Broadcom Q3 2026 Earnings: AI Chips Triple, Stock Dips

Broadcom closed out the season’s marquee chip earnings after the bell on Tuesday, September 2, 2026, reporting record fiscal third-quarter results as its custom AI-accelerator business more than tripled year over year. The company posted consolidated revenue of $29.6 billion, up 86% from a year earlier, and non-GAAP earnings per share of $3.32, up 96%. Both figures cleared Wall Street consensus, which had centered on roughly $29.4 billion in revenue heading into the print.

And yet the stock fell in after-hours trading. The reason was not the quarter Broadcom just delivered but the one it guided to next: management projected fourth-quarter revenue of about $34.8 billion, just shy of the roughly $35.03 billion analysts had modeled. In a name priced for acceleration, an in-line-to-light guide was enough to trigger a pullback.

The AI number that mattered

The single line investors came for was AI semiconductor revenue, and it did not disappoint on the quarter. AI semi sales reached $16.7 billion, up 221% year over year and 54% sequentially, and now account for 56% of total revenue — a majority-AI mix for the first time. That comfortably beat the roughly $16 billion management had guided to in the earnings preview a day earlier.

Custom accelerators — the application-specific XPUs Broadcom designs for hyperscalers — did the heavy lifting. XPU shipments were up more than 3.5x year over year and represented 73% of AI revenue in the quarter, with the balance coming from AI networking silicon. Chief Executive Hock Tan said the company’s six XPU customers are accelerating adoption of custom accelerators and that the addressable market for AI compute “continues to expand quickly.”

Broadcom also detailed which chips shipped to whom. It delivered Ironwood TPU version 7 in high volume to both Anthropic and Google, began production shipments of the next-generation TPU version 8i for Google, and shipped Jalapeno, OpenAI’s first-generation custom accelerator, which the company said outperforms the Grace Blackwell GPU on inference workloads. Broadcom’s networking and infrastructure-software franchise rounded out the mix: infrastructure software revenue rose 29% to $8.8 billion, and total semiconductor sales grew to $20.8 billion.

Guidance did the talking

The forward outlook is where the story turned. Management guided fourth-quarter AI chip revenue to about $21.7 billion, another sequential step up, and pointed toward full fiscal-2026 AI semiconductor revenue of roughly $58 billion — modestly above the $56 billion target Broadcom had been carrying into the report. Tan told investors the AI franchise is on a path to roughly double into fiscal 2027, consistent with the company’s earlier guidance toward AI revenue eventually clearing $100 billion annually.

But the total-company $34.8 billion Q4 guide landed just below consensus, and in a stock that has been one of 2026’s biggest winners, “beat-and-in-line” reads as a disappointment. The setup was exactly the one laid out before the print: a beat-and-raise was close to priced in, so the reaction would hinge on whether the forward numbers extended the runway or merely confirmed it. They confirmed it — and the stock traded lower.

Tan added a data point that reframes the customer concentration question. He said Anthropic is “on track” to become Broadcom’s largest XPU customer in 2027, and to remain so into 2028 — a signal that the AI-lab buildout, not just the hyperscalers, is now a first-order driver of Broadcom’s order book. The company has been tied to a reported $60 billion debt arrangement involving Anthropic, and the two companies’ fortunes are increasingly linked through the same custom-silicon roadmap.

The custom-silicon franchise widens

For years, “AI chip” meant “Nvidia GPU.” Broadcom’s ascent has reframed that debate. The biggest buyers increasingly want their own silicon — designed to their workloads, owned outright, and freed from GPU allocation queues — and Broadcom is the merchant partner most of them turn to.

The quarter offered fresh evidence that the customer base is broadening beyond the two or three anchor accounts that built the franchise. The OpenAI-tied Jalapeno inference chip reaching production shipments is the clearest example: a third major AI lab now has custom silicon in Broadcom’s pipeline alongside Google and Anthropic. Each additional anchor account materially changes the multi-year revenue math, because custom-silicon programs run for years once they reach volume.

That breadth is also what separates Broadcom’s story from a pure momentum trade. The company sits at the intersection of custom compute and AI networking — the switching and interconnect silicon that stitches together the sprawling GPU and XPU clusters powering the hyperscaler capex boom. As clusters scale toward the multi-gigawatt facilities now under construction, the networking attach rate rises with them, giving Broadcom a second AI growth vector that does not depend on winning every accelerator socket.

The financing shadow

There is a caveat the AI-hardware trade cannot fully escape: how the buildout is being paid for. The hyperscalers funding these orders are running historic capital-spending deficits, and the AI labs behind much of the incremental demand are financing compute with large amounts of debt. Broadcom’s order book is only as durable as its customers’ ability to keep spending — and much of that spending is credit-financed against future revenue that has yet to materialize.

That links Broadcom’s multi-year targets to the same credit conditions rattling the rest of the market, and it is why the economics of AI data centers have become as scrutinized as the chips inside them. Broadcom reported into a backdrop of elevated bond yields and periodic chip-sector selloffs, and its guidance was always going to be read as a referendum on whether the debt-financed capex cycle can keep expanding at the pace the market has priced.

What it means

Broadcom delivered a genuinely strong quarter and got sold anyway — a telling sign of how much good news is already embedded in AI-hardware valuations. The results validate the core thesis: custom silicon is a durable, majority-of-revenue business for Broadcom now, not an emerging line item, and the customer base is widening rather than narrowing.

Who wins. The custom-accelerator model itself. With AI chips at 56% of revenue and three named lab-and-hyperscaler anchors in volume, Broadcom has cemented its place as the second structural winner of the AI compute era behind Nvidia. Anthropic’s move toward becoming the top XPU customer confirms that AI labs — not only cloud giants — are now buying silicon at hyperscaler scale, a dynamic that also underpins deals like Anthropic’s custom chip work with Samsung.

Who’s exposed. Anyone underwriting the stock for perfection. A $34.8 billion guide a few hundred million below consensus was enough to knock the shares, which tells you the bar is now acceleration-on-acceleration. Concentration risk is real too: a franchise anchored to a handful of buyers is exposed to any single customer trimming its spend — and those customers are financing much of their compute with debt.

What to watch next. Three things into the fourth quarter and fiscal 2027: whether the roughly $58 billion full-year AI figure and the “double into 2027” framing hold or get revised; gross-margin trajectory as lower-margin custom accelerators grow into a larger share of the mix; and any disclosure that moves the customer count beyond six. In a market that treats a record quarter and an in-line guide as a selling opportunity, the gap between “great” and “great, but priced in” is the entire trade.