Broadcom Inc.

Broadcom does two things that have nothing to do with each other, run by one man who treats them the same way. The first is infrastructure software — VMware, the mainframe tools of CA Technologies, Symantec's enterprise security — bought cheaply from owners who had given up, stripped to the customers who cannot leave, and run for cash at an 84% operating margin. The second is semiconductors: the radio filters in every iPhone, the switch chips inside most of the world's data-centre networks, and — the reason the company is worth $1.7 trillion — custom AI accelerators, which Broadcom designs and manufactures for six customers who want a chip built for their own models rather than Nvidia's chip built for everyone.

The numbers are the kind you read twice. In the quarter to 2 August, revenue rose 86% to $29.6 billion. AI semiconductor revenue rose 221% to $16.7 billion — more than the entire company earned in a quarter two years ago. Operating margin, on the company's preferred measure, was 68%. Free cash flow was $13.7 billion in ninety days. Management guided AI revenue to roughly $58 billion this fiscal year, $115 billion next, and $230 billion the year after, and said it has secured the supply to deliver it (approx. — Q3 call). Hock Tan, who has run this company for twenty years and never once guided to a number he did not hit, called the demand roadmap "very much on target to exceed $30 in EPS in fiscal 2028."

And the stock is 27% below its June high. That is the puzzle this report exists to solve, and the honest description of a share of AVGO today is the answer: a wonderful, cash-gushing toll bridge, bolted to a growth business with six customers, priced at eleven times a year that has not happened yet. We scored Nvidia "Too Hard" in June: a magnificent business, an unknowable price. Broadcom is not too hard — it is legible in every line — but the legibility reveals something Nvidia's does not: the growth has a customer list, and the list is short.

Two lessons for an owner. First, the acquirer's record is the rarest kind: Buffett warned in 1981 about managerial "princesses" whose kisses were supposed to turn toads into princes; Tan's trick is that he buys toads and never tries to turn them into anything — he runs them for cash, and CA, Symantec and VMware all delivered. Second, and new: the AI business is the first thing Tan has ever built rather than bought, and it behaves differently from everything in the playbook. It is co-developed with the customer, re-bid every chip generation, carries a lower gross margin, and — since June — is partly financed by Broadcom itself. The company that spent twenty years harvesting cash from other people's software is now underwriting other people's capital expenditure. That is the chapter this report is about.

FY2025 revenue by reported segment (year to 2 November 2025), and what has happened to the mix since:

The bridge. Legacy Broadcom is what Buffett meant by a toll bridge: RF filters in every iPhone (Apple was ~20% of revenue before AI swamped it), the switch silicon under most data-centre Ethernet, and software that enterprises would need years and fortunes to leave. Capital expenditure is 1.4% of revenue; research is ~13%; the rest is cash. VMware alone throws off roughly $7 billion of operating income a quarter (approx.) from customers who — in Tan's own careful phrase last year — "bought licences to deploy it, it doesn't mean they are fully deployed."

The promise. The AI business is different in kind. Broadcom does not sell a GPU; it builds the customer's own chip — Google's TPU (co-developed since 2016), Meta's MTIA, OpenAI's Jalapeño — supplying the interconnect, the packaging, the memory integration and, increasingly, the whole rack. It then sells the network around it. In Q3 the custom chips were ~73% of AI revenue and shipped at 3.5× last year's volume (approx. — call). The economics are superb and getting slightly less so: because memory and packaging pass through at cost, gross margin has slid from 78.4% to 75% and is guided to ~73%, while operating margin holds near 68% because opex is only ~10% of revenue. Management's line is "focus on operating margin." Ours is that a gross margin falling five points in a year, while operating margin rises, is the signature of a business turning from a designer into a manufacturer.

