Broadcom's Hock Tan: Custom Silicon Moat Widens as AI Revenue Path to $230 Billion Hinges on Power, Not Chips
Goldman Sachs Communacopia + Technology Conference, September 8, 2026
Broadcom CEO Hock Tan used a fireside chat with Goldman Sachs semiconductor analyst James Schneider to sharpen the narrative around the company's AI ambitions, offering unusually candid detail on what is actually constraining its supply-limited forecast of $115 billion in AI revenue for 2027 and $230 billion for 2028. The bigger news for investors is not the growth math itself, which Broadcom laid out on its earnings call last week, but Tan's explanation of where the real bottlenecks sit, how deep the company's technical moat against customer in-sourcing really is, and a striking data-driven argument for why value in AI will keep concentrating in frontier closed models rather than open-weight alternatives.
Power, Not Wafers, Is the Real Constraint
Tan was explicit that chip and memory supply, while tight, is the more manageable half of the equation. "That's pretty well locked in. We can pretty well lock that in for sure in '27 and the process of locking in '28 gives us -- within some limits," he said of wafers and memory. The harder constraint is physical infrastructure: power availability, site readiness, construction timelines and permitting. Tan noted that making a site power-ready by 2028 requires customers to break ground now, and that construction at scale faces the same friction as renovating a house, just magnified. This is a meaningful data point for investors modeling the durability of Broadcom's AI backlog: the company is effectively saying its own forecast is capped not by its execution but by how fast hyperscalers can pour concrete and secure grid capacity, a slower-moving and less predictable variable than semiconductor manufacturing.
The Moat Against Customer In-Sourcing
Addressing investor concern that hyperscale customers like Google could eventually bring custom silicon design fully in-house, Tan pointed to Broadcom's breadth of IP across high-speed SerDes interconnect, multiplier density, multi-die communication and advanced packaging as barriers competitors and customers alike would need years to replicate. He also disclosed a specific technical roadmap: Broadcom is now stacking multiple dies into a single chip to work around the physical reticle limit of roughly 800 square millimeters per die. The current generation uses four dies per chip, effectively creating a 3,200 square millimeter compute package, with an eight-die generation already in sight, implying chips equivalent to 6,400 square millimeters. This is Broadcom's answer to the end of Moore's Law scaling, and it is a level of architectural detail the company has not laid out this explicitly before. Tan framed it as the core reason the Google TPU relationship, formalized through a long-term agreement extending supply through 2031, keeps deepening rather than eroding.
OpenAI's Jalapeno and the Speed of Custom Co-Design
Tan confirmed that Broadcom delivered its first ASIC for OpenAI, code-named Jalapeno, within nine months of signing a 10-gigawatt custom silicon supply agreement last October, and said the chip performs as well as or better than the best general-purpose accelerators on the market. He attributed the speed to tight engineering integration between Broadcom's design teams and OpenAI's model architects, and disclosed that the two companies are already working on the next two generations beyond Jalapeno. This is a concrete signal that Broadcom's custom silicon relationships are not one-off engagements but multi-generation programs, reinforcing the durability of the revenue base underlying its 2027-2028 guidance.
The $35 Billion Financing Vehicle Explained
Tan gave his most detailed public defense yet of Broadcom's special purpose financing vehicle with Apollo and Blackstone, which provides an initial $35 billion toward more than 20 gigawatts of compute capacity for AI labs including Anthropic, starting with one gigawatt this year and scaling to five in 2027. He pushed back directly on characterizations of circularity: "This is not circular financing... We're using financing to create demand. Demand is there." Under the structure, financial partners take on the credit risk of lending to labs like OpenAI and Anthropic, while Broadcom provides residual value guarantees on the underlying equipment. Tan's rationale is that two of Broadcom's six frontier-model customers lack the cash flow to fund compute buildout themselves, and Broadcom would rather enable that demand through structured financing than let it go unmet or migrate to a competitor.
Why Open-Weight Models Won't Win the Value Chain
The most quantitatively rich section of the conversation was Tan's breakdown of token economics using OpenRouter-Vercel data. He estimated the industry currently spends roughly $200 billion annually generating inference tokens, against roughly $150 billion in revenue attributable to model providers. Splitting that data, frontier closed models account for less than half of tokens generated but about 75% of revenue, translating to roughly $100 billion in costs against $120 billion in revenue. Open-weight models, by contrast, generate more than half of tokens but only about $30 billion in revenue against a similar $100 billion cost base. "Do you think that's a sustainable model? We don't know, but it proves one thing. The value goes to where intelligence continues to improve," Tan said. The framing is a direct rebuttal to the bull case for open-weight models displacing frontier providers, and by extension supports Broadcom's thesis that value, and therefore silicon spend, will keep concentrating with the handful of labs pushing capability frontiers forward.
Capital Allocation Signal Ahead of December Board Meeting
Tan flagged that Broadcom will end fiscal 2026 with a record cash balance and is on track to generate free cash flow in the mid-$40 billions in 2027, growing further in 2028 as AI revenue scales. With $56 billion of debt carrying favorable legacy rates that make prepayment unattractive, Tan indicated the board's December capital allocation review will likely favor increased dividends and stock buybacks over debt reduction, though he stopped short of committing to specifics.