Broadcom Inc. AVGO

Key data

The business

Broadcom is two companies under one ticker. About three-quarters of revenue is semiconductors: custom AI accelerators, called XPUs, that it co-designs with individual cloud giants, Ethernet switch chips (the Tomahawk and Jericho lines) that lash thousands of those accelerators into one cluster, and older cash franchises in Apple wireless, broadband, and storage. The remaining quarter is infrastructure software, mostly VMware, bought in late 2023 and forced from one-time licenses onto annual subscriptions, which raised prices and turned lumpy sales into recurring cash. The advantage is real in two places. In custom AI silicon and in high-radix Ethernet switching, Broadcom is one of only two credible suppliers, and a customer who has taped out a chip program cannot switch without losing years and re-spinning silicon. What the buyer actually holds is a rack of AI servers whose chips only talk to each other because Broadcom's switch fabric sits in the middle, and that middle is where the margin lives.

The numbers

The shape is a step-change, not a drift. Revenue rose 48% year over year in the April quarter to $22.2B, and operating income rose 85% to $10.8B, so the operating margin jumped from 39% to 49% as the high-value AI mix took over.

The inflection is the move from $16.0B in the summer of 2025 to $22.2B two quarters later, and it is accelerating. On the adjusted basis the company guides to, earnings ran $2.05, then $2.44, then $3.32 in the quarter reported September 2, each one ahead of consensus, the last beating $3.22. That July quarter did not restate remaining performance obligations or the maximum guarantee exposure, both of which surface only in the 10-Q, so the earlier watch on whether the roughly $160B backlog holds is still open, unresolved rather than broken. What the quarter did show was the ramp landing in the balance sheet: cash up to $24.0B, receivables doubled to $13.7B, inventory doubled to $4.5B, total assets $188.1B.

Step back and the compounding is plain. Revenue went from $27.5B in FY21 to $63.9B in FY25 and is annualizing near $90B now, operating income has tripled, and gross margin has held near 68% through the VMware integration. Because Broadcom is fabless its capital spending is trivial, about $0.6B a year, so nearly all of it converts to cash, $27.5B of operating cash flow in FY25. The VMware deal left $65B of long-term debt, but net debt is only about 0.7 times cash earnings, and buybacks resumed at $7.85B in the January quarter after a year of paydown. What this memo believes that the tape does not fully price is a second wave of custom-XPU customers beyond today's handful. The print that settles it is the AI revenue guide and any new-program disclosure next quarter.

Management

Read the record, not the bio. Insiders sold $552M across 77 disposals in the last year against a single $0.4M purchase. The largest is co-founder Henry Samueli, about $104M in three same-day March sales, plan status not disclosed in the filings on hand, so I will not call it either routine or a signal. Hock Tan's team is a disciplined operator: it bought VMware, pushed prices hard, paid down debt for a year, then restarted buybacks at $7.85B in Q1. Non-GAAP guidance has been beaten four quarters running.

How it fails or surprises you

Customer concentration. A small circle of hyperscalers drives the AI revenue, and Broadcom already co-designs Google's TPU. Any one of them deferring a program, or pulling more design in-house, removes a large block of orders at once and shows up first in that doubled receivables line reversing. This is the single largest swing factor, and at 62 times earnings it is not diversifiable.

A second wave of XPU customers (right tail). Broadcom has signaled accelerator programs beyond its current buyers, reportedly including new model labs. If a fourth or fifth program reaches volume, AI revenue re-rates well past the current guide, and consensus is not paying for it because none of it has shipped. Watch the next AI revenue guide and any named new-customer commentary.

No margin for a spending pause. At about 62 times earnings and a free-cash yield near 2.3%, the price assumes the AI capital cycle does not breathe. If the hyperscalers digest a quarter, this stock has the least cushion of any large-cap chipmaker. The tell is their own capital-spending guides, not Broadcom's.

Closing thoughts

The evidence points to a distribution the market mostly already prices. At 62 times earnings, the AI ramp visible in the receivables line is paid for, so the edge is not in the central read, it is in the tails. The fatter tail is up: a genuinely new XPU customer at volume is worth more than the concentration risk is likely to cost in any single quarter, because once a program is live the switching cost runs in Broadcom's favor. The left tail is real but slow. A hyperscaler pulling silicon fully in-house plays out over years, not a quarter, which is time enough to see it coming in the orders.

The bet is still that the giant cloud companies keep buying Broadcom's custom AI chips and the switches that wire those chips together. It breaks if the customer count stops growing while the existing buyers slow, and the first place you would see it is AI revenue guidance softening as receivables roll off faster than they build. Those two numbers tell you before the income statement does.