Anthropic is hiring an AI chip design team to co-design hardware with its models
The stated goal is co-designing silicon and models together so Claude runs faster and more efficiently. A model company building a hardware team is a statement about where it thinks the constraint is.
Co-design is the operative word. This is not a plan to compete with NVIDIA on general-purpose accelerators — it is an argument that knowing your own architecture lets you specialise the silicon around it, and that the resulting efficiency is worth owning rather than renting.
The precedent is Google's TPU programme, which took years to matter and then mattered enormously. The counter-precedent is every other company that concluded custom silicon was cheaper than buying it and discovered the schedule was measured in generations, not quarters.
It also reads as a hedge against the supply structure. When accelerator vendors are taking equity in model companies and forecasts for 2028 demand swing by $900 billion in a quarter, controlling your own roadmap starts to look less like vertical integration and more like risk management.