AWS reportedly lifts GPU capacity prices again, with a 15% hike flagged for October 7

Amazon Web Services may be raising prices on its reserved GPU compute yet again. A post on X claims AWS is lifting reserved compute prices for a fourth quarter, with a 15% increase set to begin October 7.

What AWS has actually raised in 2026

The increases center on EC2 Capacity Blocks for ML. That product lets customers lock in GPU capacity for machine learning workloads ahead of time.

The first confirmed move came in January 2026. AWS raised Capacity Block prices by approximately 15%, with the hike landing on instances that run NVIDIA H200 GPUs.

A second, larger bump followed on July 1, 2026. That one added roughly 20% to the same GPU capacity service.

The effect shows up clearly on a single machine. After the July change, the H200-based p5e.48xlarge instance went from about $34.61 per hour to about $39 per hour in most regions. In the US West region, the post-July price for that instance reached $49.74 per hour.

The October 7 increase is a different story. The 15% figure comes from the X post, and research findings indicate that number has not been publicly confirmed by AWS.

Why GPU capacity keeps getting pricier

AWS has reportedly framed these changes as a response to supply and demand.

The scope of the increases is narrow on purpose. They hit reserved capacity carved out for GPU workloads, not standard on-demand instances or general Reserved Instances.

Capacity Blocks let customers book GPU capacity up to eight weeks ahead of time. The reservations can also be shared across accounts, which helps larger organizations spread capacity between teams.

AWS itself said in March 2026 that it plans to deploy over a million NVIDIA GPUs in the next year.

What this means for AI builders and enterprises

The most direct impact falls on companies that rely on reserved GPU blocks for AI training and inference. If the reported October 7 increase does take effect, it would stack on top of the January and July hikes. Each new percentage lands on an already higher base, so the cumulative squeeze grows faster than the individual numbers suggest.

Regional gaps add another wrinkle. The spread between roughly $39 per hour in most regions and $49.74 per hour in US West is wide enough that workload placement becomes a real cost decision, not just a latency one.

Until then, the things to watch are straightforward: whether AWS publicly confirms the October 7 change, whether the increase reaches instance types beyond the H200 family, and whether reserved GPU pricing starts to look less like a fixed rate and more like a moving target.