Anthropic is building a team to design its own AI chips, with the intent of co-designing hardware and model to run together. This isn't only about saving money: it's about performance nobody else has.
Anyone buying compute every month recognises the move. An off-the-shelf accelerator is good but generic: it runs everything and is perfect for nothing. A frontier model is an extreme workload; it needs a specific memory fabric and interconnect.
Designing the chip and the software at the same time gives you an advantage that buying parts cannot. Every multiplier, every memory path shaped for your model's format. The difference in efficiency per watt is enormous, and efficiency per watt is cost.
There's a second effect, less discussed. A company controlling the chip, the weights and the inference closes the loop. Nobody outside replicates that efficiency.
AI technology concentrates further in the hands of whoever controls the factory. Meanwhile, the rest of the market keeps paying list price.