I've been tracking AWS for a long time, with a specific emphasis on pricing. "What happens if AWS hikes prices" has always been something of a boogeyman, trotted out as a hypothetical to urge folks to avoid taking dependencies on a given provider.
Over the weekend - on a Saturday, no less - that hypothetical became real.
AWS has quietly raised prices on its EC2 Capacity Blocks for ML by approximately 15 percent. The p5e.48xlarge instance – eight NVIDIA H200 accelerators in a trenchcoat – jumped from $34.61 to $39.80 per hour across most regions, while the p5en.48xlarge climbed from $36.18 to $41.61. Customers in US West (N. California) face steeper hikes, with p5e rates rising from $43.26 to $49.75. The change had been telegraphed: AWS's pricing page noted (and bizarrely, still does) that "current prices are scheduled to be updated in January, 2026," though the company neglected to mention which direction.
This comes about seven months after AWS trumpeted "up to 45% price reductions" for GPU instances - though that announcement covered On-Demand and Savings Plans rather than Capacity Blocks. Funny how that works.
For the uninitiated, Capacity Blocks are AWS's answer to "I need guaranteed GPU capacity for my ML training job next Tuesday." You reserve specific GPU instances for a defined time window – anywhere from a day to a few weeks out – and pay up front at a locked-in rate. It's popular with companies doing serious ML work who can't afford to have a training run interrupted because spot capacity evaporated. The pricing should make it abundantly clear that the people using this aren't hobbyists; these are teams with budgets measured in millions.
An Amazon spox told us via email, "EC2 Capacity Blocks for ML pricing vary based on supply and demand patterns, as described on the product detail page. This price adjustment reflects the supply/demand patterns we expect this quarter."
To be clear, AWS has raised prices before, but rarely as a straight increase to a line item. The company prefers to change pricing dimensions entirely, often spinning this as a price reduction for most customers – a claim I'd characterize as "creative." Historical straight-up price increases have been tied to regulatory actions: per-SMS charges in certain markets and the like. This is different.
The timing is curious for another reason: it hands Azure and GCP a talking point on a silver platter. Both have been aggressively courting ML workloads, and "AWS just raised GPU prices 15%" is exactly the kind of ammunition enterprise sales teams dream about. Whether the competitors can actually absorb the demand is another question – GPU constraints are hardly unique to AWS – but perception matters in enterprise deals.
For companies with Enterprise Discount Programs or other negotiated agreements, this raises uncomfortable questions. EDPs typically guarantee discounts off public pricing – so if public pricing goes up 15 percent, your "discounted" rate just got more expensive in absolute terms, even if the percentage held steady. I expect some pointed conversations between AWS account teams and their larger customers in the coming weeks.
Why are they doing this?
... continue reading