Anthropic locks multi-GW compute from AWS, Google and SpaceX

Anthropic disclosed on May 28 that its $65 billion Series H round values the company at $965 billion and that the new capital will be used largely to expand compute to meet surging demand for Claude.

In the same Series H materials the company detailed infrastructure commitments: agreements with Amazon Web Services for up to five gigawatts of new capacity, with Google and Broadcom for five gigawatts of next‑generation TPU capacity, and arrangements to access GPU capacity in SpaceX’s Colossus 1 and Colossus 2 facilities.

Anthropic framed those deals as central to its strategy, saying the funding will “expand compute to meet growing demand for Claude” and that compute scale is critical to sustain research and commercial performance.

The SpaceX agreement in particular shifts raw GPU supply. Reporting and filings indicate Anthropic arranged access to the full capacity of the Colossus 1 cluster — a facility described in coverage as more than 300 megawatts and housing well over 200,000 NVIDIA GPUs — a move that can immediately raise Claude’s service and training limits.

Financial terms tied to the Colossus deal have emerged in filings and reporting: several outlets have reported a headline figure of about $1.25 billion per month paid to SpaceX (roughly $15 billion a year) under arrangements running multi‑year, though not every public report cites identical contractual dates.

The Google/Broadcom arrangement is more complex in public filings than the Series H summary suggests. Anthropic’s announcement cites five gigawatts of TPU capacity; Broadcom’s filings and subsequent coverage describe a multi‑GW supply path that includes roughly 3.5 gigawatts scheduled from 2027 and provisions tied to future TPU designs and commercial contingencies.

Amazon remains central to Anthropic’s cloud plan: the company says AWS is its primary cloud provider and training partner even as it diversifies capacity across Cloud providers and specialized clusters. The multi‑cloud footprint lets Anthropic route workloads and manage regional availability while keeping a large presence on a primary hyperscaler.

Taken together, these contracts illustrate why compute has become a defensible moat for frontier LLM providers. Locking long‑duration access to gigawatts of training and inference capacity reduces supply risk, speeds iteration on larger models, and raises the bar for rivals that lack similarly firmed capacity.

Those deals also reshape datacenter GPU demand. Providers of racks, interconnects, memory, and power systems will see demand lumpy and concentrated, driven by a handful of large commercial commitments rather than many small cloud reservations. Analysts and datacenter reporters say that concentrates purchasing, shortens lead times for some hardware classes, and amplifies the importance of supply‑chain partners.

Strategically, the SpaceX access is notable because it converts a competitor’s idle or spare capacity into a resource for Anthropic — and because it illustrates alternative paths to GPU scale beyond the three mainstream hyperscalers. That changes the competitive map for frontier models, where raw compute and the ability to iterate on model size and training regimes matter.

There are still limits and risks. Long, expensive capacity deals raise fixed costs and hinge on reliability, deployment timelines, and contract terms — for example, Broadcom’s filing ties some capacity to future TPU generations and commercial performance, making parts of the promise conditional. Those contractual details will matter if demand or model economics shift.

For enterprise customers, the immediate effects are practical: Claude is now offered across AWS, Google Cloud, and Microsoft Azure, and Anthropic says the compute increases support higher usage limits and new product capabilities. For the industry, the headline takeaway is that companies are treating raw, guaranteed compute as a strategic resource and that capital raised in May will be funneled straight into infrastructure scaling.