Reference price: $359.10 (September 3, 2026 premarket, −2.2%, after fiscal Q3 results reported the night before). Numbers come from the earnings release and eight years of official filings; multiples are recomputed at the live price, not yesterday's close.
What this company actually does
It designs custom chips for whoever commissions them, and it sells management software to large enterprises. Two very different businesses under one roof: chips ($20.8B last quarter — of which $16.7B is custom AI silicon for the four or five giants that want their own chip instead of buying the standard one) and software ($8.8B, mostly the layer that runs big-company data centers).
The thing to understand: Broadcom doesn't sell a catalog chip. An enormous customer says "I want a chip that does exactly this," and Broadcom designs and builds it. That locks the customer in for years… and it also means the customers can be counted on one hand.
The good — and last night's quarter is one of the best we've seen
Revenue nearly doubled in a year. The summer quarter went from $16.0B to $29.6B (+86%), and operating profit from $5.9B to $16.0B (+171%). Out of every $100 of revenue the business used to keep $36.90; it now keeps $53.90 — most companies keep $10 or $15. And the engine has a name: custom AI chips went from $5.2B to $16.7B (+221%), with $21.7B guided for next quarter (+236%). They grew 54% versus the previous quarter — not year over year: versus three months ago.
The real money followed the profit, dollar for dollar. This is the check most companies fail, and here it comes out spotless: reported profit +95%, free cash that actually came in +95%. Exactly the same percentage. No bloated inventory, no revenue booked in advance, no customers who don't pay. $46 of every $100 of revenue converts into free cash.
Guidance with near-surgical aim: in eight years it has never missed its own forecast by more than 1.14% in either direction. It is paying down the debt from its big acquisition ($65.1B → $59.4B in nine months, with cash rising from $16.2B to $24.0B): interest takes one-twentieth of what the business earns. Stock compensation has been cut in half (from $29.70 per $100 of profit in 2025 to $17.80 in 2026). And the gap between adjusted and GAAP earnings is closing on its own: from 43% to 25%.
The bad
Problem 1 — and it is the finding of this analysis: the rising margin is an accounting mirage. Manufacturing costs MORE than before. Published gross margin rose from 67.1% to 69.1%. But inside that cost line live two very different things: the real cost of manufacturing rose 105.8% ($3.7B → $7.6B) while revenue rose 86%, and the accounting amortization of the big acquisition stayed flat ($1.52B → $1.50B). Translated: out of every $100 collected, manufacturing used to cost $23.20 and now costs $25.80. The product leaves LESS behind than before; the published margin rises only because a fixed charge is spread across twice the revenue. Why it matters: custom AI chips carry lower margins than the traditional business, and as they go from a third to 56% of everything sold, the real product margin dilutes — and that is invisible in the headline.
Problem 2: it spends LESS on R&D than a year ago, in dollars. Research: $3.05B → $2.90B (−5.1%) while revenue nearly doubled — from $19.10 to $9.80 per $100 of revenue. The fair part, because there is one and it's strong: when you design a custom chip for a giant, the customer pays for much of the engineering, and the software business barely needs new research. This can be a smart business model, not a cut. But it is a fact to keep in front of you: it is winning the AI race while spending less on inventing than a year ago. If the customer ever decides to build the chip itself, what's the defense?
Problem 3: strip out the acquisition and book equity is negative. Reported equity: $99.7B; remove the acquired company ($97.8B) and its intangibles ($26.3B) and what's left of real things is −$24.4B. It doesn't affect its ability to pay bills — debt is falling and interest is covered twenty times — but any "price-to-book" measure here tells you nothing.
Problem 4: four or five customers are 56% of revenue. Custom AI chips, by definition, aren't sold to thousands of companies. If one customer switches supplier or cuts its plan, revenue doesn't dip — a large chunk disappears.
