On August 26, Anthropic signed a $45 billion, six-year deal with UK-based Nscale to rent compute capacity — taking the exact spot Microsoft had walked away from weeks earlier, on the same site in West Virginia. It's Anthropic's fourth major infrastructure commitment announced in 2026, and the episode reveals more about the industry's uncertainty than about its confidence.
What Changed
The deal gives Anthropic roughly 460 megawatts of power — enough to supply about 345,000 US homes — at Nscale's Monarch campus in Mason County, West Virginia, running Nvidia's Vera Rubin chips starting in late 2027. Before that, in March, Microsoft had signed a non-binding letter of intent for 1.35 gigawatts at the same site, then walked away mid-year without a public explanation, according to Bloomberg. Anthropic stepped in for the first of three planned buildings. Added to three other deals closed this year — $50B with Fluidstack, $10B with Volta Infrastructure, and $45B with SpaceX — Anthropic's total compute commitments for 2026 now exceed $169 billion, according to TechCrunch.
Why It Matters
A company switching cloud vendors isn't news. What stands out is Microsoft — which has the cash, a long-standing OpenAI relationship, and every reason to expand capacity — walking away from a 1.35GW commitment without explanation, while Anthropic doubled down on the very same site weeks later. That's a signal that even among the companies that best understand future AI demand, forecasting how much compute will actually be needed remains a bet, not a certainty. Anthropic is making this move ahead of an expected October IPO, reportedly valued at more than $2 trillion — which raises the pressure to justify, with results, a capital commitment of this scale.
The Impact for Brazil
For anyone in Brazil budgeting AI spend for the next few years based on today's prices and availability, the takeaway is caution: the scale of these commitments shows that AI compute supply is still being designed in real time, with billion-dollar bets and plan reversals happening within weeks of each other. That helps explain the pricing volatility we've already covered here — from DeepSeek to Gemini — and reinforces that a vendor's "guaranteed" contracts and capacity today aren't a solid basis for long-term cost projections. Finance and technology teams should treat AI cost as a still-unstable budget line, not a fixed line item.
Entercast's Take
This is the fifth chapter of a storyline we've been building all August: Anthropic paying for its own grid infrastructure, Google Cloud formalizing agent governance, AT&T cutting cost through routing, OpenAI testing its own chip against Nvidia — and now the very top of the chain (the capital financing all of it) showing real signs of uncertainty about market size. The practical lesson for whoever leads AI adoption at your company hasn't changed: negotiate contracts with exit flexibility, avoid single-vendor dependency on any one model or cloud provider, and treat the current phase of AI for what it is — a capital-intensive industry still discovering its own real demand, not a stably priced commodity.