Anthropic has recruited former Google manager Amir Salek for its compute team, advancing the establishment of its own chip division. Salek led Google’s Tensor Processing Unit program from 2013 to 2022, delivering seven generations of chips during that time. He will now report to compute chief James Bradbury.
Salek was responsible for seven TPU generations at Google
Salek is considered one of the architects of Google’s Tensor Processing Unit program, specialized chips that the company uses to train and run its own AI models. From 2013 to 2022, he led the business and was responsible for delivering the first seven TPU generations. Before his time at Google, Salek worked at graphics chip maker Nvidia. After leaving Google, he joined the investment firm Cerberus Capital Management, where he most recently served as Senior Managing Director.
The news of his move was first reported by Bloomberg reporter Dina Bass on the platform X. At Anthropic, Salek joins the compute team and reports to James Bradbury, who is responsible for the company’s compute strategy. Engineers with his experience are rare: only a handful of specialists worldwide have been responsible for an entire TPU chip generation from architecture to mass production – let alone seven of them. This kind of knowledge of tightly interlocked chip and software development – known in the industry as co-design – is considered a prerequisite for wringing meaningful efficiency gains from specialized chips compared with general-purpose graphics processors.
Anthropic orders additional chips from Fractile
The move fits into a series of billion-dollar compute deals Anthropic has been closing since the summer of 2026. Already in August, the company confirmed it was building its own chip design team – led by former OpenAI engineer Clive Chan. With Salek, that effort now gains a second high-profile lead with decades of experience in TPU development.
Anthropic has also placed an initial order worth around $250 million with British chip startup Fractile, according to a report by Business Standard. The figure has not been independently verified. A later expansion of the order is reportedly planned. The AI provider currently sources compute from Nvidia, Google, and Amazon; Google recently linked roughly $200 billion in credit financing with partners for further Anthropic chip purchases. Anthropic also holds capacity contracts with Riot Platforms and Volta Infra Holdings. Rival OpenAI is pursuing a similar strategy with the Broadcom-codesigned inference chip Jalapeño, expected to enter use later this year.
The timing is no coincidence. Shortly before Salek’s move, Fortune reported that Nvidia had told customers to expect price increases of more than 15 percent for upcoming chip generations like Vera Rubin and Grace Blackwell. That points to rising cost pressure across the industry. What matters now is whether Anthropic can actually turn its new hire’s experience into a market-ready chip, or whether the effort – like many AI labs’ chip projects – needs years before a first delivery.
