By On Saturday, January 4, 2026, Amazon Web Services raised EC2 Capacity Block GPU prices by approximately 15% without a separate customer announcement. The p5e.48xlarge instance featuring eight NVIDIA H200 accelerators jumped from $34.61 to $39.80 per hour—a $5.19/hour increase that adds $3,737/month for teams running continuous GPU workloads. AWS implemented the change over the weekend and buried it in a pricing page update. The Register captured the sentiment perfectly: “AWS raises GPU prices 15% on a Saturday, hopes you weren’t paying attention.” The timing stands in stark contrast to AWS’s June 2025 announcement of “up to 45% GPU price reductions,” which the company publicized through blog posts and press releases—those cuts, however, applied to On-Demand and Savings Plans instances, not Capacity Blocks.
The Weekend Stealth Move
AWS chose Saturday for implementation with zero customer communication beyond a generic pricing page note that “prices are scheduled to be updated in January, 2026.” No direction specified. No email to customers. No blog post. Tech journalists discovered the increase—AWS didn’t announce it.
This matters because AWS shouts about price cuts but whispers about increases. When AWS reduced On-Demand GPU prices by up to 45% in June 2025, they published a blog post, issued press releases, and generated media coverage. When they raised Capacity Block prices 15% seven months later, they updated a pricing page on a Saturday. Selective transparency undermines trust in cloud pricing commitments.
Why Capacity Block Lock-In Makes This Worse
EC2 Capacity Blocks are reserved GPU access: you pay upfront for guaranteed GPU availability at a scheduled future date, typically for ML model training or generative AI experiments. You can reserve 1 to 64 instances (up to 512 GPUs) for up to six months, and once reserved, you’re locked in.
Customers who reserved blocks before January 4 are safe; they already paid the old price. New reservations starting January 4, 2026 cost 15% more. You can’t easily switch to On-Demand instances because GPUs are often sold out during peak demand. You can’t switch cloud providers mid-project without expensive data migration. AWS forces a choice: pay 15% more or risk not getting GPUs when you need them.
For an enterprise running 10 instances 24/7, that’s $37,370/month in additional costs.
GPU Shortage Gives AWS Pricing Power
AWS raised prices because they can. NVIDIA received orders for 2 million H200 chips for 2026, but current inventory stands at just 700,000 units. TSMC, the sole manufacturer, prioritizes newer Blackwell and Rubin architectures over H200 production. Samsung and SK hynix raised HBM3E memory prices by 20% for 2026, plus China demand from ByteDance and Alibaba further strains supply.
Developers building large language models or running deep learning experiments need H200/H100 GPUs. Alternative cloud providers (Azure, GCP) face the same supply constraints. AWS knows developers are stuck, which is why they felt comfortable implementing a weekend price hike with no customer announcement.
The June 2025 Deception
AWS’s June 2025 “up to 45% price reduction” announcement covered P4 (A100) and P5 (H100/H200) instance types—the exact same instance families they just increased Capacity Block pricing for. The June cuts applied to On-Demand and Savings Plans. The January increases hit Capacity Blocks.
AWS gets credit for lowering prices (June 2025 headlines: “AWS cuts GPU costs by up to 45%!”) while quietly raising them (January 2026: weekend pricing page update, no headlines). Selective transparency creates a misleading impression that AWS is making GPUs more affordable when, for Capacity Block customers, they’re doing the opposite. This isn’t illegal. It’s just dishonest.
Alternatives Exist, But Come With Trade-Offs
Azure and Google Cloud offer H200 instances at roughly $10.60/hour per GPU, comparable to AWS’s new pricing. Specialized GPU cloud providers like GMI Cloud offer H200 at $2.50/hour—50% cheaper than hyperscalers. On-premise GPUs become cheaper after 12-18 months for sustained workloads, though they require $10K-$30K upfront per GPU plus data center costs.
Most developers stick with AWS despite higher prices due to ecosystem lock-in. If your data lives in S3, your workflows integrate with Lambda and SageMaker, and your enterprise has existing AWS commitments, switching providers means expensive re-architecture and data egress fees.
What Developers Should Do
Check your AWS bills closely—the next pricing review is April 2026, and nothing stops AWS from raising prices again with the same weekend stealth approach. Evaluate alternatives if you’re not deeply locked into the AWS ecosystem. Specialized providers offer meaningful savings for teams willing to trade polish for cost.
Demand pricing transparency from cloud vendors. AWS’s weekend announcement without customer notification is unacceptable for infrastructure teams managing budgets and planning roadmaps. When cloud providers bury price increases but publicize cuts, they’re not acting in good faith. Vote with your wallet. Long-term, adopt a multi-cloud strategy to reduce dependency on any single provider. The GPU shortage will continue through 2026, and cloud providers will keep testing how much they can raise prices before customers leave. Make it easy to leave.
