Nvidia Blackwell GPU Cloud Costs Jump 20% in New AWS Pricing

For machine learning engineers and AI researchers, the sticker price on guaranteed Nvidia Blackwell compute just got meaningfully bigger. AWS’s quiet 20% hike on EC2 Capacity Blocks for ML—effective July 1, 2026—means that locking in a P6-B300 instance now costs $14.04 per accelerator hour, up from roughly $11.70. This isn’t just a rounding change; it’s a direct hit to training budgets and a clear signal that the GPU supply crunch isn’t easing anytime soon.

Amazon Web Services has quietly raised the price of EC2 Capacity Blocks for ML, the reservation product that lets customers lock in accelerator capacity for machine learning workloads. Effective July 1, 2026, rates for some of the most sought-after Nvidia GPU families—including Blackwell, H100, and H200—are climbing by roughly 20%, according to AWS pricing data and market reports. The move, while not a broad increase across all AWS products, targets the very heart of the AI infrastructure economy: guaranteed, dedicated compute for model training and fine-tuning.

“

AWS’s quiet 20% hike on EC2 Capacity Blocks for ML—effective July 1, 2026—means that locking in a P6-B300 instance now costs $14.04 per accelerator hour, up from roughly $11.70.

The specific numbers tell a clear story. In non-GovCloud Regions, the P6-B300 capacity will move to $14.04 per accelerator hour, and the P6-B200 will move to $12.355 per accelerator hour. These are the rate cards for the top-tier GPU instances that power large-scale AI workloads. AWS’s own explanation is straightforward: “Amazon EC2 Capacity Blocks for ML reservation prices are updated periodically based on supply and demand.” That simple statement encapsulates a market reality that has been building for two years. The AI arms race has not abated. GPU supply remains constrained, and major cloud providers—AWS, Microsoft Azure, and Google Cloud—are wielding pricing power as they ration the latest hardware.

For investors, the quiet hike is a signal. Amazon is telling Wall Street that the AI boom comes with a very real price tag, and that the company can monetize its infrastructure aggressively. The 20% increase on reservation prices suggests that demand continues to outstrip supply for Nvidia’s most advanced accelerators. It also hints at margin expansion opportunities within AWS’s high-performance compute segment. Nvidia itself benefits indirectly: every Price increase on cloud instances that run its GPUs reinforces the value of its hardware, and the high utilization rates confirm that customers are willing to pay a premium.

The impact extends across the AI ecosystem. Startups and smaller AI labs, which often depend on reserved capacity to control costs, will see their compute budgets squeezed. A 20% jump on a critical input can trim runway by months if not managed carefully. Meanwhile, large enterprises with multi-year commitments may be insulated for now, but the direction of travel is clear: cloud AI infrastructure is getting more expensive, not less. This could accelerate the search for efficiency—model optimization, quantization, and smaller architectures that require less brute force. It may also push some organizations to explore alternative cloud providers or even on-premise GPU clusters, though the scarcity of Blackwell chips makes that a challenging route.