AI Chips Tripled to $16.7B and the Stock Still Fell: Inside Broadcom's Massive Q3 and Its $230 Billion Bet

Broadcom just delivered one of the most paradoxical earnings reports of the AI boom. On September 2, the company reported fiscal Q3 2026 results in which AI semiconductor revenue more than tripled year over year to $16.7 billion, grew 54% just from the previous quarter, and pushed total revenue up 86% to $29.6 billion — comfortably beating Wall Street’s targets. Management then raised its fiscal 2027 AI chip revenue forecast to roughly $115 billion, up from the $100 billion figure it had been repeating since June, and sketched out a path to $230 billion in fiscal 2028. The stock fell anyway, sliding more than 6% at points after the company’s fiscal Q4 revenue guidance of $34.8 billion came in below analyst estimates.

That disconnect — record AI numbers on one side, a disappointed market on the other — is the real story of this quarter, and it says a great deal about where the AI infrastructure buildout now stands.

The numbers behind the blowout

Start with what Broadcom actually reported for its fiscal third quarter, which ended in early August:

• Total revenue: $29.6 billion, up 86% year over year, beating consensus estimates of around $29.4 billion.
• AI semiconductor revenue: $16.7 billion, up 221% year over year and 54% sequentially — meaning AI chips alone now account for well over half of the entire company.
• Adjusted EPS: $3.32, above Wall Street’s targets, keeping the company on track toward management’s claim of more than $30 in annual earnings per share.
• Operating margin: a record 67.9%, an extraordinary figure for a semiconductor business of this scale.
• Free cash flow: $13.7 billion for the quarter.
• Net income: more than tripled to $13.09 billion.

Against that backdrop, the fiscal Q4 guidance of $34.8 billion in revenue — which would itself represent continued sharp growth — landed below what analysts had modeled, and that single number drove the sell-off. Bloomberg’s take was blunt: the two-year AI chip sales forecast “failed to impress investors, a sign it’s still in the early stages” of what they had priced in.

The $115 billion question: who is buying all this?

The engine behind the tripling is Broadcom’s custom accelerator business — the XPUs it designs for hyperscalers and AI labs — together with the networking silicon (Tomahawk and Jericho switch chips, optical interconnects) that stitches large GPU clusters together. The customer list has become the envy of the industry: Google, Meta, and OpenAI are all locked in as custom-silicon partners through at least 2028, with Anthropic also signing on more recently. TIKR’s summary of the quarter captured it well: “Google, Anthropic and OpenAI just locked in Broadcom’s Q3 earnings through 2028.”

The OpenAI relationship deserves particular attention. Broadcom is the co-design partner behind OpenAI’s “Jalapeño” custom chip effort, a deal first detailed earlier this year and one that reportedly commits OpenAI to purchase tens of billions of dollars of custom silicon. Every incremental commitment like this converts directly into the backlog that lets CEO Hock Tan stand up on an earnings call and guide AI revenue from $58 billion this fiscal year to $115 billion next year and potentially $230 billion the year after.

In other words, Broadcom’s forecast is not a demand prediction in the abstract — it is largely contracted, multi-year capacity commitments from a handful of the largest AI spenders on Earth.

Why the market shrugged

If the backlog is real, why did the stock fall? A few threads stand out.

First, expectations had become untethered. Broadcom shares have run enormously on the AI custom-silicon thesis, and even a guide implying AI revenue roughly doubling again in FY27 was read as a disappointment against whisper numbers. When a stock is priced for perfection, “merely excellent” gets punished.

Second, the Q4 number fed doubts about the near-term slope. A $34.8 billion Q4 guide implies sequential growth that some read as a deceleration from the Q3 surge, and in a momentum-driven sector, the second derivative matters more than the level.

Third, concentration cuts both ways. With Google, Meta, OpenAI, and Anthropic representing the bulk of AI chip demand, Broadcom’s fortunes are now tied to the capital-expenditure discipline of four or five companies. Any sign that hyperscaler capex growth is flattening — and there have been scattered warnings all summer — hits Broadcom harder than a more diversified chipmaker.

Fourth, the VMware software business is no longer the story. Analysts noted that the AI revenue surge is masking softer dynamics elsewhere in the portfolio, including VMware, where the easy synergies from the acquisition have largely been harvested. The market increasingly values Broadcom as a pure AI infrastructure play, and it now grades the company on that curve alone.

The NVIDIA question

Broadcom’s rise is inseparable from the AI compute race with NVIDIA. The two companies are not yet head-to-head competitors in most racks — NVIDIA sells merchant GPUs and full-system platforms, while Broadcom sells custom silicon that hyperscalers design themselves precisely to reduce dependence on NVIDIA — but the strategic collision is coming. NVIDIA’s own recent moves, including its $12.93 billion acquisition of Hugging Face, signal that it intends to defend the ecosystem side of the stack, while Broadcom’s networking franchise means it actually earns revenue from both NVIDIA-based clusters and its own XPU deployments.

For the AI industry at large, the more important signal is what Broadcom’s backlog implies: the largest AI companies are now contractually committing to buy compute hardware years in advance at a scale that dwarfs most countries’ GDP. The $230 billion FY28 scenario would make Broadcom’s AI business alone larger than most semiconductor companies’ entire revenue, NVIDIA excepted.

What to watch next

The market’s skepticism and management’s confidence cannot both be right forever. The next checkpoints: whether fiscal Q4 actually lands at or above the $34.8 billion guide; whether OpenAI’s Jalapeño silicon moves from design wins into volume shipments on schedule; whether Google’s TPU generation cadence keeps absorbing Broadcom’s capacity; and whether any of the large customers wobble on capex. Broadcom has told a story of contracted, compounding demand. The next two quarters will show whether the market’s discount or Hock Tan’s forecast was the better model of reality.