vertical all the way down
The safety lab needs its own chips.
Anthropic is building a custom silicon team. The company confirmed it’s hiring engineers across hardware and software, and the job listing used the phrase “custom silicon team” without ambiguity. Samsung is being scouted as a manufacturing partner, according to The Information — Anthropic hasn’t confirmed it.
This follows a pattern so consistent it might be a law: every AI company that reaches sufficient scale discovers it cannot rent its way to competitive margins. Google figured this out first — TPUs shipped in 2016. Amazon followed with Trainium and Inferentia. Meta built MTIA. Microsoft and OpenAI are both designing custom silicon. Now Anthropic joins the list.
The arithmetic is straightforward. Anthropic is serving billions of tokens daily at a $30 billion revenue run-rate. At that volume, the spread between market-rate GPU costs and what a purpose-built inference chip could deliver is not a rounding error. It’s the margin. When your product is computation sold by the token, the cost of each computation is the entire business model.
What makes this worth noting is not the hardware play itself. It’s who’s making it.
Anthropic was founded to be the careful lab. The one that publishes alignment research, staffs a Responsible Scaling Policy, and talks about existential risk in investor decks. None of that changes the cost curve. Caution and capability both run on silicon, and silicon purchased at market rates from a single vendor is a dependency that scales linearly with your success.
This is not ambition. It’s survival math dressed up as a job posting.
The deeper pattern: vertical integration in AI is not a strategy some companies choose. It’s a phase transition that sufficient scale forces. You start buying GPUs. You start designing chips. You start looking for a fab. The stack wants to be owned, top to bottom, by anyone processing tokens at this volume.
Anthropic’s safety mission and Anthropic’s chip project are not in tension. They have the same root cause. The models got good enough that people want to use them. Everything else follows.
Sources: Anthropic reveals plan to develop custom chip, SiliconANGLE ↗ · Anthropic in talks with Samsung to manufacture custom AI chip, The Information via Bloomberg ↗