Desperately seeking GPUs? AWS will let you reserve instances in advance – no refunds

Better make sure you really need it on the day. Paid upfront and no modifications to orders allowed

AWS has come up with a new money-making scheme – letting customers desperate for GPU resources pay to reserve them for scheduled dates and times, paid upfront, and with no order modification allowed.

The new consumption model is known as Amazon Elastic Compute Cloud (EC2) Capacity Blocks for ML. It lets customers access highly sought-after GPU compute in order to run short machine learning workloads.

The growth in demand for GPU capacity to train and inference machine learning models has outpaced industry-wide supply, making GPUs a scarce resource.

What AWS doesn't admit, of course, is that this is because hyperscaler are at the front of the queue for GPUs, hoovering up supply, including AWS itself. Nvidia reportedly said 22 percent of its calendar Q2 revenue was driven by a single cloud service provider.

EC2 Capacity Blocks are initially available for Amazon EC2 P5 virtual machine instances, of which there is currently only one type, the p5.48xlarge. This features 192 vCPUs, 2 TB of system memory, and eight of Nvidia's H100 GPUs, making it a pretty hefty instance.

According to AWS, EC2 Capacity Block reservations work like hotel room reservations. You specify the date and duration and size of your room.. In the same way, AWS lets you select the date and duration you will need GPU instances and the number of instances required.

Customers can reserve an EC2 Capacity Block comprising just a single instance, or up to 64, and these can be reserved for one to 14 days as required. AWS will allow these to be reserved up to eight weeks in advance, claiming that customers can then plan for their machine learning deployments with certainty, knowing they will have the GPU capacity when they need it.

AWS says EC2 Capacity Block prices are dynamic and depend on the total available levels of supply and demand at the time the customer buys. AWS claims it will show the lowest-priced offering available that meets the range the customer has specified. A screenshot on the company's blog shows an example of a single instance for a single day costing $2,344.

But here's the kicker: the total cost of an EC2 Capacity Block is charged up front, billed to your account within 12 hours, and AWS does not allow them to be modified or cancelled after purchase. So you'd better be sure you really need that GPU capacity on those dates.

And once your time is up, your workload will be unceremoniously halted. EC2 will emit an event through Amazon EventBridge to alert that the reservation is ending so the user can checkpoint the workload. Running instances will go into a shutting-down state 30 minutes before the reservation ends, but once the time expires, any instances still running will be terminated.

EC2 Capacity Blocks are available now, but initially only in the AWS US East (Ohio) Region. Availability is planned for additional Regions and Local Zones in future, AWS said.

The cloud biz reported revenue of $23 billion for Q3 of this year, up from $20.5 billion a year earlier, reflecting the continued growth of the cloud market but at a slower pace as customers seek ways to rein in spending. Like the other big clouds, AWS sees demand for AI services as a way to combat that trend. ®

More about


Send us news

Other stories you might like