The P6e pricing stays the identical at $761.904 for 72 B200 accelerators within the Dallas Native Zone for p4d.24xlarge.
“EC2 Capability Blocks for ML pricing are dynamic and fluctuate primarily based on provide and demand patterns, as described on the product element web page,” an AWS spokesperson mentioned. “This worth adjustment displays the provision/demand patterns we anticipate this quarter. AWS’s dedication to not increase pricing on fastened pricing fashions like On Demand and Financial savings Plans stays unchanged.”
“Essentially the most defensible clarification is solely market-based pricing tied to produce and demand,” mentioned Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting. “Because the demand for H100 and H200 GPUs outstrips provide, AWS is successfully making use of a shortage premium to assured stock. AWS is making an attempt to recuperate greater infrastructure and capital prices from pressing capability relatively than general capability.”
Assured GPU capability turns into the brand new battleground
Assured entry to GPU clusters permits enterprises to de-risk AI infrastructure planning and construct resilience towards future provide volatility. Acknowledging the steep demand for high-end GPUs leading to a scarcity of Nvidia H100 and H200s, large clouds are more and more providing assured capability to clients.
Apart from AWS, Google and Microsoft even have related choices, however introduced in additional conventional reservation fashions and scheduling frameworks.
As an illustration, Google Cloud has launched a calendar-based scheduling instrument that lets clients reserve GPU capability in fastened blocks forward of time. “On paper, that appears loads like what AWS is doing with Capability Blocks. However the framing is completely different. Google is treating it as a part of its broader useful resource scheduler, not a premium SKU. The assure continues to be there, however the pricing doesn’t really feel as segmented or dynamic. It’s nearly as in the event that they’re utilizing scheduling to compete, not worth. And since they will additionally steer some workloads onto TPUs as a substitute of GPUs, they’ve received slightly extra flexibility constructed into the system,” mentioned Sanchit Vir Gogia, CEO and chief analyst at Greyhound Analysis.
