The war for AI talent has entered a new theater: silicon. According to Bloomberg, Anthropic has poached Amir Salek — the veteran chip architect widely known as the “Father of Google TPU” — who will report to James Bradbury, Anthropic’s Head of Compute. The high-profile hire lands just after Anthropic formally established an internal Custom Silicon division, signaling that the battle among AI giants has officially spread from the model layer down to the chip layer.
One Hire That Rewrites the AI Battlefield
Amir Salek is no ordinary executive. He joined Google in 2013 and spent nine years as the core soul of the TPU project, building Google’s custom silicon team from scratch and leading the delivery of the first seven generations of TPUs. The TPU (Tensor Processing Unit) is the compute engine Google built specifically for machine learning to escape its dependence on Nvidia — without it, Google would not have the confidence to rival Nvidia today.
Salek’s resume also spans both technology and capital. Before Google, he led Nvidia’s SoC design team under Jensen Huang. After leaving Google in 2022, he served as Senior Managing Director at top private equity firm Cerberus Capital Management, mastering the capital operations behind enterprise-scale chip projects. A rare combination of technical depth, management chops, and business acumen.
Not Just One Hire: OpenAI’s Chip Core Was Poached Too
Anthropic’s silicon play has been in the works for months. In June, Clive Chan — the second hardware employee on OpenAI’s secretive in-house chip project codenamed “Jalapeno” — quietly joined Anthropic. In other words, Anthropic now holds both the founding head of Google’s TPU effort and a core expert from OpenAI’s custom chip program, topped off by a brand-new Custom Silicon department established this month.
Why Custom Chips Are Non-Negotiable
Backed by Amazon AWS and Google Cloud, with seemingly unlimited funding, why would Anthropic bother building its own chips? The answer lies in the financials.
Anthropic’s estimated compute spending for 2026 is $19 billion, and its inference costs overran budget by 23% in 2025. The company’s 2025 gross margin of 40% sits far from the 77% target it told investors it plans to reach before an expected Nasdaq listing. With the “Nvidia tax” weighing heavily on margins, custom silicon becomes the inevitable path to cost recovery.
The second driver is hardware-software co-design. Just as Apple’s M-series chips outperform traditional processors at lower power through tight integration, Anthropic wants a “TPU for Claude”: engineers who understand Claude’s algorithms working hand-in-hand with chip designers, so future Claude versions could deliver far higher inference speed and intelligence at far lower cost than peers.
Beyond Chips: A Global Compute Land Grab
The poaching is only one facet of Anthropic’s infrastructure blitz. The company recently issued an initial commitment order worth up to $250 million to Fractile, a UK AI chip startup whose novel architecture promises extremely low-power LLM inference — locking in third-party capacity while its own chips mature. It has also struck a series of infrastructure deals involving Riot Platforms, one of the largest crypto miners in the US, and Volta Infra Holdings, snapping up repurposed mining sites and power capacity at bargain prices.
Anthropic hasn’t cut off its existing suppliers — it still buys large volumes of chips from Nvidia, Amazon, and Google. But with self-development, startup investment, and asset acquisition advancing on three fronts, an exclusive compute fortress is taking shape.
Conclusion: Whoever Masters Compute Defines AGI
The competitive logic of frontier AI has been rewritten. It used to be about algorithms and data, then about fundraising — now the war has reached chip manufacturing. OpenAI has Jalapeno, Google has TPU, and Anthropic now has a custom silicon team led by Amir Salek. Whoever first closes the loop of “model architecture + underlying silicon” will gain a cliff-edge advantage in inference cost.
As the market puts it: whoever masters compute defines AGI. This chip war has only just begun.