The new line item. In June, Broadcom launched a $35 billion financing platform with Apollo and Blackstone to fund its customers' racks — the first tranche for Anthropic's first gigawatt. The 10-Q discloses a maximum exposure of roughly $29 billion on that first transaction, a five-year lease-payment backstop (approx. — 10-Q). The CFO said there is "nothing new to announce regarding residual value guarantees or backstops" and that future tranches are "deal-by-deal." A chip company guaranteeing the leases on its own chips is not unheard of; it is how the telecom-equipment boom ended in 2001. We are not saying that is where this goes. We are saying it is the line to read first in every filing from now on.

Buffett's test for a moat is whether a competitor with money, talent and will can replicate the advantage. For the bridge — the filters, the switch silicon, the mainframe code — the answer has been no for a decade, and it is still no. For the promise, the honest answer is "partly, and the customer is trying". Google designed the TPU; Broadcom implements it; and every generation Google asks whether someone cheaper can implement it too. In 2026 the answer arrived: MediaTek for inference, Marvell for the ecosystem around it, AMD reportedly for a future part. That is not the moat collapsing — Broadcom keeps the training chip, the networking, and five other customers — but it is the moat being priced by the customer, which is what happens to every supplier whose customer owns the architecture. Nvidia's customers cannot do this to Nvidia, because Nvidia owns CUDA. Broadcom's can, because Broadcom sells implementation. Score: 8 — the same as Nvidia's and Alphabet's, wide and contested, and for the first time contested by the customer rather than the competitor.

Every business on our board carries one question the price turns on. Broadcom's is not whether AI demand is real — Nvidia's $96 billion quarter settles that — nor whether custom chips work — Google has run on them for a decade. It is narrower and harder: can a $230 billion revenue line rest on six customers, two of whom are private companies that lose money, one of whom is financed partly by Broadcom itself, and the largest of whom is hiring second sources?

The durability case. The contracts are long: Google through 2031, Meta through 2029, OpenAI through 2029, Anthropic to ten gigawatts by 2028. Bookings ran three times shipments in the spring (approx. — press estimates). Hyperscaler capital budgets for 2026 are near $725 billion (approx.), and a co-designed chip runs a customer's own model at, in Tan's words, "less than half the cost" of a GPU. The FY2027 guide of ~$115 billion exceeds the top of the $60–90 billion "serviceable market" Tan set in December 2024 — he has raised, not cut, at every print. And his own record on guidance, over twenty years, is unblemished.

The concentration case. "Depth is not breadth." Six boardrooms determine most of the growth; the top five customers were ~40% of revenue before the AI ramp and are more now. The largest customer by 2027–28 will be Anthropic — not a cash-rich hyperscaler but a private lab with a large and unprofitable revenue base, whose racks are leased through a vehicle Broadcom backstops. Meanwhile the cash-rich hyperscaler that built the business is diversifying: Google's Marvell warrant vests against up to ~$120 billion of purchases (approx. — 8-K). Tan's own words on the call: even as the chips ship, "are they going to be deployed on a timely basis?" — the outlook is "highly sensitised" to land, power and shell. Add a gross margin that has given up five points, a Q4 guide $200 million light, and a ten-year Treasury near 4.8% discounting promises for 2028.

Our read: the ramp is real and the concentration is the price of it. A business that sells implementation to customers who own the design will always be re-bid; that is not a flaw in Broadcom, it is the definition of the product. What changed in 2026 is that the re-bidding became visible (Google) and the customer list became riskier (Anthropic first, financed in part by the supplier). Neither breaks the thesis; both raise the discount rate. The growth dial is a 10 — nothing on our board is growing faster — and the valuation dial is a 4, because the growth is priced and the customers are six.

The pattern is unusual and worth stating plainly: Broadcom's most dangerous competitors are its customers. Nvidia competes with it; Marvell and MediaTek compete with it; but the entity that decides whether Broadcom implements the next TPU is Google, and Google has every incentive to keep two suppliers honest. The mirror image is the software: there, the customers cannot leave in under three years, and the company prices accordingly — which is why the European cloud providers have taken it to the Commission. Broadcom is, in the same quarter, the price-taker in silicon and the price-maker in software. The stock trades on the first; the cash comes from both.

