Amazon raises EC2 Capacity Block prices ~20% as GPU demand and memory costs soar

AWS announced that hourly rates for EC2 Capacity Blocks for ML — the service that lets customers reserve GPUs in advance to guarantee uninterrupted training and fine-tuning workloads — will jump roughly 20% starting in July, on top of an approximately 15% increase in January. 'Reservation prices are updated periodically based on supply and demand,' the company said, stressing the change applies to one purchasing option and that alternatives with fixed pricing remain.

The move is a clean data point for the week's compute-scarcity theme. Capacity Blocks are popular with serious AI developers running big-budget projects like training new models, so a 20% hike directly raises the cost of frontier-scale work and signals that GPU demand continues to outstrip supply. AWS explicitly tied it to high GPU demand, and the increase parallels memory-price pressure that pushed Apple to raise device prices the same week.

Competitively, the pricing power cuts both ways: it pads AWS margins but also strengthens rivals' pitches — notably Google's TPU cost advantage (Pichai's claimed 78% Gemini serving-cost reduction) and xAI renting idle Colossus GPUs into the shortage. Customers facing serial AWS hikes have fresh incentive to diversify across clouds or invest in custom silicon.

The broader signal, as Business Insider framed it, is an AI affordability squeeze: tech giants are passing memory and chip-cost pressure to customers across the stack. Watch whether sustained Capacity Block hikes accelerate enterprise interest in TPUs, Trainium, or alternative providers, and whether AWS holds the line on its 'fixed price' on-demand options as demand intensifies.