On Monday, Anthropic announced a massive long-term deal with Google and Broadcom for custom TPU chips. By Wednesday, Reuters reported that Anthropic is also exploring the possibility of designing its own chips entirely from scratch.

Three sources told Reuters the plans are early-stage. Anthropic might still decide to stick with buying chips from partners and never build its own. But the fact that it is even exploring custom silicon tells you everything about where the AI industry is heading: every major lab is becoming a hardware company whether it wants to or not.

The context makes this move obvious. OpenAI just told investors it plans to have 30 gigawatts of compute by 2030. In a shareholder memo that leaked this week (because of course it did), OpenAI directly attacked Anthropic for operating on a "meaningfully smaller curve," estimating Anthropic will have just 7 to 8 gigawatts by the end of 2027. OpenAI called its own approach a "compounding advantage."

That is a lot of trash talk from a company burning $85 billion a year. But the compute gap is real. And Anthropic knows it.

Anthropic's $50 billion commitment to US computing infrastructure, announced alongside the Google/Broadcom deal earlier this week, is the company's answer. Custom chips could be the next step. Google designs TPUs. Amazon designs Trainium. Microsoft has its own Maia chips. Apple has been designing silicon for over a decade. The playbook is clear: if you depend on someone else's hardware, you are at their mercy.

For Anthropic specifically, there is also a Pentagon-shaped problem. The US government just won an appeals court ruling keeping Anthropic's supply-chain-risk designation in place. The DC appeals court said Anthropic "has not satisfied the stringent requirements" to remove the blacklist label, directly contradicting a San Francisco judge who found the Pentagon acted in bad faith. Being locked out of government contracts while also being compute-constrained is a two-front war Anthropic cannot afford to fight forever.

Custom silicon would give Anthropic something it desperately needs: control. Control over cost per token. Control over supply chain. Control over the pace at which it can scale Mythos and future models without waiting for Nvidia, Google, or Broadcom to deliver. In a world where OpenAI is publicly mocking your compute capacity to investors, "early stage exploration" of chip design sounds a lot like survival planning.

The AI chip war just got another entrant. And if even the "safety first" lab is now designing hardware, you can stop pretending this industry is going to slow down anytime soon.