AWS raises prices for AI GPU reservations
Amazon Web Services has posted a new pricing update for EC2 Capacity Blocks for ML, the reservation product companies use when they want guaranteed accelerator capacity for machine-learning work. The change takes effect on July 1, 2026. It is not a flashy model launch, but for many AI builders it is the kind of news that shows up directly in budgets.
On its pricing page, AWS says hourly rates per accelerator will move to $14.04 for P6-B300 capacity, $12.355 for P6-B200, $5.191 for P5 in U.S. regions, $4.72 for P5 outside the U.S., $5.97 for P5e, $6.865 for P5en in U.S. regions, $6.241 for P5en outside the U.S., and $2.214 for P4de. The company says other prices remain unchanged. Business Insider, which flagged the update on June 26, described it as another increase in the price of AI cloud capacity.
Capacity Blocks matter because they are designed for planned training, fine-tuning, experimentation and batch inference jobs where a team cannot simply hope that GPUs will be available when a run starts. Customers reserve capacity in advance and pay the prevailing rate at the time of purchase. That makes the product a useful signal for the broader AI market: when reserved accelerator capacity becomes more expensive, scarce chips, power, data-center space and cloud demand are still shaping the economics of AI.
For AI users, the immediate effect may be invisible. Chatbots and apps will still work. For makers and companies, the implications are more concrete. Training schedules, proof-of-concept budgets, pricing for AI features and commitments to customers may need another check before July. Teams that assumed compute would gradually get cheaper may have to be more careful about model size, run length and whether a workload really needs top-end Nvidia accelerators.
The bigger message is that the AI boom is not only a software story. Model quality, agents and applications get most of the attention, but cloud infrastructure prices still decide what can be built profitably. AWS is telling the market that premium AI capacity remains valuable enough to reprice.