AWS Cuts Prices on NVIDIA GPU Instances

So I wrote my last article on GPUs and the Capacity Blocks AWS was requiring you to purchase in order to access precious Blackwell-based B6-B200 instances.   Capacity Blocks are no fun, requiring up front commitments of one day to 26 weeks with no ability to modify or cancel them upon ordering them.   But AWS clearly got a little nervous about all the neoclouds like Lambda, CoreWeave and co. offering Blackwells on demand and has decided to sell them now within Compute Savings Plans.  And if that’s not enough, they’ve cut prices on p4’s and 5’s - older gen Ampere and Hopper GPUs- which are now available for up to 45% less than they were yesterday, along with making them available on demand in more regions across the globe.  AWS does not cut prices 45% unless something is wrong, and something clearly is.

AWS generates $120 billion of annualized revenue and prices many of its AI-related services with things like Bedrock “Model Units” and “OpenSearch Compute Units” figuring you won’t ask what those units actually represent.  Moreover, it’s figured out that it doesn’t need to cut prices to stimulate demand like it used to…except with GPUs.

AWS has realized it's a little behind on packaging and pricing GPUs.   Moreover, generating all that revenue doesn’t do much to defeat competitors when anyone with access to GPUs can raise more than enough capital to rent them out to eager buyers.  As such, rigid Capacity Blocks are not one of the best ideas to come out of Seattle - asking you to reserve capacity with no ability to modify or cancel a reservation with an effective rate up to $65 per hour per instance, or roughly 100x the cost of a typical EC2 CPU-based server.

AWS accounts for approximately 60% of Amazon’s $18 billion quarterly operating profit per its last earnings announcement, which in turns supports a $2 trillion market cap.   Don’t expect this price cut to extend to more traditional services like S3, RDS, EBS, and especially not Data Transfer.  Nonetheless, this is good news for the GPU rental market and further innovations in agents, foundation models, and plenty of other advancements that will fill its customers' press releases in coming months.

Next
Next

AI Capex Grew over 80% Annually in Q1 to $90 billion