· 9 min read

Broadcom's AI Chip Business Grew 221% in a Quarter. The Stock Fell Anyway. Here's What That Actually Tells a Solo AI Builder.

AI Stack

On September 2, 2026, Broadcom reported fiscal third quarter revenue of $29.6 billion, up 86 percent year over year, with AI semiconductor revenue up 221 percent to $16.7 billion. That's the biggest AI quarter the company has ever had. The stock fell anyway, down roughly 4 to 6 percent over the following session, because Broadcom's Q4 revenue guide came in a hair under what Wall Street wanted. A company posting record numbers and a market shrugging is a more useful signal for anyone building on top of API pricing than the next "the AI bubble is popping" thread you'll see this week.

What Broadcom actually reported

The headline numbers, straight from the company's own filing: net revenue of $29.591 billion for the quarter ended August 2, 2026, up 86 percent from a year earlier. Non-GAAP operating income hit $20.1 billion, up 92 percent. Non-GAAP diluted EPS came in at $3.32, ahead of the roughly $3.23 analysts were modeling, up 96 percent year over year. Free cash flow was $13.7 billion, 46 percent of revenue, also a record.

Inside that, the number that matters for this blog: AI semiconductor revenue was $16.7 billion, up 221 percent year over year and 54 percent from the prior quarter. CEO Hock Tan said it plainly on the earnings call: "Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter." Semiconductor solutions overall, the segment that includes the AI chips, made up 70 percent of total revenue this quarter. Two years ago that would have been an infrastructure-software company with a chip side business. Now it's the reverse.

The two Q4 numbers that get mixed up

Here's where a lot of the coverage this week got sloppy, and it's worth untangling because the two figures aren't measuring the same thing. You'll see "$21.7 billion" and "$34.8 billion" both attached to Broadcom's Q4 guidance, and they read like a contradiction until you check which line of business each one describes.

Straight from the press release: Hock Tan guided AI semiconductor revenue specifically to $21.7 billion for Q4, up 236 percent year over year. CFO Amie Thuener separately guided total consolidated company revenue, everything Broadcom sells including its infrastructure software business, to approximately $34.8 billion, up 93 percent year over year. One is a business line. The other is the whole company. They're both real, they're both from the same press release, and they're not in tension with each other.

The reason the stock moved is the second number, the $34.8 billion total guide, landed just below the roughly $35.0 billion analysts had penciled in. A miss of well under one percent against consensus, on a quarter where the company beat on every headline metric, is what took the stock down.

Why a record quarter still sold off

This is the part that's genuinely useful if you've never watched a chip earnings cycle up close: at this valuation, "beat, but not by enough" gets priced like a miss. Broadcom didn't guide down. It guided up 93 percent and the market still took points off, because the stock had run up on the assumption that AI-linked names would keep clearing the bar by a wide margin every quarter. There's also a real margin story underneath the sentiment one: Broadcom guided Q4 consolidated gross margin to roughly 73 percent, down from 78 percent a year ago, because custom silicon carries thinner margins than the software business it's increasingly diluting. Investors who read gross margin compression as an early warning sign aren't wrong to watch it, even if the operating margin (guided flat at 66 percent) says the mix shift is being managed, not stumbled into.

The custom silicon bet, and why it's the actual story

Broadcom doesn't sell GPUs. It designs custom ASICs, XPUs, for specific hyperscaler customers who want AI compute that's cheaper at scale than general-purpose Nvidia silicon, in exchange for giving up some flexibility. What's new this quarter is that Broadcom named names on the earnings call in a way it historically hasn't: Google's TPU v7 (Ironwood) is now shipping in high volume to both Google and Anthropic, TPU v8i has started production shipments for Google, and OpenAI's first custom accelerator (reportedly codenamed Jalapeno) has begun shipping. Meta's custom inference chip is also in the pipeline. Broadcom says it now has six XPU customers total, and it's guiding full fiscal 2026 AI revenue to roughly $58 billion, with multi-year outlooks of about $115 billion in fiscal 2027 and $230 billion in fiscal 2028.

Take those multi-year numbers for what they are: a public company's own forward guidance, delivered on the same call where it's trying to reassure investors after a stock drop. Not an audited fact. But even discounted heavily, four of the largest AI labs and platforms on earth are locking in years of custom silicon capacity because it's cheaper per unit of compute than buying more Nvidia GPUs. That's a demand signal that doesn't depend on whether AVGO stock is up or down this week.

The honest take

None of this changes your OpenAI or Anthropic API bill today, and it won't next quarter either. Broadcom's chips ship to hyperscalers, get built into data centers, and show up as lower marginal compute cost eighteen to thirty-six months later, filtered through whatever pricing strategy the labs choose. But sustained, multi-year hyperscaler capex on cheaper-per-unit custom silicon is a real leading indicator that inference costs keep trending down over the next year or two. It's a much stronger data point than a single stock chart, and it argues against the recurring "the bubble's about to pop, prices are about to spike" narrative that shows up in solo-operator circles every time a chip stock has a red day.

What I'd actually do with this: build for the pricing you have today, not pricing you're hoping shows up. Concretely, that means don't defer real cost engineering, prompt caching, batching, picking a smaller model where it's good enough, because you're betting tokens get 10x cheaper next quarter. Do that work now; if prices drop later, you just save more. On the flip side, don't over-architect your stack around brittle assumptions about any single price staying exactly where it is either. Keep your model layer swappable, don't hardcode pricing assumptions into your unit economics spreadsheet, and revisit them quarterly rather than never.

Where I could be wrong: if hyperscaler capex actually hits a digestion phase, meaning Google, Meta, and the AI labs pause spending to prove out returns on what they've already built, Broadcom's own multi-year figures could get walked back hard, and the "prices keep falling" thesis weakens with it. Custom silicon yield problems or a slower-than-expected ramp from any of the four named customers would also blunt this. I'm reading a strong signal here, not a guarantee.

Author

Lukas

Sources

More AI articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts