Broadcom CEO: AI Firms Can't Go It Alone
Broadcom's CEO asserted that "AI companies can’t make their own silicon," pushing back on the trend of tech giants developing custom ASICs. The statement argues that the deep expertise, tooling, and capital needed for advanced chip design creates a persistent moat for established semiconductor firms.
Broadcom CEO Hock Tan's assertion came during the company's Q1 2026 earnings call, where he revealed major deployments with key AI players. He specified that Anthropic will implement a one-gigawatt Broadcom-based TPU system, with plans for three gigawatts in 2027, while OpenAI is set to deploy over one gigawatt of custom XPUs in the same year. The statement is backed by significant financial momentum, with Broadcom's AI-related silicon revenue surging 106% year-over-year to $8.4 billion in the first quarter. Tan projected a clear "line of sight" for AI chip revenue to surpass $100 billion in 2027, highlighting major deals with clients like Google for its next-gen TPU and Meta for its custom MTIA accelerators. This pushback targets a growing trend of "in-house" silicon development at major tech firms, including Google's Tensor Processing Units (TPUs), Amazon's Trainium and Inferentia chips, and Meta's own AI accelerators. Interestingly, Tan confirmed that Meta's custom accelerator roadmap is "alive and well" and that Broadcom is currently shipping chips for the project, illustrating the deep integration between the firms. The core of the argument rests on the immense barriers to entry in advanced semiconductor manufacturing. Development and productization costs for new chips have grown exponentially, with projections for 2nm designs exceeding $1 billion. Establishing a single leading-edge fabrication plant now costs around $10 billion, a figure that doesn't include the complex supply chains for specialized materials and extreme ultraviolet (EUV) lithography equipment. Broadcom's business model for AI is not just selling off-the-shelf chips; it specializes in custom ASIC design and execution, holding over 60% of the AI ASIC market. The company acts as a crucial partner, translating a client's design, like Google's TPU, into a physical chip layout that can be mass-produced by foundries like TSMC. Tan emphasized the difference between a working lab prototype and mass production, questioning if a company can "produce 100,000 of those chips quickly, at yields that you can afford." To this end, Broadcom has secured its own supply chain, including critical high-bandwidth memory (HBM), with enough capacity to meet demand through 2028.
