Broadcom prices the custom silicon decade

The interesting number in an AI chip forecast is rarely the revenue. It is whether the company has already secured the manufacturing inputs required to deliver it, because everything upstream of a finished accelerator is booked years in advance.

Broadcom's forecast, given by chief executive Hock Tan alongside third-quarter fiscal 2026 results, would take the company's AI-related chip sales from $2.6 billion in fiscal 2022 to $230 billion in fiscal 2028. The Next Platform, which laid the figures out, notes that this is two orders of magnitude in six years.

The line covers custom accelerators, AI networking, and the surrounding components that go into AI clusters. Grouping them is not padding. A custom accelerator that cannot be connected to forty thousand others at line rate is a research project, and the networking is where a large part of the difficulty lives.

What makes the forecast more than a slide is the procurement behind it. Broadcom has secured the high-bandwidth memory, the wafers and the packaging capacity to fulfil the revenue stream, which converts the number from a demand estimate into a supply commitment.

That distinction is the one to apply to every AI infrastructure projection this year. Demand for accelerators is not the constraint and has not been for two years. HBM allocation, advanced packaging slots and foundry capacity are the constraint, and they are contracted well ahead.

The company's position was not built during the boom. Broadcom had been helping Google design its tensor processing units and move them through Taiwan Semiconductor Manufacturing Company's foundry and packaging for seven years before generative AI arrived.

Seven years of doing unglamorous co-design work for one customer is what a moat looks like before anyone calls it one. The capability that matters is not a chip design; it is the accumulated knowledge of getting somebody else's chip through a fab.

The strategic consequence is that the counterweight to a merchant silicon monopoly is not another merchant vendor. It is the set of companies large enough to design their own accelerators and the partners who can get those designs manufactured.

For buyers, custom silicon changes the negotiation rather than the price. A hyperscaler with a credible internal part has an alternative to cite, and the existence of the alternative does more work in the conversation than the part does in the cluster.

For everyone below hyperscaler scale, the effect is indirect and mostly unhelpful. Custom parts absorb the same constrained memory and packaging capacity that merchant accelerators need, which means the largest buyers are competing with the rest of the market for the identical bottleneck.

The forecast is a forecast, and a six-year projection given to Wall Street is a statement of intent as much as an estimate. The part that is already real is the supply Broadcom has bought, and that is the part competitors cannot answer quickly.