Broadcom says AI companies can't make their own silicon

"Our custom accelerator business is progressing very well across five customers. Anthropic will soon implement one gigawatt of Broadcom-baked TPUs, and we expect the AI company plans a three-gigawatt deployment in 2027. Meta will install multiple gigawatts of Broadcom's XPU accelerators in 2027 and beyond. OpenAI will deploy over one gigawatt of compute capacity based on custom XPUs in 2027."

"They face tremendous challenges in terms of attracting silicon design talent capable of creating chips tuned to particular workloads, managing the production process, developing packaging expertise, and then networking their chips. Homebrew chipmaking efforts must create chips that are competitive with not just NVIDIA, but all the other LLM platform players that you are competing against."

"Anybody can design a chip in a lab that works well. Can you produce 100,000 of those chips quickly, at scale? That's the challenge that hyperscalers and AI companies cannot overcome for many years to come."

Broadcom reported 106 percent year-over-year growth in AI-related silicon revenue, reaching $8.4 billion in Q1 2026. The company is deploying custom accelerators across five major customers, including Anthropic's one-gigawatt implementation with plans for three gigawatts in 2027, Meta's multiple-gigawatt XPU deployment starting 2027, and OpenAI's over-one-gigawatt custom XPU capacity in 2027. CEO Hock Tan stated Broadcom has secured necessary supplies including high-bandwidth memory through 2028. Tan argued that hyperscalers and AI companies face insurmountable challenges in developing competitive custom silicon, lacking talent for chip design, production management, packaging expertise, and networking capabilities. He emphasized that producing chips at scale differs fundamentally from laboratory design, positioning Broadcom as the dominant supplier for years.

Read at Theregister