Broadcom’s Custom Chips Work Like Long-Term Licenses

Ninety-three percent of the analysts covering Broadcom rate the stock a buy. Its quant score, a model-driven read on business quality, sits at D, unchanged from a year ago even as the stock has climbed well off its 52-week low. Two ways of grading the same company, pointing in close to opposite directions, and I want to work out which one is closer to right before trusting either.

The quarter behind that split was Broadcom’s third fiscal quarter of 2026, reported September 2. Revenue came to $29.6 billion, up 86% from a year earlier and 33% from the prior quarter. AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially, according to CNBC’s coverage of the print. Numbers that size do not fit neatly into either a simple buy case or a simple quant score.

My thesis: Broadcom’s custom AI business behaves less like selling chips and more like collecting a royalty on the compute roadmaps of a handful of very large technology companies, and that royalty read explains the numbers better than a cyclical-chip read does. Whether it survives a single large customer changing its mind is the open question.

Record quarter, and a guide that keeps climbing

For the fourth fiscal quarter, management guided to AI semiconductor revenue of $21.7 billion, up 236%, and total revenue of $34.8 billion, up 93%, with non-GAAP operating margin holding near 66%. Broadcom also raised its full-year outlook to about $58 billion of AI revenue for fiscal 2026, and laid out, for the first time in this level of detail, targets of roughly $115 billion for fiscal 2027 and $230 billion for fiscal 2028.

I would treat those two out-year numbers as a roadmap the company believes it can hit, not as figures to underwrite a valuation today. Guidance for the next ninety days is one thing. A number for a fiscal year that has not started yet is a different kind of claim, and I want at least two more quarters of the current trajectory before I lean on it.

Why a custom chip can’t be swapped like a GPU

Here is the concept that matters most in this story. A general-purpose GPU can come from more than one vendor, and a buyer can shop around on price and availability. A custom accelerator, which Broadcom calls an XPU, is designed over roughly eighteen months to two years around one customer’s specific workload, memory layout and power envelope. Once that chip is running in a data center, switching vendors means redesigning the software stack around a new architecture, not just swapping a part.

That is why I keep coming back to the royalty comparison rather than a supplier comparison. A royalty is sticky because the thing it is attached to, in this case years of joint engineering, is expensive to replace. Nvidia’s own moat looks similar in spirit, built on software rather than raw chip performance, though the two companies are selling different products to overlapping customers.

There is a second layer to the stickiness that gets less attention. Broadcom licenses the underlying IP blocks, verification tools and packaging know-how that go into each generation, on top of designing the silicon itself. A customer that walked away would lose a multi-year design partner who already understands its specific workload, and rebuilding that relationship with a new vendor from scratch would cost more time than most hyperscalers can afford in the current capacity race.

That is the switching cost in one sentence. Years of joint engineering, not a purchase order, are what keep a customer in place.

Five customers carry the whole AI segment

Broadcom keeps AI revenue by customer out of its filings. Reporting on the custom-silicon business consistently names the same five companies anyway: Google, Meta, OpenAI, Anthropic and ByteDance. Google is the oldest relationship by a wide margin, on its seventh generation of co-designed chips going back roughly a decade, which is the closest thing to a track record this business has. OpenAI struck a multi-year agreement in October 2025 for custom accelerators to be deployed between the second half of 2026 and the end of 2029, adding a second large anchor customer to the book.

Management has said it expects to end fiscal 2027 with seven or eight qualified customers, roughly double today’s count. More names would matter less for near-term revenue, most of which still comes from the oldest relationships, than for reducing how much the stock hinges on any single one of them.

Margins are widening as the mix shifts

Gross margin came in at 67.8%, up from 63.0% a year earlier. Operating margin sits at 41%, with an EBIT margin of 40.6%, and net margin of 36%. Net income for fiscal 2025 was $23.1 billion, against $5.9 billion the year before, a jump that outran revenue growth of 24% by a wide margin.

