Broadcom Q3 FY26 — The Gigawatts Are Settled. The Conversion Rate Is Not.

A record on every line, and a red tape

Revenue of $29,591M, up 86%. Non-GAAP operating income of $20,095M at a record 67.9% of revenue, up 240bps. Non-GAAP EPS of $3.32, up 96%. Free cash flow of $13,665M at 46% of revenue. AI semiconductor revenue of $16.70B against a $16.00B guide — a $0.70B beat, or 4.38% — up 221% year on year and 54% sequentially. Fiscal 2026 AI revenue raised to $58B from $56B. Fiscal 2027, which the June call had only framed as "in excess of $100 billion," given a figure: approximately $115B. Fiscal 2028 introduced at approximately $230B. And AVGO closed at $357.16 against the prior $367.24, down 2.74%, inside an options-implied move of 6.79%.

That is the second consecutive quarter in which this pattern has printed. The Q2 dive published here in June carried the same shape — AI revenue up 143% to $10.80B, Q3 guided to $16B, and a stock that fell. A single red day after a blowout is noise. Two in a row, on prints that beat and raised both times, is the tape telling you the variable it cares about is not on the page it is reading. So the question worth spending this piece on is not whether Broadcom is winning. It is what, exactly, the market thinks it is being asked to pay for — and that turns out to be answerable from the numbers management gave.

The guide's own arithmetic

Start with the enumeration, because Broadcom did something unusual on this call: it named the gigawatts. Anthropic at 1 gigawatt of Ironwood in 2026, 5 gigawatts of TPU version 8i in 2027, an incremental 10 gigawatts in 2028. OpenAI at 1.3 gigawatts of Jalapeno in 2027 and over 5 gigawatts in 2028. Meta at 3 gigawatts of MTIA across three generations through 2028. Asked by Bernstein to confirm the aggregate, Tan did: "it will be 30 gigawatts between the customers we have, the 6 customers we have."

Then he gave the dollars, and the dollars do not match the gigawatts. Fiscal 2027 at $115B and fiscal 2028 at $230B is $345B — Tan rounded it to "about $350 billion" on the call — which against 30 gigawatts is $11.5B per gigawatt. Bernstein got there first and said so, computing "$11 billion and $12 billion per gigawatt" from the guidance. Tan's answer named a different number entirely: "what we are seeing per gigawatt is in the range of less than $30 billion, $20 billion to $30 billion per gigawatt. And we expect that to be very sustaining in that level."

Both figures are his, and they only reconcile one way. If Broadcom's content is $20B to $30B per gigawatt and the two-year revenue is $345B, then between 11.5 and 17.25 gigawatts of the enumerated 30 actually convert to shipped revenue inside fiscal 2027 and fiscal 2028 — 38.3% to 57.5% of the line of sight. Tan said as much without quantifying it: "we're not saying that over the next 2 years, '27, '28, that there are 30 gigawatts that will go into production... We think we judge it conservatively to be somewhat less." The chart carries the gap, because a sentence cannot.

This is the gap between the narrative and what the data supports in its purest form, and it is not a criticism of the guide — it is a description of what the guide measures. Tan was explicit about the mechanism when Truist asked: "some of this part on the site's power shell has a long lead time, the construction project... we do that not development, but analysis with our customer, and we reflect that in the forecast outlook we're giving you today." Read plainly, the $115B and the $230B are not semiconductor demand forecasts. They are construction-schedule forecasts with a silicon content multiplier applied. Broadcom is telling you it can build the chips; what it is forecasting is whether anyone will have a building to put them in.

The margin number the CFO had to correct

Consolidated gross margin was 75%, down 210bps sequentially, and guided to approximately 73% in Q4 against 78% a year ago — roughly 500bps of decline in four quarters. Amie Thuener attributed it in one clause: "as XPUs become a larger proportion of our total revenue mix, it impacts our margin," with "increasing memory content" named as the cause. That is for the picks-and-shovels vendor, the funding question is a margin question stated by the company itself. Broadcom is not carrying the customers' capex on its balance sheet; it is carrying it in the bill of materials, because the HBM inside an XPU is bought at someone else's price.

