The third hike of the year
AWS is reportedly preparing to raise prices on EC2 Capacity Blocks for machine learning, the reserved-capacity product that lets customers book guaranteed access to Nvidia GPU instances for a defined window, by 15 percent effective October 7. The move follows a 15 percent increase in January and a 20 percent increase in July, which means this would be the third upward adjustment to Capacity Block pricing inside a single calendar year for the same underlying hardware generation.
The specific instance most affected is the p5e.48xlarge, built around Nvidia's H200 GPUs, which remain one of the most sought-after accelerators for large-scale training and inference workloads even as newer chips come to market. AWS has not publicly confirmed the exact October figure at the time of reporting, which is itself a pattern: these capacity pricing adjustments have tended to surface through customer reports and secondary coverage before AWS issues an official statement, leaving procurement teams to react on shorter notice than a typical enterprise pricing change would allow.
Why compounding matters more than any single number
A 15 percent increase sounds manageable in isolation, and considered on its own it often is. The problem is that this is the third increase in ten months, and each one lands on top of the prior increase rather than against some fixed original baseline. The p5e.48xlarge rate reportedly moved from approximately 34.61 dollars an hour to around 39 dollars after the January increase, then continued climbing through the July adjustment to as high as 49.74 dollars an hour in the US West region, depending on exact configuration and region.
That trajectory represents roughly a 44 percent increase from January's starting point before this reported October adjustment is even applied, which is a materially different number than three isolated headline percentages would suggest to a finance team skimming quarterly cost reports. Anyone modeling 2027 AI infrastructure spend off a snapshot of current Capacity Block pricing is almost certainly underestimating the number, based on how this year has actually played out.
The demand story behind the price
AWS has characterized these increases as a response to supply and demand, which lines up with the company's own stated plan, announced in March, to deploy more than a million additional Nvidia GPUs within a year. Capacity that scarce commands a premium, and AWS is not alone in raising prices on exactly this kind of reserved GPU capacity. Reports over the same period describe competing neocloud providers raising their own Nvidia GPU prices by as much as 21 percent, suggesting this is an industry-wide pricing dynamic rather than an AWS-specific decision driven by its own margin targets alone.
That broader pattern matters for buyers evaluating alternatives. Shopping a workload to a different cloud or neocloud provider in search of cheaper GPU capacity is a reasonable instinct, but it is not guaranteed to produce meaningfully better pricing if every major supplier is adjusting prices upward in roughly the same direction and on a similar timeline, driven by the same underlying chip scarcity rather than by any one vendor's pricing strategy.
What this does to AI project economics
Training and inference budgets built on an assumption of flat or slowly rising GPU costs are the ones most exposed here. A project greenlit in January using that month's Capacity Block pricing as its cost basis would already be running against a materially higher number by October, before accounting for however much compute the workload actually ends up consuming as scope grows, which it very often does once a pilot proves out.
That gap between planning assumptions and actual spend is exactly the kind of variance that turns a well-reasoned AI pilot into a budget overrun story at the next board meeting, not because the project failed technically but because the underlying infrastructure cost moved substantially during the build. Finance leaders sponsoring AI infrastructure investment should build an explicit price-escalation assumption into any multi-quarter GPU capacity commitment, rather than anchoring a forecast to whatever the published rate happens to be on the day a project gets approved.
The practical response for infrastructure buyers
Locking in longer-term commitments, where a vendor offers them, trades flexibility for price certainty, and that trade looks considerably more attractive in an environment of repeated, compounding mid-year increases than it did when GPU pricing was comparatively stable. Enterprises with predictable, sustained training or inference workloads should revisit whether a reserved, multi-quarter commitment now beats the cumulative cost of continuing to pay on-demand or short-term Capacity Block rates through another one or two rounds of increases.
For workloads that genuinely are variable and hard to commit to in advance, the more durable response is architectural: invest in making training and inference jobs portable across GPU generations and cloud providers where feasible, so that a 2027 price shock on one specific instance family and vendor does not become a forced, costly migration decided under pressure rather than a deliberate, planned one.