That gap between earnings growth and revenue growth is the tell. When a growing share of revenue comes from custom silicon and infrastructure software rather than merchant chips, fixed engineering costs get spread across more volume, and the incremental dollar drops through at a much higher rate than the last one did.

Infrastructure software, the VMware-based half of the business, grew noticeably slower than the AI semiconductor line this quarter, and it is easy to overlook next to an 86% headline number. I would not overlook it. It is the steadier, subscription-priced half of Broadcom’s revenue, and it is what keeps operating margin from swinging as hard as the AI chip number does from quarter to quarter.

Networking is no longer the sideshow

For the fourth quarter, Broadcom expects AI networking revenue to roughly triple year over year alongside the XPU business, built on the newer Tomahawk switch chips and a growing bet on optical components. It is easy to treat networking as a footnote next to the headline AI chip number. I think that is a mistake. Every additional XPU shipped needs switching and optical capacity to actually move data between chips, so the networking business grows for the same reason the chip business does, and it carries software-like margins once the initial design work is done.

I wrote separately about why the broader semiconductor cycle isn’t dead so much as hidden behind AI capital spending, and Broadcom’s networking growth is one of the cleaner pieces of evidence for that argument. It is capacity spending that shows up outside the GPU headlines.

The concentration risk this creates

Trade press estimates, not company disclosure, put a single hyperscaler at more than half of AI segment revenue today. Broadcom will not confirm a number like that, and I would not treat it as precise. But the direction is not really in dispute: this is a business built on a small number of very large relationships, closer in shape to an aerospace supplier’s customer list than a typical chip vendor’s.

If that top customer decided to bring its next chip generation fully in-house, or shifted a meaningful share of its build toward a competitor such as AMD, which is chasing the same AI compute wave from the merchant-silicon side, the growth curve above would break quickly and the royalty framing I am using would need a rewrite. That is the specific way I would be wrong about this stock.

A ninety-three percent buy rating against a D quant score

Twenty-nine analysts cover Broadcom, 93% of them at buy, with an average target of $525 implying 47% of upside and a high estimate of $630. The stock trades at 45.6 times trailing earnings, actually below its own five-year average of 61.6, and at a forward multiple of 24.5 against expected earnings growth of 86%.

The quant model is not wrong to be skeptical on its own terms. It tends to weight customer concentration, revenue lumpiness and a market cap north of $1,707.1 billion as risk factors regardless of what is driving the growth. Analysts weight the backlog and the design wins instead. I lean toward the analyst side here, mainly because the margin expansion described above is already showing up in reported numbers rather than sitting in a forecast.

Metric	Value	Context
Price	$357.61	52-week range $290-$494, 27.6% below high
Market cap	$1,707.1 billion
P/E (TTM)	45.6	five-year average 61.6
Forward P/E	24.5	assumes EPS of $14.58, up 86%
Q3 FY2026 revenue	$29.6 billion	up 86% year over year, 33% sequential
AI semiconductor revenue	$16.7 billion	up 221% year over year, 54% sequential
Q4 FY2026 guide	$34.8B total / $21.7B AI	up 93% / 236% year over year, guided
Analyst target	$525 avg (29 analysts, 93% buy)	47% upside; range $400-$630

What I’d need to see by December

Broadcom’s fiscal fourth quarter closes in late October and should be reported in December. The number I am watching is whether AI semiconductor revenue lands close to the guided $21.7 billion, and whether operating margin holds near 66% as the mix keeps shifting toward custom silicon. The headline growth rate matters less than either of those. A print meaningfully below that revenue number, or a margin slip past a point or two, would tell me the quant model has been seeing something the buy ratings were not pricing. Until then, I think of this as a business earning a royalty on the largest capital spending cycle in technology, and I would value it accordingly rather than as a conventional chip supplier trading on a cycle.

Analysis and opinion only, not investment advice. Figures come from Broadcom’s filings on SEC EDGAR and its investor site; customer and guidance figures are sourced in the text above and were checked on September 23, 2026.