Anthropic Hires Former Google TPU Chief for Custom AI Chip Push

Salek isn't a random senior hire off a chip-design org chart - he founded and led Google's TPU program from 2013 to 2022, shepherding the first seven generations of the chip that became the backbone of Google's AI infrastructure^[1]. After leaving Google, he spent roughly four years at Cerberus Capital Management, serving as Senior Managing Director and Partner of Tracker Ventures, the firm's stage-agnostic venture platform focused on deep-tech investments including semiconductors and AI/ML^[2][3]. Financial Times' Tech Tonic podcast put it bluntly, calling him "the brains behind Google's TPU chip"^[2]. Recruiting the person who built TPU from a standing start - rather than a chip-program manager brought in to execute someone else's roadmap - is a much stronger signal than a routine hiring push: Anthropic is betting on someone who has already solved the model-hardware co-design problem once, not someone learning it for the first time. On X, the news was framed less as a surprise and more as confirmation of an escalating "AI infrastructure war" that had already been rumored for months.

Weeks before the Salek hire became public, TechCrunch reported Anthropic was in early-stage discussions with Samsung about manufacturing a custom AI chip^[6]. Notably, as of that report Anthropic hadn't yet decided what the chip would be used for, how it would fit into the server, or how powerful it would be^[6]- the people are being assembled before the architecture exists, not the other way around. Analysts estimate a realistic launch window of 18 months to roughly 3 years out, landing around 2028, since new manufacturing capacity isn't expected until then and TSMC's advanced-node capacity is booked years in advance^[7]. That timeline puts any Anthropic-designed silicon well behind Google's seven TPU generations and Amazon's already-in-production Trainium chips - this is a multi-year bet, not a near-term Nvidia replacement.

Anthropic CFO Krishna Rao has pointed to "the exponential growth we have seen in our customer base" as the backdrop for the company's compute strategy - Anthropic's revenue run-rate is reported at roughly $30 billion, up from about $9 billion at the end of 2025, with more than 1,000 customers each spending over $1 million annually^[5]. To build out the team, job listings advertise salaries between $320,000 and $485,000 and explicitly want engineers "comfortable making consequential calls without a large organization behind them" - a tell that the group is still small^[5]. Anthropic isn't acting alone here: Google, Amazon, Microsoft, and Meta have all built custom AI silicon, and OpenAI unveiled its own Broadcom-built inference chip earlier in 2026 - the same industry-wide hedge against Nvidia pricing and GPU supply constraints^[4][6].

Every report on this story carries the same caveat: Anthropic says a diversified hardware stack including chips from Google, Amazon, and Nvidia "will continue to be pivotal to its compute strategy," and the company has separately committed to up to 1 million Google TPUs and as much as 3.5 gigawatts of TPU capacity beginning in 2027^[5][6]. That framing sits in some tension with how the news landed elsewhere online. A widely discussed Reddit thread asking why every AI lab is suddenly building its own chips took a more cynical view, debating whether in-house silicon is a genuine engineering necessity, a vendor-lock-in escape hatch, or - as some commenters argued - a narrative move timed around a possible Anthropic IPO, with Taiwan supply-chain and export-control risk raised as a driver most coverage doesn't dwell on. A more technical Reddit thread distinguished between GPUs, dedicated tensor-core chips, and "weights burned into silicon" designs (citing Cerebras and Taalas as real examples) - a reminder that "custom chip" spans a wide design space, and Anthropic hasn't said which end of it Salek's team is actually targeting.