AWS Raises Nvidia GPU Reservation Prices 20%

In 2014, Amazon Web Services cut S3 prices by up to 22%. It has since raised advance-reserved Nvidia GPU prices in EC2 Capacity Blocks by 20%, while leaving Trainium prices unchanged.

Storage still rewards the price-down playbook

AWS also cut EBS prices by up to 50% in 2014. Nine years later, S3 Express One Zone offered writes ten times faster and request costs 50% lower than standard S3. The latter is a request-price comparison, not a claim that every storage bill fell. Both moves show how AWS could pair lower prices for a defined unit with a more specialized service.

An Nvidia reservation has a constraint that a storage request does not: the customer needs the right machines available together at the appointed time. AWS cannot make an unavailable block useful by lowering the price of some other instance.

A reservation insures the schedule, not the spot price

EC2 Capacity Blocks lets businesses rent AI compute in advance. The buyer commits to a block rather than assuming that equivalent capacity will be available later on demand. This is operational insurance against a delayed run, not a financial option that lets the buyer walk away if prices fall.

On-demand prices can move while a buyer holds an advance booking. In April, the Ornn Compute Price Index put a Blackwell GPU rental at $4.08 an hour, up 48% from $2.75 two months earlier. That index measures reported on-demand rentals, not AWS Capacity Blocks; it cannot establish AWS’s margin or the cause of its price change. It does show why a buyer with a fixed workload date might value certainty separately from an hourly discount.

OpenAI’s Guaranteed Capacity makes the same scheduling concern visible from another part of the stack. OpenAI offers discounted access to its compute in exchange for one- to three-year commitments tied to spending. A provider can charge for dependable access while rewarding a customer that makes demand easier to plan. CME Group and Silicon Data have also announced futures tied to on-demand GPU rental benchmarks. Those contracts may help hedge price exposure; a benchmark contract alone does not deliver the machines.

Trainium makes substitution an engineering decision

A customer can take advantage of AWS’s unchanged Trainium price only after establishing that its software, performance requirements and operating process work on the chip. The posted price alone cannot tell the buyer the cost of a completed run.

AWS has a large-scale example. It said Anthropic’s Claude Opus 4 launched on Trainium2 and that Project Rainier, built for Anthropic, uses more than 500,000 Trainium2 chips. AWS can manage more of that system’s design and supply than it can with Nvidia hardware. The example does not prove that every Nvidia workload can move, or that unchanged Trainium pricing means unchanged cost per completed task.

Multiyear deals put delivery risk on the balance sheet

Recursive Superintelligence signed a $410 million multiyear AWS compute deal after raising $650 million in May. The reported deal does not disclose a Capacity Blocks schedule, so it cannot be treated as another Nvidia reservation. It does show the scale at which a model builder is willing to commit to a compute provider rather than assemble every future run from short-interval rentals.

AWS is making longer-term supply arrangements too. In September, it signed a multigeneration customized-silicon and inference partnership with Qualcomm. Qualcomm issued Amazon warrants to acquire up to $4 billion worth of its stock as part of the arrangement.

For AWS, customer commitments help support procurement and infrastructure planning; for the customer, they concentrate dependence on the provider’s ability to deliver. Buyers of contracted AI capacity must weigh that delivery obligation alongside the cost of power and chips.

The premium has a test, not a permanent franchise

After Amazon set up Project Greenland to manage GPU allocation across its retail unit, it said it had “ample” GPU capacity. That internal allocation effort does not establish availability for every external customer or every accelerator, but it weakens any claim of an AWS-wide shortage.

Demand can also fall per task without a new data center opening. OpenAI engineers reportedly told colleagues they had found a way to more than halve inference costs. Cerebras, Groq and other chip suppliers are pursuing alternatives for inference. Neither a reported efficiency gain nor a rival chip makes a customer’s existing workload portable on its own; both put pressure on the amount a customer should pay to lock in one hardware type.

If customers can move workloads across accelerators and reliably obtain powered capacity at short notice, an Nvidia reservation becomes insurance for exceptional scheduling risk rather than the default way to secure a run. Until those conditions hold, the date belongs on the price tag.