Anthropic has hired Amir Salek, the founder of Google’s tensor processing unit (TPU) chip programme, to lead its push into custom AI silicon. Salek built and delivered the first seven generations of Google’s TPUs before leaving in 2022, and will now report to Anthropic’s James Bradbury as the company lays the groundwork for in-house chip design.

This is a meaningful strategic shift. Until now, Anthropic has relied on chips from Nvidia, Google, and Amazon to run Claude’s training and inference workloads. The company’s ambition to build its own silicon signals that those external dependencies are becoming a constraint at the scale it is targeting.

Why This Matters

Anthropic is not doing this for prestige. It is doing it because the arithmetic of running billions of tokens per day at a $30 billion revenue run rate makes compute costs one of the biggest variables in the business. Building custom silicon is the same play Google, Apple, Amazon, and Meta made before it, and for the same reasons: better efficiency, lower cost per query, and control over the performance roadmap.

The timing matters too. Anthropic is preparing to file for an IPO that it expects to rival or exceed SpaceX’s record offering in size, targeting a valuation around $2 trillion. Investors in a company at that scale want to see a credible path to margin improvement, and owning the hardware layer is one of the clearest levers available.

The hire of Salek specifically is a credibility signal. He did not just work on chips at Google, he built the TPU programme from nothing. Delivering seven successive generations of custom AI accelerators is a serious engineering track record, and Anthropic is not hiring him to study the idea.

What This Means for Business

For enterprise buyers evaluating Claude for production workloads, this development has practical implications.

Cost trajectory is improving. As Anthropic develops its own silicon, the cost of inference will fall. That means the per-query or per-token pricing that enterprise customers pay should decrease over time, making large-scale AI deployments more economically viable.

Claude’s capabilities will be co-designed with hardware. The real advantage of custom chips is not just cost reduction, it is the ability to optimise model architecture and silicon together. Apple’s Neural Engine and Google’s TPUs both improved model performance as well as efficiency because the hardware was built with the software in mind. Anthropic will likely pursue the same path with Claude.

Lock-in risk is reduced, not increased. There is a counterintuitive upside for enterprise customers here. Anthropic’s current dependence on third-party chip providers means that capacity constraints or supply chain issues at Nvidia, Google, or Amazon can affect Claude’s availability. A proprietary chip programme reduces that fragility over time.

The AI infrastructure gap is closing. The fact that Anthropic, still a relatively young company, is now building custom silicon alongside OpenAI (which has also signalled hardware ambitions) shows that the leading AI labs are maturing into full-stack technology companies, not just model providers. That changes how businesses should think about AI vendor relationships over a five-year horizon.

The Broader Picture

Anthropic’s move is part of a wider pattern. Microsoft, Amazon, Google, and Meta all invested heavily in custom silicon specifically to reduce their dependence on Nvidia. Now the model providers themselves are doing the same thing. The AI chip market is fragmenting, with more players building their own hardware to control more of the performance and cost curve from model architecture down to silicon.

For businesses building on AI platforms today, the key takeaway is that the economics of AI inference are going to improve faster than most current cost models assume. The companies investing in AI-powered operations now will benefit from that cost deflation without needing to rebuild their workflows.

Enterprise DNA works with businesses on exactly this kind of forward planning, helping leadership teams understand how the underlying economics of AI platforms translate into practical investment decisions. If you are evaluating AI infrastructure for your business, talk to the Omni Advisory team to make sure your roadmap accounts for where the technology is heading, not just where it is today.