AWS raises EC2 GPU Capacity Block prices ~20%, wins Warner Bros. Discovery and Fox

AWS raised prices for its EC2 Capacity Block reservations covering Nvidia P5 (Hopper) and P6 (Blackwell) GPU instances by roughly 20%, effective July 1, citing tight supply. It's the second increase in six months, following a 15% hike in January — a clear signal that GPU-training capacity remains scarce and that AWS has pricing power over customers who need guaranteed access to high-end Nvidia silicon.

The hikes prompted developer complaints of a compute-cost squeeze and thin inference ROI, summarized in community discussion as 'infrastructure pricing catching up to AI demand reality.' The tension is real: while model vendors compete furiously on dollars-per-task (GPT-5.6 Luna, Sonnet 5 cuts, DeepSeek V4), the underlying compute is getting more expensive, squeezing the margins of anyone building on rented GPUs. AWS's launch of the cheaper G7 (RTX PRO Blackwell) inference instances the same week is partly a response — offering a lower-cost on-ramp that doesn't compete for scarce P-series training capacity.

On the demand side, AWS scored major enterprise wins: Warner Bros. Discovery and Fox named it their preferred AI cloud provider for agentic ad-tech, validating AWS's Bedrock and agent-infrastructure strategy against Azure and Google Cloud. The combination — raising prices while winning marquee customers — reflects a seller's market where AWS can push rates and still grow enterprise share. The broader context is the capital-intensity theme dominating the week: Amazon planning a $25B raise, Nvidia renting back inventory, Meta and DeepSeek building chips. The strategic risk for AWS is that sustained price hikes accelerate customers' motivation to self-host on open models (as Hugging Face's CEO argued) or build custom silicon — the very dynamic eroding Nvidia's, and by extension AWS's, long-term pricing power.