Anthropic Books 250K Next-Gen GPUs in $10B Cloud Agreement

Anthropic has signed a $10 billion infrastructure deal with AI cloud startup Volta, securing dedicated access to 250,000 next-generation GPUs over the next five years. The agreement shifts the company’s compute strategy away from exclusive reliance on primary backers Google and Amazon, ensuring guaranteed capacity for upcoming training cycles.

Dedicated Infrastructure for Claude 4

The strategic partnership centers on Volta’s upcoming HyperScale-G clusters. Anthropic will utilize 250,000 GPUs built on NVIDIA’s Blackwell-successor architecture. Volta is constructing three new sovereign-grade data centers across North America and Northern Europe to house the hardware. These facilities are explicitly optimized for Anthropic’s proprietary Claude-4 training requirements.

Training models with parameters in the tens of trillions requires massive parallelization and minimal interconnect latency. Volta’s clusters feature V-Link 3.0 technology, delivering 1.6 Tbps of bandwidth per node. The deployment also relies on Volta’s proprietary Liquid-Grid cooling system. This thermal management design allows for a 30 percent increase in rack density compared to standard hyperscale environments, minimizing the physical distance between compute nodes to accelerate distributed training loops.

Infrastructure Metric	Volta HyperScale-G Specifications
Total Compute	250,000 next-generation GPUs
Interconnect Bandwidth	1.6 Tbps per node (V-Link 3.0)
Cooling Technology	Liquid-Grid (30% density increase)
Power Source	100% Small Modular Reactors (SMRs)

High-density GPU clusters require massive power provisioning. The agreement mandates that 100 percent of the operational energy for these facilities will come from dedicated small modular reactors. Volta is developing these nuclear power sources alongside its energy partners, providing uninterrupted baseline power for sustained AI workloads.

Compute Diversification

By committing $10 billion to a young cloud startup, Anthropic signals a move toward strict cloud-agnosticism. The company has historically secured hardware through deep partnerships with hyperscalers, including arrangements that trade equity for Google TPUs. While the company still relies heavily on AWS for Claude inference, owning dedicated training hardware prevents vendor lock-in.

Hardware bottlenecks at major cloud providers routinely delay frontier model development. Securing independent infrastructure guarantees that Anthropic will not have to compete with Amazon or Google’s internal teams for queue priority during critical training runs. Following the announcement, shares of major AI infrastructure providers dipped slightly as investors assessed the rising competition from specialized AI cloud startups.

Analysts project this deployment effectively doubles Anthropic’s available training compute for the 2027 fiscal year. This capacity aligns Anthropic directly with OpenAI’s Stargate infrastructure roadmap, ensuring both companies have the physical hardware necessary to test the limits of scaling laws.

The first phase of the HyperScale-G cluster goes online in Q1 2027 with Anthropic as the exclusive first tenant. If you allocate resources based on frontier model development cycles, the timeline indicates that the largest Claude 4 checkpoints will likely not begin full-scale training until these sovereign clusters are operational early next year.