AWS to Hike GPU Reservation Prices in July, Signaling AI Compute Crunch Has No End in Sight

Amazon (AMZN.US) is flexing its pricing power in the cloud computing market, announcing a sweeping increase to its GPU capacity reservation service that underscores the relentless demand for artificial intelligence infrastructure. Amazon Web Services (AWS) will raise prices for its EC2 Capacity Blocks for machine learning workloads effective July 1, 2026, marking the second such adjustment this year as enterprises scramble to lock down scarce Nvidia (NVDA.US) chips.

The new pricing structure will push hourly rates per GPU accelerator higher across multiple instance families powered by Nvidia's latest silicon, including the Blackwell B300 and B200, as well as the older but still heavily utilized Hopper-based H100 and H200. AWS confirmed the changes in its official documentation, stating that "Amazon EC2 Capacity Blocks for ML reservation prices are updated periodically based on supply and demand."

According to the updated price list, the P6-B300 instance, which runs on Nvidia's cutting-edge Blackwell B300 GPUs, will now cost $14.04 per GPU hour. The P6-B200 instance follows at $12.355 per GPU hour. For the widely adopted P5 series built on H100 architecture, the rate in US regions climbs to $5.191 per GPU hour, while non-US regions will see a rate of $4.72. The P5e and P5en instances, which utilize H200 GPUs, are priced at $5.97 and $6.865 per GPU hour in US regions respectively. Even the older P4de instances equipped with A100 GPUs will see rates rise to $2.214 per GPU hour.

The scale of enterprise spending on these reservations becomes clear when looking at full node configurations. A P6e UltraServer composed of 72 Blackwell GB200 GPUs will command an hourly rate of $761.90 in the Dallas local zone, translating to roughly $10.58 per GPU. A more common eight-GPU P5 instance running H100 chips will cost about $34.61 per hour in major US regions, while an eight-GPU P5en node with H200 chips runs approximately $45.77 per hour.

This pricing action arrives at a moment of extraordinary financial momentum for AWS. The cloud division posted revenue of $37.6 billion in the first quarter of 2026, a 28% year-over-year surge that represents its fastest growth rate in more than three years. Amazon has committed roughly $200 billion in capital expenditure this year alone to AI infrastructure, and Reuters reported in March 2026 that the company is set to receive 1 million Nvidia GPU chips by the end of 2027 under a cloud supply agreement.

Instance Type	GPU Architecture	New US Price (per GPU hour)
P6-B300	Blackwell B300	$14.04
P6-B200	Blackwell B200	$12.355
P5en	H200	$6.865
P5e	H200	$5.97
P5	H100	$5.191
P4de	A100	$2.214

Note: Non-US pricing differs for select instance types. P5 non-US is $4.72 per GPU hour; P5en non-US is $6.241 per GPU hour.

Beyond Nvidia-powered instances, AWS also disclosed pricing for its proprietary Trainium AI chips. The Trn1 instance, which uses first-generation Trainium accelerators, costs approximately $0.596 per accelerator per hour, while the newer Trn2 instance commands about $2.235 per accelerator per hour. The wide gap between Trainium and Nvidia GPU pricing illustrates both the premium that Nvidia silicon commands and the potential cost savings available to customers willing to adopt AWS's custom silicon.

EC2 Capacity Blocks function as a reserved-capacity product, allowing enterprises to secure GPU instances for future dates to handle time-bound, large-scale model training workloads. Because availability guarantees are baked into the reservation model, customers have historically been willing to pay a premium over on-demand rates. The July 1 increase represents a significant step-up in that premium, and it raises questions about how enterprise buyers will respond.

The tight supply environment shows no signs of easing. Nvidia's H100, H200, and Blackwell series GPUs remain in a state of persistent shortage, driven by the rapid proliferation of generative AI, large language models, and autonomous AI agents. This supply-demand imbalance has given cloud providers considerable leverage to adjust pricing upward without immediate risk of customer defection, particularly for workloads already locked into specific GPU architectures.

For Nvidia, the AWS price hike serves as a dual-edged signal. On one hand, it confirms robust end-market demand for its silicon. On the other, rising reservation costs could accelerate customer interest in alternatives. Google Cloud has been actively marketing its Tensor Processing Units as a cost-competitive option, and AWS itself continues to push Trainium as a lower-cost alternative for certain AI workloads. Whether customers begin shifting workloads away from Nvidia instances in meaningful numbers remains an open question.

The competitive dynamics among hyperscale cloud providers will be closely watched in the coming weeks. Microsoft Azure remains AWS's nearest rival in enterprise cloud infrastructure, and a unilateral price increase by AWS could either trigger matching hikes across the industry or create an opening for Azure and Google Cloud to attract cost-sensitive AI workloads. It also remains unclear whether existing Capacity Block reservations placed before July 1 will be honored at prior rates or billed at the new schedule from that date forward.

With less than a week before the new rates take effect, enterprise buyers face an immediate decision point. Those with pending AI workloads may rush to lock in remaining capacity at current prices, while others will simply absorb the higher costs as a structural feature of the AI infrastructure landscape. Either way, the message from AWS is unmistakable: the AI compute crunch is not a temporary bottleneck but a defining feature of the current technology cycle, and the price of entry is still climbing.