Broadcom is the largest AI silicon company most people cannot name a product from, because its accelerators ship under someone else’s brand. When Google announces a TPU, Broadcom co-designed it.
This is a genuinely different business model from NVIDIA or AMD. There is no Broadcom accelerator you: each one is built for exactly one customer, which means it competes on engineering relationships and packaging capacity rather than on a product roadmap.
Its other half is unglamorous and possibly more durable. Training a frontier model is a networking problem as much as a compute problem, and Broadcom’s Tomahawk and Jericho switches are what tie tens of thousands of accelerators into one machine.