Broadcom closed its fiscal third quarter on August 2 with record revenue of $29.6bn, up 86% year over year. AI semiconductor revenue was $16.7bn, up 221% year over year and 54% sequentially. Management guided fourth-quarter AI semiconductor revenue to $21.7bn — up 236% against last year — and total revenue to roughly $34.8bn.
It was the ninth consecutive quarter Broadcom has beaten consensus. The stock fell, trading down roughly 3–4% by midday Thursday.
The reason is one line in the guidance. Fourth-quarter consolidated gross margin is expected to be about 73%, against 78% in the same quarter a year ago. Broadcom did not bury the explanation: the company attributed the compression to the growing mix of XPUs and, specifically, to the increasing memory content those XPUs carry.
The mix is the message
XPUs — Broadcom’s custom AI accelerators, built for a handful of hyperscale customers — shipped more than 3.5 times prior-year volumes and accounted for 73% of AI revenue in the quarter. That is the growth engine. It is also structurally the lowest-margin thing Broadcom sells.
The mechanism is memory. A modern accelerator is a logic die surrounded by high-bandwidth memory that Broadcom does not make and has to buy. When memory content per unit rises, the bill of materials rises with it, and a large slice of each incremental dollar of revenue is effectively pass-through to a memory supplier. The revenue line looks spectacular. The margin line records what actually stayed in the building.
This is the opposite of the software-and-networking mix that earned Broadcom a 78% gross margin in the first place. The company is trading percentage points of margin for enormous absolute dollars — which can be an excellent trade, and is a different business from the one the multiple was originally set against.
Our take: “Beat and fell” is not investors being irrational. It is investors correctly noticing that a 221% growth rate and a 500-basis-point margin decline are the same fact described twice. The interesting question is not whether Broadcom is winning custom silicon — it obviously is — but who captures the value inside an AI accelerator as memory content keeps climbing. Right now a growing share of it is being captured by companies whose names were not on this earnings release.
What to watch
- Whether 73% is a floor or a waypoint. If memory content per XPU keeps rising through fiscal 2027, the same sentence will produce a lower number next year. Management commentary on where the mix stabilises matters more than any single quarter’s beat.
- Absolute gross profit dollars, not the percentage. At $34.8bn of revenue, 73% is far more gross profit than 78% was on a much smaller base. The percentage is the headline; the dollars are the business.
- Memory pricing. HBM supply is concentrated among very few vendors. Any move in that pricing flows straight through this margin line, in either direction, without Broadcom having much say in it.
- Customer concentration disclosure. Custom accelerators are, by definition, built for named buyers. The revenue is spectacular and the counterparty list is short.
- The networking and optics attach. Ethernet and optics carry better economics than XPUs. How fast that side scales alongside the accelerator ramp is the swing factor on blended margin.
The general lesson
Every AI infrastructure supplier is now running the same experiment: revenue growth that would have been implausible three years ago, purchased with a mix shift toward hardware that carries more bought-in content. Growth rate and margin rate are moving in opposite directions across the whole supply chain, and the market has started reading the second number first.
When a company posts 86% revenue growth and the stock goes down, the market is not disputing the growth. It is asking what fraction of it the company gets to keep.
The Sharp Brief covers markets as news and analysis. Nothing here is investment advice or a recommendation to buy or sell any security.