Anthropic is no longer just renting compute. It is building the machine.

Sana Ouji, who led strategic investments and energy alliances for Google's data center division, has left to join Anthropic's newly formed energy team. She will work alongside Ariel Horowitz and Tim Hughes on what Anthropic describes as a global energy strategy aimed at "scaling responsibly and swiftly" its data center portfolio. But Ouji is not an isolated hire. She is the tip of a systematic raid.

According to DatacenterDynamics, Anthropic has spent months recruiting Google veterans across every layer of data center operations. Winnie Leung now leads data center infrastructure. Liwen Mao handles data center design. Adam Johnson runs electrical engineering. Zach Miller oversees operations. Tim Hughes and Matt Wanner both came from Stack Infrastructure, bringing build-and-operate experience. The job postings tell the same story: Data Center Portfolio Planning Lead, Senior Capacity Delivery Manager, Electrical Engineer, Resource Efficiency Engineer.

Three layers of compute, simultaneously

Anthropic is now running three compute strategies at once. Layer one: contracted cloud with hyperscalers, maintaining high-volume relationships with Google and others. Layer two: dedicated infrastructure, including the $50 billion deployment announced in November 2025 with Fluidstack, spanning data centers in Texas and New York with more locations planned. Layer three: the one that matters most. An internal team capable of negotiating energy deals, planning portfolios, executing construction, and operating facilities like a major data center company.

The scale of contracted capacity is staggering. Reuters reported in October 2025 that Anthropic signed with Google for over 1 GW of TPU-based compute, expected operational in 2026. By April 2026, the company announced a new deal with Google and Broadcom for several gigawatts of next-generation TPU capacity starting 2027. Industry sources estimate this expansion around 3.5 GW. For context, that is more power than some small countries consume.

The real war is megawatts, not models

This strategic pivot comes with a sharp edge. An internal OpenAI memo leaked to The Verge accused Anthropic of a "strategic error" for not securing enough compute capacity, claiming it had impacted service availability and reliability. Anthropic's Claude has indeed suffered notable outage periods. But the accusation reads differently now. Anthropic was not ignoring the problem. It was building an entire internal division to solve it permanently.

Dario Amodei has publicly defended Anthropic's measured approach. In February, he warned that if a company misjudges growth by even a year, or if the market grows five times faster than expected instead of ten, buying too much capacity too early could bankrupt it. This is not caution for its own sake. It is the calculation of a CEO who watched competitors light money on fire and decided to hire the people who actually know how to build and operate the infrastructure instead.

The AI race used to be about who had the best model. Then it became about who had the most GPUs. Now it is about who controls the power. Anthropic just hired the team that knows where the power is and how to get it. That may matter more than any benchmark score.