AWS GPU price increase 2026: EC2 Capacity Blocks jump 20% as memory crunch tightens

AWS is raising GPU reservation prices again. Starting July 1, 2026, EC2 Capacity Blocks for machine learning will cost about 20% more, marking the second AWS GPU price increase 2026 in six months. For AI builders who rely on reserved Nvidia GPU capacity to train or fine-tune large models, the cost of locking in compute is climbing sharply.

What happened

AWS increased prices for EC2 Capacity Blocks for ML by approximately 20%, effective July 1, 2026. This follows a roughly 15% increase in January of the same year. The combined effect means the cost of reserving GPU capacity has jumped significantly since the start of the year. AWS said prices change periodically based on supply and demand. Affected instance families include P6-B300, P6-B200, P5, P5e, P5en, and P4de.

The increase is specific to the reserved capacity block purchasing option. Other pricing models for EC2 remain fixed, according to AWS. Amazon's in-house Trainium chip was spared in this hike.

Why AI builders should care

This price increase is not just a vendor adjustment. It reflects a physical constraint: high-bandwidth memory shortage. The chips stacked beside AI processors are in tight supply, which limits how many GPUs can be built and how many data centers can be deployed. As BCA Research chief economist Peter Berezin noted, less memory means fewer GPUs, and fewer GPUs means fewer data centers. That scarcity gives cloud providers pricing power. AWS, Microsoft, Google, and Oracle can pass higher costs through because demand exceeds supply.

For teams that depend on EC2 Capacity Blocks to guarantee long training runs without interruption, this directly affects AI training costs. The reserve button now costs more to press, and the trend line is not in the builder's favor.

Practical implications

If you use reserved GPU blocks for training or fine-tuning, your budget needs a buffer. Consider mixing reservation types: use reserved blocks only for critical, long-running workloads and shift interruptible or less urgent jobs to on-demand or spot instances. Evaluate Trainium-based instances if your framework supports it. And monitor whether other cloud providers follow with similar increases.

Caveats

The price change applies to select GPU instance families and only to the EC2 Capacity Block reservation model. Prices may vary by region. AWS stated that other pricing options remain unchanged, but no guarantee was given for how long that holds. The reported $14.04 per hour figure for P6-B300 instances is based on one source and may not cover all regions or discount programs. The high-bandwidth memory shortage is widely cited but AWS itself described the pricing change as based on supply and demand, without explicitly naming memory as the sole cause.

FAQs

Why is AWS raising GPU prices for EC2 Capacity Blocks?

AWS stated that prices change periodically based on supply and demand. The broader context is a high-bandwidth memory shortage that limits GPU production and data center buildout. With demand outstripping supply, cloud providers have pricing power over reserved capacity.

When do the new GPU reservation prices take effect?

The new pricing takes effect July 1, 2026 for affected EC2 Capacity Block for ML reservations.

Which AWS GPU instance families are affected by the price increase?

The increase applies to select GPU instance families used in EC2 Capacity Blocks for ML, including P6-B300, P6-B200, P5, P5e, P5en, and P4de. Other EC2 pricing options remain unchanged.

Is Trainium affected by the price increase?

No. AWS's in-house Trainium chip was spared in this price increase, according to reporting by The Information. This may make Trainium instances more attractive for builders who can adapt their workflows.