AWS posts new Capacity Block GPU rates effective October 7.

Amazon Web Services has posted new reservation rates for EC2 Capacity Blocks for ML, effective October 7, 2026. The prices are hourly, per accelerator: $16.146 for P6-B300 ($16.819 in AWS GovCloud), $14.208 for P6-B200 ($14.801 in GovCloud), $7.895 for P5en, $6.866 for P5e, $5.970 for P5, $2.546 for P4de, and $1.696 for P4d. AWS did not state a percentage.

The company says Capacity Block prices move with supply and demand and are updated from time to time. On its pricing page it says the new rates apply across available regions for those families. It also says every other EC2 Capacity Block price stays the same, and that prices on fixed models such as On-Demand and Savings Plans do not change. UltraServer and Trainium Capacity Block rates are on the current tables and are not in the October 7 list.

A Capacity Block is charged at the rate in effect when the customer buys it, AWS says, even if the block is scheduled to start after a later update. The fee is paid up front.

News outlets, citing Wells Fargo rather than quoting a public research note, have described a reserved-GPU increase of about 15 percent on October 7 and called it a fourth straight quarterly hike. Cointelegraph wrote that AWS is raising reserved GPU prices by 15 percent on Oct. 7, “marking its fourth straight quarterly hike, per Wells Fargo.” Roic AI, attributing the figure to Wells Fargo, said reserved GPU compute prices would rise 15 percent on October 7 after a 20 percent increase in July, and described the step as the fourth consecutive quarterly increase. No Wells Fargo note was available to quote directly, and none of those percentages appears on the AWS page. AWS’s own line is the dollar rates above, not “15 percent.”

The analyst shorthand is also wider than the product AWS named. The public write-ups say “reserved GPU” prices. AWS limits the posted change to the Capacity Block families listed, and it says other Capacity Block prices, On-Demand, and Savings Plans are unchanged. The current tables still show more than one regional rate for some families, including P5 and P5en, while the October 7 notice states a single rate for each of those names. AWS does not say every current regional row rises by the same share.

This rate card is not Amazon’s reported talks to move about $8 billion of Nvidia Grace Blackwell chips into an investor vehicle, and it is not Amazon’s pledge to spend more than $1 billion in data-center communities. Those are separate stories.

Stakes: GPU reservations are one of the few AI inputs with a public price tag, and AWS just printed a higher one for its top Capacity Block accelerators. How large the step is, in percent, is an analyst claim. The number on Amazon’s own page is the new hourly rate.