The segment number is where it gets interesting, and it took two tries. Thuener first read out that "gross margin for our Semiconductor Solutions segment was approximately 76%," then came back later in the Q&A: "I want to just make sure I correct what I said before, just to be super crisp. Gross margin for our Semiconductor Solutions segment was approximately 67% in Q4 -- or Q3." Nine points. The corrected figure is the right one and it checks against the consolidated line: at 70% of revenue from semiconductors and 30% from software at 94%, the blend lands on the reported 75%. What the correction does is remove the softest reading of the print. On the first number, the segment carrying $16.70B of AI revenue was still a 76-point-margin business. On the corrected one, the segment producing 70% of revenue earns a gross margin nine points below the corporate average, and the entire consolidated figure is being held up by an infrastructure-software business growing 29%.

Tan's response to the third analyst who raised it was to change the instrument: "stop focusing on gross margin is what we're saying, look at where it matters, operating margin at the bottom of the -- at the end of the day." He is not wrong, and the numbers back him — semiconductor segment operating expenses were $1.2B, 6% of segment revenue, growing 22% against revenue growth of 127%, which is why segment operating margin went to 61%, up 440bps. But note the timing. The instruction to change the measuring stick arrived in the same session the old stick moved against him, and the new one is guided a single quarter at a time. Thuener said so directly: "We guide one quarter at a time. So we'll tell you each quarter how -- what our margin is going to look like."

The fabless company that is buying fabs

Here is the line nobody asked a question about. Capital expenditure was $532M in the quarter — 3.75x the $142M of a year ago and 3.11x the quarter's $171M of depreciation — and Q4 is guided to $1.4B. That single guided quarter is 1.38x the $1,013M Broadcom spent across the first three quarters combined, 2.63x Q3, and equal to 44.53% of the entire $3,144M net fixed-asset base. As a share of revenue, capex goes from 0.89% in Q3 FY25 to 1.80% this quarter to 4.02% of the guided $34.8B.

Four consecutive step-ups is a series, not a quarter, which is the only reason this qualifies as a finding at all — a one-period change is a phase, not a decision, until you have the series, and this newsletter got that wrong on CXMT's capex two days ago. The series here points one way, and the leading disclosures agree with the lagging cash. Charlie Kawwas named the programmes: "our EML, CW, and VCSEL factories, our indium phosphide factories, both in the U.S. as well as in Singapore, we're actually more than tripling them year-on-year," plus a substrate fab in Singapore going into production in fiscal 2027. Tan gave the reason without softening it: substrates and HBM memory are bottlenecks, "which is leading us to build our own substrate capacity at scale in our factory in Singapore."

Decode that. A company whose entire multiple rests on being asset-light — $13,665M of free cash flow on $532M of capex is 46% of revenue — is putting money into fabs because it cannot buy what it needs. In absolute terms this is nothing: $1.4B against $14,197M of quarterly operating cash flow does not move the model. The information is in the direction and the stated reason. Vertical integration is what a company does when the merchant market has stopped clearing, and it is the second place in this print, after the gross margin, where the cost of the AI ramp shows up on Broadcom's side of the ledger rather than the customer's.

What "$30 in earnings per share" actually claims

Tan closed the outlook with a number that reads like an additional promise: "we are very much on target to exceed $30 in earnings per share in fiscal 2028." It is not an additional promise. It is the same promise with the margin held still, and the arithmetic is short enough to do here.

$30 on the guided 4.94 billion diluted shares is $148.2B of non-GAAP net income. At the guided 16% tax rate that is $176.43B pre-tax, or roughly $178.8B of operating income once interest is added back. Hold non-AI semiconductors and infrastructure software at the run-rates Broadcom guided for Q4 — $4.3B and $8.7B a quarter, or approximately $17.2B and approximately $34.8B a year — and fiscal 2028 revenue with $230B of AI is approximately $282B. That implies an operating margin of approximately 63%, against the 66% guided for Q4 FY26. Run it the other way and the point is the same: fiscal 2026 non-GAAP net income derives to approximately $57.42B, so $148.2B is 2.58x over two years, against total revenue growth of 2.66x from approximately $105.89B.

Earnings and revenue grow at almost exactly the same rate. That is the whole content of the $30 figure: it asserts that two more years of memory-driven gross-margin dilution get fully absorbed by operating leverage, costing about 300bps at the operating line and nothing else. It contains no information the $230B did not already contain. Which means the entire forward model on this name reduces to one line item — the operating margin — and that is the line management guides one quarter at a time.

