AWS to Raise GPU Computing Rental Prices Across Multiple Regions Starting July 1
0xBroomberg
Amazon Web Services will raise reserved GPU pricing across its full Nvidia lineup — from H100 to B300 — starting July 1, marking the second hike in six months and signaling that enterprise AI training costs are structurally climbing.
What exactly is getting more expensive?
AWS is raising prices on EC2 Capacity Blocks — a service that lets customers reserve GPU compute in advance, essentially "booking" GPU time slots before they need them.
Per-accelerator hourly rates now range from $2.214 for the older P4de to $14.04 for the flagship P6-B300. The top tier costs over 6× more than the bottom.
This means → whether you're on last-gen or cutting-edge Nvidia GPUs, the bill is going up. This is a blanket increase, not a new-model surcharge.
Why can AWS raise prices twice in six months?
AWS stated explicitly that reserved pricing will be adjusted periodically based on supply and demand — a clear signal that further hikes are possible if demand holds.
This is the second increase since January 2025, just six months apart.
This reflects a supply-demand imbalance that still favors the seller: enterprise appetite for AI training and inference compute is growing faster than AWS can expand capacity. In plain terms = more buyers than sellers, so the seller sets the price.
How much does region choice matter?
The same instance type carries notably different prices by region. P6-B200: $10.296 per accelerator in the US, $12.3552 in Asia-Pacific Mumbai, $12.870 in GovCloud — Mumbai runs roughly 20% above US pricing.
P5 instances (H100 GPUs) show a smaller gap: $4.326 in the US vs. $4.72 outside the US.
This means → for large-scale training jobs, the region you deploy in now makes a material difference to the total bill — enough to reshape infrastructure decisions.
What does the most expensive tier look like?
The top-end P6e UltraServer — a single machine packing 36 or 72 accelerators — runs $380.952–$761.904 per hour, or $10.582 per accelerator.
In plain terms = renting a 72-accelerator UltraServer for one day costs over $18,000 in compute alone.
This reflects AI training's shift from "rent a few GPUs to experiment" to "book entire cluster-machines by the time block" — an industrial-scale pricing model to match an industrial-scale workload.
What does this mean for enterprises and the market?
The direct consequence of back-to-back hikes: hard compute costs are rising structurally for enterprise AI projects, especially mid-to-large training runs that rely on reserved capacity.
This means → for large enterprises still debating whether to build their own compute, the steady climb in cloud rental costs will force a fresh build-vs.-rent reassessment.
The next validation point: whether this round of increases passes through to end customers smoothly, or pushes enterprises toward more efficient inference architectures or alternative cloud providers.
Content is for reference only, not financial advice.