Anthropic Custom AI Chip Push Adds Google TPU Founder Amir Salek

A TPU architect changes sides

Anthropic has hired Amir Salek, who founded Google's Tensor Processing Unit program, to join the compute team behind its new in-house chip effort. Bloomberg first reported the hire on August 21. Salek reports to James Bradbury, Anthropic's head of compute.2123 He led Google's TPU work from 2013 to 2022 and oversaw the first seven generations of the chip.2224 Before Google, he spent eight years at Nvidia, where he founded and grew its system-on-a-chip organization.2430 After leaving Google he worked as a senior managing director at Cerberus Capital Management, investing in semiconductor, AI and other deep-tech companies.2129

Anthropic did not announce the hire with a press release. The Economic Times, carrying Bloomberg's story, says the company itself disclosed the move on Friday.26 Runtime Wire likewise describes the hire as confirmed by Anthropic.29 One trade outlet checked Anthropic's newsroom several days later and found no post about Salek or hardware hiring. It concluded the story rests on Bloomberg's reporting, which Anthropic apparently engaged with.28 With Bloomberg carrying the company's statement and several outlets repeating it, the hire looks solid. What is still missing is any official word on the program's scope, budget or timeline.

Part of a deliberate build-out

Salek's hire is the latest step in a hardware push that has been taking shape for months, not a one-off. On August 5, Anthropic confirmed for the first time that it is building a "custom silicon team." Business Insider had spotted a job listing for a senior engineer who had shipped chip designs.31 That listing pays $320,000 to $485,000. It asks for someone with a "realistic relationship with schedules" who can make major decisions "without a large organization behind them."3 That wording points to a small team at the start.

The talent hunt started before that. In June, Clive Chan, who led OpenAI's custom chip program, left to join Anthropic.5 Memeburn reports that Chan spent more than two years on OpenAI's Broadcom-built Jalapeño inference chip and earlier worked on Tesla's Dojo supercomputer.6 TheStreet notes that Anthropic has now hired senior people from the two companies furthest along in custom AI chips, and argues the pattern reflects a strategy rather than opportunism.25 That reading holds up. Chan brings recent experience taking a chip to tape-out, the point where a finished design goes to manufacturing. Salek brings more than a decade of running a chip program through repeated generations at data-center scale. These are the two skills a new chip effort needs most.

Manufacturing is less settled. The Information reported that Anthropic had early talks with Samsung as a possible manufacturing partner.13 Memeburn says those talks focused on Samsung's 2-nanometer process and packaging.6 Other coverage stresses that no foundry has been confirmed.7

The real target is the cost of each token

Many headlines framed the move as Anthropic trying to break free of Nvidia. The more careful coverage puts the cost of inference, meaning the cost of running models for users, at the center. Anthropic's spokesperson described the goal as co-design: building chips and Claude models together so Claude runs "faster and more efficiently" at the scale customers need.310 Forbes argues the move is driven less by leaving Nvidia than by the math of serving billions of tokens a day. In Forbes' reading, custom chips are a cost program aimed at the price of each token.2

The size of the possible savings is reported inconsistently. TheStreet, citing TechCrunch, says the goal is to cut inference costs by roughly half.25 Memeburn repeats the 50% figure.6 The company's own statements to Business Insider and others mention speed and efficiency but no number.3 Treat the 50% figure as a reported goal, not a promise.

The reasoning holds even without a specific figure. Chips built to run every kind of workload carry circuitry that a single company's models may never use. A chip designed around one model family can drop that overhead.2 The catch is cost. Reuters, cited in several reports, estimates that designing an advanced AI chip can run about $500 million before manufacturing starts.98 Only companies with very large revenue can absorb that. Anthropic now qualifies: its annual revenue run rate passed $30 billion this spring, up from about $9 billion at the end of 2025.1418 One venture-capital outlet adds that Anthropic reported more than $11.5 billion in second-quarter revenue and positive adjusted operating income. It frames custom chips as a margin story that fits Anthropic's reported IPO preparations.4

The Google problem

The most unusual part of the story is that Google is both where Salek came from and one of Anthropic's biggest suppliers. In October 2025, Anthropic agreed to use up to one million Google TPUs, with more than a gigawatt of capacity coming online in 2026.1229 In April it expanded that deal with Google and Broadcom to multiple gigawatts of next-generation TPU capacity starting in 2027. Anthropic called it its largest compute commitment so far.1113 A Broadcom securities filing put the Anthropic portion at about 3.5 gigawatts. The filing also said Anthropic's use of that capacity depends on its "continued commercial success."1618 Mizuho analysts estimated Broadcom could earn about $21 billion in AI revenue from Anthropic in 2026 and $42 billion in 2027.15

TheStreet points out that Google is supplying a direct competitor to its own Gemini models while that competitor hires the person most responsible for Google's chip program.25 Because Salek left Google years ago, there is probably no contract or legal barrier to the move.25 Futurum, writing before the hire, argued that Anthropic's main advantage was the close feedback loop between Claude's software and Google's TPU hardware, built up since 2022.14 Hiring Salek suggests Anthropic wants to own that kind of tuning itself instead of relying on a partner's roadmap. That reading is our analysis, not something the company has said.

Buying from everyone while building its own

None of this means Anthropic is pulling back from outside suppliers. The company calls its chip team "the latest step in our multi-chip approach" and says it will keep using a diversified hardware stack.5 It says hardware from AWS, Google, Nvidia and AMD remains central to its growth.3 Amazon is still its primary cloud and training partner.18 Anthropic is also adding suppliers. It placed an initial order of about $250 million with UK inference-chip startup Fractile, with room to expand, and signed data-center capacity deals with Riot Platforms and Volta Infra Holdings.2123 Memeburn also cites a $5 billion AMD arrangement covering up to 2 gigawatts of MI450-series compute.6

This is where coverage splits on what it means for chip stocks. 24/7 Wall St. reads the chip team as a sign that demand is so high no single supplier can meet it, which it calls good news for the whole sector, Nvidia included.9 TheStreet sees slower pressure building: each AI lab that designs its own chips shrinks the group of top customers fully dependent on Nvidia.25 Forbes takes a middle position. It argues that custom chips shift spending within the chip industry rather than removing it, toward chip-design partners like Broadcom, Arm-based processors, networking gear and TSMC's factories.2 Forbes has the most convincing case. Anthropic's spending is moving across suppliers, not shrinking.

When the chip arrives

The biggest gap in the story is timing. TheStreet says Anthropic's chip is not expected before 2028 at the earliest.25 Memeburn cites typical three-to-five-year development cycles and points to 2028 or 2029.6 Another analysis suggests 18 to 24 months from design to production.7 A venture-capital note warns that if the timeline slips past an IPO, Anthropic's margin story will depend on Nvidia and TPU pricing in the meantime.4 Rivals are ahead. OpenAI says its Jalapeño chip went from first design to tape-out in nine months, with first shipments expected by the end of 2026.5 TheStreet adds that Microsoft and Amazon already run their own chips and that Meta's Iris chip is scheduled for production.25

The Salek hire won't change Claude's pricing or availability this year. One outlet tells buyers to expect no near-term effect.28 What it does show is that Anthropic now treats chip design as a core skill rather than something to outsource. Runtime Wire puts the next test well: whether Anthropic can build a chip program that keeps producing new generations, rather than one expensive experiment.29 Salek's record at Google is about exactly that. Anthropic's next major milestone will be a confirmed manufacturing partner or a chip sent to production.