Broadcom CEO Hock Tan said the company expects AI revenue to reach $115 billion in 2027 and $230 billion in 2028, while free cash flow should reach the mid-40s billions of dollars in 2027 and be higher in 2028. Speaking at the Goldman Sachs Communacopia + Technology Conference on Tuesday, September 8, 2026, in a session with Goldman Sachs semiconductor analyst Jim Schneider, Tan said demand from frontier AI developers remains strong but is being constrained by power availability, packaging limits and the cost of silicon respins.
Over the last twelve months, Broadcom reported revenue of $89.1 billion, up nearly 49% year over year, and generated $39.4 billion in levered free cash flow. The company carries about $56 billion in debt, with a debt-to-equity ratio of 0.6. Broadcom stock closed at $368.56, up 2.98% from the previous close of $357.90, and traded at $368.43 after hours. The dividend yields 0.73%, and the company has raised its dividend for 16 consecutive years.
To support compute buildouts, Broadcom created a special purpose vehicle with Apollo and Blackstone to provide $35 billion in initial financing for more than 20 gigawatts of compute capacity. Tan said Broadcom works closely with a concentrated group of six frontier model developers, including OpenAI, Google and Anthropic. The company has a long-term agreement with Google through 2031 for TPU and networking products, and signed a supply agreement with OpenAI for 10 gigawatts of custom silicon capacity in October. Anthropic is expected to receive 1 gigawatt of capacity in 2026, rising to 5 gigawatts in 2027. The company supplies custom silicon, accelerators and networking products to these customers, and the discussion also referenced Claude Fable 5 and Mythos Preview.
Tan said preparing power sites for use by 2028 requires about two years of construction lead time. A single silicon respin can cost about $30 million and delay a project by six months. The discussion referenced XPU accelerators, the Jalapeño ASIC, SerDes and TPU products. Broadcom delivered a Jalapeño ASIC sample to OpenAI within nine months of the initial agreement. On packaging, Tan said a single die has a practical maximum size of about 800 square millimeters. Current chips use two dies per chip with about 1,600 square millimeters of total area, while development is moving to four dies and future versions could reach eight dies and more than 6,000 square millimeters in equivalent area.
Tan also outlined the economics of the AI market. He said global inference token generation costs about $200 billion a year, while total AI-related revenue is about $150 billion a year. Frontier model providers generate about 50% of tokens but capture 75% of revenue, roughly $120 billion, while open-weight model providers generate about 50% of tokens but capture only about $30 billion in revenue. Both groups spend about $100 billion on infrastructure. One gigawatt of frontier-model compute capacity generates about $30 billion in annual recurring revenue, with operating costs of about $10 billion a year, leaving $20 billion for model and application developers and $10 billion split among power, cloud, chip and memory providers.
Internally, Broadcom evaluated generative AI tools over a one-year period involving 15,000 of its 30,000 engineers. Tan said the tools "help very smart people become smarter," but added: "The biggest issue is finding use case that generates what you call a return on investment that is clearly meaningful and significant. That's not easy to do, and I suspect a lot of companies out there face the same issue." He said value will accrue over time to companies producing the best frontier models as long as large language models continue improving. On cash generation, Tan said: "We're going to generate a lot of cash. Your point exactly. More so in 2028."