Who is funding the gigawatt

Broadcom answered who ultimately pays for the capex more explicitly than most vendors ever do, and then declined to size it. In June it established the AI XPV platform with Apollo and Blackstone to enable more than 20 gigawatts of compute infrastructure for OpenAI and Anthropic by the end of 2028, closing a first $35B tranche for Anthropic's 1 gigawatt. Thuener described the structure precisely: "we partner with sophisticated third-party financial partners to independently underwrite and capitalize the assets rather than providing the direct financing ourselves. Where necessary, we may provide modest residual value guarantees, which are contingent liabilities we view as low risk."

Two words are load-bearing there, and both are qualitative. "Modest" is not a number, and "low risk" is a view. Asked by Melius to put a maximum on the off-balance-sheet exposure — the analyst citing a prior 10-Q figure of about $29 billion on the first tranche the $29B figure is the analyst's characterisation of a prior filing and is not verified here — Thuener said there was nothing to announce and would not give an overarching figure. That is a fair answer to an unanswerable question and it is also the answer that leaves the exposure unbounded in the reader's model.

This is the vendor becomes the credit, engineered one step further than the version this newsletter described on Dell yesterday. Dell put $6.67B of financing receivables on its own balance sheet and took the free-cash-flow hit in the quarter. Broadcom routed the same economic function through third-party capital and retained only the residual-value tail — the piece that never appears in EPS, never appears in the multiple, and only shows up if the assets are worth less than someone underwrote them at. The instrument is better. The question it answers is identical, and Tan answered it out loud when Bank of America asked why the two labs get financing help the other four customers do not: "every gigawatt of compute they deploy, they could achieve $30 billion of ARR. That's a hell of a business model."

Hold that against his other number. Broadcom's content is $20B to $30B per gigawatt. The customer's annual recurring revenue from that gigawatt is $30 billion. So the silicon bill for a gigawatt is roughly one year of the revenue that gigawatt will eventually produce — payable before the first token is served. That single ratio is why the XPV platform has to exist, and it is the honest form of the funding question on this name: not whether Broadcom gets paid, but how many years of a customer's future revenue have to be financed up front for it to get paid on schedule.

How this reads across the complex

The most consequential sentence for everyone else in the chain is Tan's supply claim: "In 2027, we have secured the supply to again double AI revenue to approximately $115 billion... Here again, we have secured the supply to meet this outlook." Supply that is secured is supply that is not available, which makes this a commitment is an allocation claim before it is a cost — the model this newsletter wrote against NVIDIA's memory pre-commitment in August, now with a second buyer of comparable scale standing in the same queue.

• Memory and HBM. Tan named the bottleneck twice: "we all know about memory, HBM memory, and beyond HBM memory, the system memory that goes into AI servers, which we don't supply necessarily, but our customers have to secure too." Broadcom securing HBM against $345B of two-year AI revenue is a scheduling statement about the DRAM makers' order books, and it is constructive for them for the same reason it is dilutive for Broadcom's gross margin. The memory content moving from the supplier's revenue line into Broadcom's cost line is the single mechanism behind both.
• Substrates. A fabless designer building its own substrate capacity at scale in Singapore, in production from fiscal 2027, is a scarcity disclosure about the merchant substrate market, not a strategy announcement. Read it as tightness confirmed and, over time, as one large buyer removing itself from that queue.
• Optical. "Demand for lasers, whether it's EML lasers, CW lasers, is far surpassing supply out there in the industry," with Broadcom's own EML, CW, VCSEL and indium phosphide capacity "more than tripling" year on year. Near-term that confirms the scarcity premium across the merchant laser and optical-DSP names; on the two-year view, the largest incremental buyer is also becoming an incremental supplier.
• Networking. AI networking revenue up over 2.5x to approximately $4.51B, Tomahawk 6 shipping in both 100-gigabit and 200-gigabit SerDes versions across "pretty much all of the AI hyperscalers that are building XPUs with us" and some that are not, and Tomahawk 7 taped out at 200 terabits per second. Tan said AI networking "is expected to grow just as fast as XPUs over the next few years," which is the part of the guide with the least memory content in it and therefore the best margin.
• Marvell. Last week this newsletter published the rebate that indexes to your own stock on Marvell's Google warrant — 58,970,907 shares vesting against Custom Products revenue. Broadcom announced a long-term agreement with the same customer covering "multi-tens of billions of dollars of TPUs annually over the next several years" and gave away no equity to get it. Same buyer, two prices. The comparison is the cleanest available read on which vendor is negotiating from scarcity and which from need.