Integrity: no veto. Tan's guidance record is the cleanest of any large-cap CEO we have scored, and the VMware price increases, however unpopular, were announced, not hidden. Capital allocation is textbook and slightly complicated. Textbook: FY2025 free cash flow of $26.9B went ~$11.7B to dividends (approx.), $6.3B to buybacks and $2.8B to net debt reduction; gross debt is down from ~$76B at the VMware peak to ~$58B after the post-quarter paydown, at a ~4% coupon; the dividend policy — half of the prior year's free cash flow — has produced fifteen consecutive raises. Complicated, on three counts. First, the share count rose 27% in a decade (3.83B → 4.85B) because VMware was paid partly in stock and because stock compensation runs at $7.6B a year, 9.5% of revenue — buybacks here offset dilution; they do not shrink the company. Second, the acquisition machine has stopped: no deal since VMware, and the next one would need to be $100B+ to matter, which is Buffett's "curse of scale" arriving at Broadcom. Third, the financing vehicle: capital that used to buy toads now backstops customers' leases. Dial: 7. A superb operator, an expensive one, at 74, with a $29B footnote.

Four feed faults, named. First, the DCF: our model prints an intrinsic value of $184, −49%. Yesterday, at Lululemon, the same model said +216% because it capitalised earnings the company was leaving; here it capitalises trailing free cash flow while revenue is doubling — the identical fault, mirrored. Neither number has read the guidance. Second, the P/E of 45 is computed on GAAP earnings that carry ~$8.8B a year of amortisation of acquired intangibles; on the non-GAAP run-rate the multiple is nearer 28×. We show both and note, in the other direction, that non-GAAP also excludes $7.6B of stock compensation, which is real pay. Third, the dividend feed prints 2023's payout as $19.05 a share — it is $1.905; the pre-split history was never adjusted, so the 10-year dividend growth the feed implies is wrong by a factor of ten. Fourth, the estimates feed's FY2030 row shows revenue of $139B below FY2029's $336B, from ten contributors who are plainly not modelling the same company; we stop the forward table at FY2028.

The dividend is the one place where the concentration risk does not reach. It is paid out of last year's cash, which is already in the bank; at 31% of free cash flow it is covered three times; and the policy — half of the prior year's free cash flow — points to a December raise that, applied mechanically, would roughly double the payout (approx.). Tan's language ("increase dividend and probably do stock buyback") suggests the board will not be mechanical, and a 0.7% yield will not attract income investors either way. But it answers the sixth question in the rare direction: this is Chubb and Altria — a growth rate turning up — not Rexford's decay. Write the exception when the data says so.

Here is the way to think about the price. The bridge — software plus non-AI chips — earns roughly $35B of revenue a year at margins that put its operating income near $30B (approx.); a slow-growth, unassailable, 84%-margin software franchise deserves 18–20× its ~$24B of net profit, call it $450–480B, or ~$95–100 a share. The promise — everything AI — is therefore being bought for about $265 a share, ~$1.26 trillion, against ~$58B of revenue this year and a guide of $115B next. If FY2028 arrives as guided, with ~$30 of EPS, the whole company at 15× is worth ~$450 and the stock is 20% cheap. If the AI line stalls at FY2027's level — $115B, no doubling to $230B — the earnings power is nearer $19–20 a share, and at 16–18× the stock is worth $300–360: today's price, with no cushion. If Google's diversification and a private-lab funding winter cut FY2027 to, say, $80B, you are looking at ~$15 of earnings and a stock in the low $200s.

So the arithmetic is honest and uncomfortable: at $362 you are paid for 2027 and asked to trust 2028, from six customers. Buffett's margin-of-safety tiers ask 20–30% for a business of very high certainty, 40–50% where uncertainty is real. We would call the bridge very high certainty and the promise real uncertainty; blended, we want ~35% off a ~$430 blended fair value, which is ~$280 — a hair below the March low, and ~15× FY2027 consensus. Above that, the stock can absolutely go up — the street's $510 says it will — but it goes up on the promise, and we do not pay for promises with six signatures.