Broadcom’s Q3 FY2026 Results: Custom AI Silicon Is Now a Franchise, Not a Hedge

At $16.7 billion in a single quarter — 56% of total revenue — AI semiconductors have transformed Broadcom’s identity; the open question now is who captures the margin pool in custom silicon.

What Happened

Broadcom’s own Q3 FY2026 press release, published September 2, 2026, reported total revenue of $29.6 billion, up 86% year over year, with AI semiconductor revenue of $16.7 billion — up 221% year over year and 54% sequentially. That sequential figure is the one that demands attention: a $5.9 billion quarter-over-quarter addition is not the smooth ramp of mature programs. It implies at least one newer custom accelerator moving into volume production. The identity of that program is a call-only detail; this analysis will not speculate on it.

Non-GAAP earnings came in at $3.32 per share, up 96%, on $13.7 billion of free cash flow. Hock Tan’s statement in the release was direct: “Demand for our custom AI accelerators and networking continues to be very strong.” The Q4 guide extends rather than moderates the trajectory — roughly $34.8 billion in total revenue with AI semiconductor revenue of $21.7 billion, up approximately 236% year over year, which would take AI to roughly 62% of the company.

Two standing hedges before the analysis. All figures cited here are from Broadcom’s press release; the earnings call was still in progress at time of writing, so no call-only items appear — no newly confirmed XPU customer names, no updated backlog or bookings figures, no change to any 2027 AI revenue target. Prior-quarter figures that continue to circulate — including bookings and backlog numbers from earlier periods — are not current and are not printed here. The often-cited FY2026 AI total of roughly $57.6 billion is a derived figure, the arithmetic sum of four quarterly prints, not a line Broadcom reported. And nothing here is investment advice.

The Structural Read

The merchant-GPU-versus-custom-silicon debate has resolved its first question and opened a more interesting second one. For two years the live argument was whether hyperscaler-designed accelerators — the XPUs Broadcom builds with its cloud customers — would take meaningful share from off-the-shelf GPUs. At $16.7 billion in a single quarter, up 221%, guided to $21.7 billion next, that question is closed. The live question is different: how much of the AI chip margin pool does the custom-silicon designer actually keep?

The “commodity ASIC” framing argued that when a hyperscaler designs its own chip, it captures the economics and the silicon vendor becomes a low-margin, foundry-adjacent contractor. Broadcom’s numbers argue the structural opposite. Non-GAAP operating margin ran near 68% in Q3 and the Q4 guide holds at 66% — roughly two points of compression for a roughly six-point climb in AI revenue mix. That is a very small margin cost for a very large compositional shift.

Two explanations compete, and which one is correct determines what multiple the AI revenue deserves. First: XPU economics are structurally better than the commodity-ASIC framing assumed — Broadcom’s design IP, networking stack, and packaging expertise create genuine pricing power even when the customer owns the architecture. Second: VMware’s software margin is absorbing the AI mix, flattering the blended number. The release alone cannot adjudicate between them. But the margin print at this scale and this mix shift is the fact that forces the question onto the table — and that question, framed through the Map of AI, is really about where in the AI stack durable margin lives: with the merchant GPU vendor, the custom silicon designer, or the hyperscaler that owns both the model and the chip spec.

Three Implications