Cloud customers who locked in graphics-processing capacity with Amazon Web Services are about to pay more to keep it. The company is raising prices on capacity reservations for most of its GPU instance types by 15% or more, a change that takes effect October 7, Yahoo Finance reported October 2.
The increase applies to the reservations that let customers secure access to scarce GPU capacity ahead of time, a service that has become essential for AI companies that need guaranteed compute for training and inference. Under the new pricing, the cost of holding that reserved capacity will rise across most GPU instance types, reflecting what AWS says is the higher cost of the memory and processors behind it.
This is the second time in six months that Amazon Web Services has adjusted prices on these capacity reservations, the report said. The repeat move shows how persistently the economics of AI infrastructure have shifted against the cloud providers themselves, which buy chips and memory at the front of the supply chain and then pass the cost through to the customers renting them.
The timing matters. Capacity reservations have become the mechanism by which AI labs and enterprises guarantee themselves a place in line for the most powerful chips, which have been in short supply since the generative-AI boom began. When AWS raises the price of those reservations, it is effectively raising the price of certainty: a customer who wants to know its GPUs will be there next quarter now has to pay more for that assurance.
Amazon’s GPU lineup spans the P4 and P5 families built on Nvidia chips as well as its own Trainium and Inferentia processors, and demand for the Nvidia-based instances has outstripped supply almost continuously since large language models went mainstream. Capacity reservations were originally a way for enterprises to lock in predictable capacity for steady workloads; they have since become a rationing tool, with the price signaling how tight the queue is at any given moment.
The underlying pressure comes from two directions. Memory prices, particularly for the high-bandwidth memory that sits next to AI accelerators, have climbed as data-center demand has outstripped supply. GPU procurement costs have moved the same way. AWS, like its rivals, has responded by building more capacity, but the chips it buys to fill those data centers have grown more expensive, and the company has shown it is willing to pass those increases along.
For customers, the choice is rarely simple. Many have signed multiyear commitments tied to specific instance families, and the cost of moving a trained model or a production workload to a rival cloud is not trivial. AWS knows this, and the price increase appears calibrated accordingly: enough to recover costs, not enough to push a large customer into the arms of Microsoft Azure or Google Cloud in a single quarter.
The distinction between reservation prices and instance prices matters for understanding what is changing. A capacity reservation is a commitment to hold specific compute for a customer, separate from the per-hour rate that customer pays when it runs workloads on that compute. Raising the reservation price does not change the hourly rate, but it raises the cost of simply guaranteeing access, which is the thing most in demand when chips are scarce.
Enterprise customers, rather than the largest AI labs, are the ones most exposed to the change. The biggest model developers negotiate bespoke deals and multiyear contracts that can absorb or offset reservation fees. Smaller companies that buy capacity on standard terms have less room to bargain, and analysts said they are the most likely to feel the October increase in their monthly infrastructure bills.
The broader industry is watching. The major cloud providers have all raised prices or introduced surcharges on AI-related services as the cost of running them has climbed. AWS is the largest of the three, and its pricing decisions tend to set the tone for the rest of the market. A second capacity-reservation increase inside six months signals that the company does not expect the cost of AI infrastructure to fall soon, and that it plans to keep customers carrying a share of the burden.
The dynamic is unlikely to reverse quickly. Building new data centers takes years, and the components that go into them are ordered far in advance at prices set by the current shortage. Until that new supply arrives, the scarcest resource in cloud computing will remain the GPU, and the price of locking one down will keep reflecting that scarcity. For the AI companies racing to ship products, the October 7 increase is a small line item with a large meaning: the compute they depend on is getting more expensive, and the provider selling it to them is no longer absorbing the cost.