Problem 5: your slice gets slightly smaller every year (share count 4,841M → 4,884M, +0.9%): it pays a large dividend and barely buys back stock, so the paper it hands employees is never retired. Problem 6: the years that make it look affordable (2028 at 13x, 2029 at 11x) are signed by 28 and 2 analysts respectively; the two that matter — 2026 at 31x and 2027 at 18.4x — are signed by 49 each, one of the best coverages we've seen. Problem 7: the business yields you 2.25% a year (3.11% at last quarter's pace; 3.0–3.8% with the dividend) against a 5–6% cost of money.
The math that actually works
Here the numbers add up, and we want to show the calculation because it's the best news in the analysis. To earn the $19.49 expected for 2027, on $168B of revenue it would need to keep $66 of every $100 — exactly what the company itself just guided. What 2027 expects fits if Broadcom keeps growing 21% above the pace it will close this quarter at; with custom AI chips growing 54% quarter over quarter, that 21% a year looks conservative today. It isn't being asked to improve the margin — it's being asked to keep selling.
So the right question isn't whether the numbers work. It's this: Is it really growing? Yes — +86%, guiding +93%. Does cash follow profit? Exactly — both +95%. Does the balance sheet hold? Yes. Is what's asked achievable? Yes, without margin improvement. Does every dollar sold leave the same behind? No — manufacturing costs $2.60 more per $100 than a year ago. Is it priced in? In good part.
Only one of two things can happen: either custom chips keep growing like this, and the diluting margin doesn't matter because volume covers everything; or the pace normalizes, and you're left with a product that leaves less behind, frozen research, and four customers. Not both. And today's price is paying for the first.
What the chart says
For six years (2018–2023) price and earnings moved almost hand in hand: Broadcom was a normal company at a normal price. In 2024–2025 the price detached violently — at the end of 2025 it traded at $369.64 while earnings justified $102.30: 3.6x above. And now the good part, the one that explains this year: earnings are up 73% and the price is down 3%. The multiple has compressed 44% — from 54 to 30 years of earnings. The expected-earnings line crosses above the price in 2028 ($415.65 vs today's $359.10). Translated: if expectations hold, within two fiscal years what it earns is worth more than what you pay now. The company is catching up to its price — and it's halfway there.
What we would do
Not buy it at $359 — but, for the first time in weeks, not because anything is failing. This quarter answered every open question: growth is real, cash tracks profit exactly, debt is falling, stock comp has halved, and next quarter's guide is even bigger. The old reason to wait no longer exists: the company delivered.
| Price | What you would be paying for |
|---|---|
| Below $253 | Very attractive. |
| $253–312 | Attractive. |
| $312–468 ← it is here, at $359 | Fair value, middle of the band. |
| Above $468 | Expensive. |
Neither cheap nor expensive: exactly where it should be for what it is. What holds us back at this specific price are two things, and neither is the business: the cash it generates yields 3.0–3.8% with the dividend while borrowed money costs 5–6% — being in costs you one to three points a year while you wait; and the real product margin is diluting as custom silicon eats the mix. Not a problem today with volume growing 86%. It will be the day it doesn't. And the entry price is no fantasy: twenty months ago this traded at $169.77.
The three things we would watch
1. The true cost of manufacturing, without the accounting charge. THE number this quarter uncovered: from $23.20 to $25.80 per $100 collected. If it rises another two points next quarter, the published margin starts falling too and the whole thesis changes.
2. The research budget. Down 5.1% in dollars with revenue nearly doubled. Flat for two more quarters: margin keeps rising and the question is postponed. A sudden spike: the "customer pays" model has stopped working.
3. Custom AI chip orders. Guided from $16.7B to $21.7B. The variable that rules everything, and it doesn't depend on Broadcom — it depends on four or five companies continuing to commission their own silicon. The quarter that number stops accelerating, a 30-year multiple has nothing to hold on to.
In one sentence
It nearly doubled revenue in a year and the money truly came in, dollar for dollar — but every chip it builds leaves less behind than before and the research budget is smaller than twelve months ago. It is winning the AI race while spending less on inventing.