The strongest case against this read

The best argument against everything above is that gross margin is simply the wrong denominator, and that the tape is mispricing a mix shift as a deterioration. On this reading, an XPU carrying expensive HBM and sold at a lower gross margin but requiring no incremental engineering drops more absolute operating profit than a higher-margin chip that does not exist — and the filing supports it. Non-GAAP research and development was $1,551M in the quarter against $1,600M in Q2 and $1,477M a year ago; across nine months it grew 5.35%, from $4,432M to $4,669M, while revenue over the same period grew 54.97%, from $45,872M to $71,089M. Total non-GAAP operating expenses were $2,096M against $2,044M. That is roughly flat spending against revenue up more than half, and it is why operating margin can be guided flat at 66% while gross margin falls 500bps. On that view the gross-margin decline is not the cost of the ramp; it is the evidence the ramp is working, and it should be welcomed rather than tracked.

I think that argument is correct for fiscal 2026 and probably for fiscal 2027, and untested for fiscal 2028 — which is exactly the year the $230B and the $30 sit in. It requires engineering spend to stay near flat while the programme count rises: 6 XPU customers, successive TPU generations for Google, three MTIA generations for Meta between now and the end of 2027, Jalapeno plus a successor approaching tape-out plus a third-generation part in development for OpenAI, Tomahawk 7 taped out and Tomahawk Ultra ramping. R&D falling sequentially from $1,600M to $1,551M against that programme list is either extraordinary leverage or a timing artifact, and one quarter does not tell you which — the same discipline I applied to the capex series has to apply here, pointed at my own argument.

Valuation, the trade, and what settles it

At $357.16 on 4,887M GAAP diluted shares, Broadcom is worth approximately $1.75T. Reported non-GAAP EPS for the first three quarters was $2.05, $2.44 and $3.32, or $7.81. Holding every guided Q4 input — $34.8B of revenue, 66% operating margin, a 16% tax rate, 4.94 billion shares, with non-GAAP interest and other income held at Q3's $703M and $98M — Q4 derives to approximately $3.80 and fiscal 2026 to approximately $11.61. Call it 30.8x this year.

And approximately 11.9x the company's own fiscal 2028 figure. That is the number to sit with. A business guiding AI revenue to quadruple in two years, with the supply secured, is being asked to trade at under twelve times the earnings it says that produces. The market is not disputing fiscal 2026 — it beat and the stock fell anyway. It is discounting the conversion rate, which is precisely what the $11.5B-per-gigawatt arithmetic says management is discounting too. The disagreement between the tape and the guide is smaller than it looks; they are arguing about the same variable from opposite sides.

On structure — analyst framing, not advice — the post-print setup is unusual in one respect worth naming. Thirty-day implied volatility collapsed to 0.359 from 0.490 the prior session, with an IV rank of 0 and a 7.0% implied move. Optionality on a name with a dated catalyst is priced at the bottom of its own one-year range, which argues against selling premium here and in favour of owning dated exposure into the December print for anyone who wants the conversion-rate question expressed rather than the direction. The residual-value guarantees and the unquantified XPV tail are the reason not to size it as if the balance sheet were still purely fabless.

Bottom line. Nothing in this print argues against the demand. AI revenue beat its own guide by 4.38%, the segment producing it grew 127%, operating margin hit a record 67.9%, and free cash flow was 46% of revenue. What changed is which number carries the risk. It moved off the top line — where it has been for two years — and onto the two lines that translate a gigawatt into a dollar: the gross margin the memory chain is now taking a share of, and the schedule on which a data centre gets built. Both are outside Broadcom's control, and management priced both into the guide before the market did.

The test is dated. Q4 and fiscal 2026 report after the close on December 9, 2026. If the fiscal 2027 figure moves above approximately $115B with the same six customers, supply was the binding constraint, this read is wrong, and the right model is the old one: demand, uncapped. If approximately $115B holds or is re-cut while Q4 AI revenue meets the $21.7B guide, the constraint is site readiness — and every quarter from here should be modelled on gigawatts converted, not chips sold.