over the past year. In June 2025, AWS cut on-demand pricing for P5 instances by up to 45 percent, P5en by up to 26 percent, and P4d and P4de by up to 33 percent. That cut is still in effect for those product types.
Put those two facts together and the picture sharpens: AWS didn't raise GPU prices across the board, it raised the price of the guarantee specifically. Reserving capacity ahead of time, the exact thing enterprises do when a training run absolutely cannot slip because hardware wasn't available, now costs meaningfully more than it did six months ago, even as flexible, non-reserved pricing has gotten cheaper over the same broader period.
AWS's public explanation has been identical both times: Capacity Blocks for ML pricing varies based on supply and demand patterns, and each adjustment reflects the patterns AWS expects for the quarter ahead. That explanation is not implausible on its face. Amazon has committed roughly $200 billion in capital expenditure to AI infrastructure in 2026, and Reuters reported in March 2026 that Amazon is set to receive 1 million NVIDIA GPU chips by the end of 2027 under a cloud supply agreement, a deal that itself signals just how constrained high-end GPU supply remains industry-wide.
Not everyone reads it that way. Cloud economist Corey Quinn offered a pointed counter-argument after the January hike, characterizing it as AWS updating its published base rates rather than responding to a specific supply shock, a policy decision dressed up as a market response rather than a market response itself. Both readings can be partly true at once: genuine supply constraints are real and well documented, and a company facing genuine constraints still has discretion over how much of that constraint it passes through as margin versus absorbs. Nothing about the two hikes settles which explanation is doing more of the work, and reasonable people who track AWS pricing closely land on different sides of it.
Whatever the underlying cause, the practical planning implication is the same either way.
If your organization runs multi-quarter AI training programs on AWS Capacity Blocks, build a 15 to 20 percent price increase assumption into any budget spanning more than two quarters. Treating the current rate as fixed is no longer a safe assumption based on the last six months of AWS's own pricing behavior.
Capacity Blocks exist to solve a specific problem: guaranteeing GPU availability for a training run that cannot slip. If a workload can tolerate some scheduling flexibility, on-demand or Savings Plan pricing, both cheaper than a year ago, may now be the more economical path, especially since reservation pricing is the piece that just got more expensive twice.
With reserved AWS GPU capacity now costing meaningfully more than it did in late 2025, this is a reasonable moment to get current quotes from neoclouds and specialized GPU providers rather than assuming AWS remains the default by inertia. The comparison only makes sense if you also vet ownership and reliability, not price alone, but price alone has moved enough to justify running the comparison again even if you ran it six months ago.
If either competitor holds its equivalent reserved-capacity pricing flat while AWS continues raising Capacity Blocks rates, that's a meaningful signal AWS's moves reflect company-specific margin decisions rather than an industry-wide supply reality, and a strong argument for diversifying reserved capacity across providers rather than concentrating it with AWS.
One price increase is an adjustment. Two in six months is a policy. Budget accordingly.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
