Broadcom delivered the quarter and guided the deceleration in the same release. Q3 revenue was a record $29.59 billion, up 86% year over year, with non-GAAP operating income of $20.1 billion, up 92%, and non-GAAP EPS of $3.32, up 96%. Gross margin came in at 75%, down 210 basis points sequentially on AI mix but above the 74% guided, and operating margin rose 240 basis points year over year to 67.9%. Free cash flow was $13.7 billion, 46% of revenue. AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially, with the semiconductor segment at $20.8 billion, up 127%, now 70% of the company. Every one of those figures is a record and none of them is the number that moved the stock, which fell about 5% in extended trading against a close near $368.
The number that moved it sits inside the Q4 guide. Broadcom guided consolidated revenue to approximately $34.8 billion and non-GAAP operating income to approximately 66% of revenue. Apply the second figure to the first and Q4 operating income lands near $22.97 billion, which is $2.87 billion above the $20.1 billion just delivered on $5.21 billion of additional revenue. That is an incremental operating margin of 55.1% against a delivered average of 67.9%. The marginal dollar of Broadcom revenue is now worth about four-fifths of the average dollar, and the company is guiding that relationship down for the first time in this ramp rather than defending it.
The composition explains the mechanism and makes it structural. AI semiconductor revenue is guided from $16.7 billion to $21.7 billion, which is $5.0 billion of the $5.21 billion sequential increase. Infrastructure software is guided to $8.7 billion, roughly flat sequentially despite 25% year-over-year growth, and non-AI semiconductors move by a few hundred million. So essentially all of the growth arrives in the lowest gross margin line in the company, and management already told you this quarter that AI mix cost 210 basis points of gross margin sequentially. The operating leverage that produced the 240 basis point year-over-year margin expansion in Q3 is being spent to buy AI revenue in Q4. That is a defensible trade at these growth rates. It is still a different margin structure than the one the multiple was built on.
On the revenue guide itself, the $34.8 billion sits modestly below the roughly $35.03 billion LSEG consensus, so this was not a beat that failed to impress. It was a light guide accompanied by a guided margin step-down, into a stock that had already fallen roughly 26% from its June 3 record of $495 and carried a market capitalization near $1.76 trillion. The after-hours reaction was proportionate to what was actually said.
The part of the release the tape appears to have ignored is the one that matters over any horizon longer than a quarter. Broadcom raised its AI outlook to roughly $115 billion of AI semiconductor revenue in fiscal 2027 and roughly $230 billion in fiscal 2028, against a fiscal 2026 AI base tracking near $55 billion. That is a four-fold increase in two years, from a company that will finish this year around $104 billion in total revenue. At $115 billion of AI revenue alone, fiscal 2027 consolidated revenue plausibly runs near $165 billion, and the current market capitalization is about eleven times that. The stock is not pricing a company that hits those numbers. It is pricing a company that guides them.
The reason for the discount is not demand and it is not the six XPU customers Hock Tan referenced on the call. It is that $230 billion of accelerator revenue in fiscal 2028 is a supply forecast and a financing forecast wearing a demand forecast’s clothes. It requires wafer, advanced packaging and HBM allocation on a scale that has not been publicly committed by anyone, in a year when memory pricing is already the binding constraint across the industry. It requires the customers to raise the debt. Broadcom has already disclosed the mechanism it uses to help them do that, in the form of the residual value support attached to the initial AI infrastructure transaction, and the max exposure line in the 10-Q is the place where a bigger denominator shows up first. If $230 billion is the right number for fiscal 2028, that exposure does not stay where it is.
The moat argument is unchanged and it is stronger than the quarter suggests. Broadcom’s durable position is not the custom accelerator socket, which is competitively contestable and always has been. It is the networking attach around the socket, the Tomahawk and Jericho franchise that ships whether the compute die is Broadcom’s or somebody else’s, plus a balance sheet that can underwrite customer deployments in a way no other merchant supplier can offer. That second item is a genuine product feature and Marvell cannot match it. On the competitive question, Marvell’s Google warrant implies roughly $17 billion a year of custom revenue at that account against Broadcom’s own $230 billion fiscal 2028 target. Either a full Google defection is immaterial to the target, or the target is not serious. Both readings are available and they cannot both be dismissed.
For the stock, the base case is $340 to $400 into the January quarter, with the guide met, margins near 66% and no re-rating until the fiscal 2027 numbers stop being a forecast. The bull case is a retest of $495 and a walk toward the roughly $500 street average, and it requires the Q1 guide to show incremental operating margin back above 60%, or a named seventh XPU customer, or both. The bear case is a cohort derating to $280 to $310 alongside Marvell and AMD, triggered not by a demand miss but by the market applying a lower multiple to revenue that arrives at 55% incremental margins and depends on customer leverage.
The decision-relevant number in December is the same one that matters today: apply the guided operating margin percentage to the guided revenue for the January quarter and check whether the implied incremental margin is above or below 55%.