Broadcom Clears the Bar: The Demand Ratio Gets a Quarter of Evidence in Its Favor

Correction, September 6, 2026: the original same-night version of this note, relying on press coverage of the call, said fiscal 2027 AI semiconductor guidance was reiterated at "in excess of $100 billion" rather than raised, and framed part of the after-hours selloff as a repeat of June's reaction to a reiterate. Having now reviewed the call transcript, that was wrong. Management raised fiscal 2026 AI revenue guidance to approximately $58 billion, guided fiscal 2027 AI revenue to approximately $115 billion ("we have secured the supply to again double AI revenue"), and outlined roughly $230 billion for fiscal 2028. Three passages have been corrected and are marked where they occur. The selloff explanation now rests where this note already placed most of its weight: the total fourth-quarter guide of $34.8 billion against the street's $35.03 billion compilation. The corrected facts strengthen the demand evidence this note scored, which is irrelevant to the standard: they were wrong. We disclosed at publication that no transcript had been reviewed; the disclosure was honest, and the process it described was not good enough. A transcript-verification pass is now a required follow-up on every earnings note, and this page is its first product.

The Print

Broadcom reported fiscal third-quarter 2026 results after the close on September 2. Consensus figures below are per press previews compiled before the print; actuals and guides are per the company's release and call.

The release itself is plain: record revenue, record operating profit, and free cash flow of $13.7 billion, about 46 percent of revenue, per the company. Infrastructure software added $8.8 billion, up 29 percent. The AI semiconductor line did the work: $16.7 billion in the quarter and $21.7 billion guided for the fourth quarter, which annualizes to a run rate well above the fiscal 2027 target that management declined to raise in June. On the call, per press coverage, CEO Hock Tan said the company now has six core custom-chip customers driving AI revenue, including Anthropic, Google, Meta, and OpenAI, and that Anthropic is on track to become the largest XPU customer in 2027, with deployment commentary running from roughly 1 gigawatt in 2026 toward 5 gigawatts in 2027 and beyond 10 gigawatts incrementally by 2028.

What It Says About the Scoreboard

We publish break conditions on every framework we maintain, and we score new evidence against them in public. The September Setup flagged this print as September's first dated evidence on whether AI compute demand is still outrunning its own price deflation. The context that makes tonight's AI number matter: on June 3, 2026, Broadcom guided third-quarter AI revenue to $16.0 billion against a consensus near $17.2 billion, per press reports of the call, and on June 5 the Nasdaq fell 4.2%, its worst day since April 2025. Tonight is the quarter that guide covered.

Two more pieces of context the reader deserves. First, the June pattern: after the June 3 report, shares fell 12.6 percent in one session despite beating on revenue and EPS, because management reiterated rather than raised its fiscal 2027 AI semiconductor target of "in excess of $100 billion," per press coverage of the report. The same question governed tonight, and options priced a move of roughly plus or minus 8.7 percent for this print against a ten-year average near 7.4 percent, per options-market compilations published before the release. Second, the tape: this print landed one day after a global rate shock, with the 10-year Treasury near 4.8 percent and chip stocks leading a selloff that was largely bought back into the September 1 close, per market coverage. A rate-driven multiple squeeze maps to our capex-throttle branch, not the demand-ratio branch, and we grade the guidance against the conditions rather than against the stock's first move. The swing factor, corrected September 6 after transcript review: fiscal 2027 AI semiconductor guidance did not stay at "in excess of $100 billion" as this note originally said. Management guided approximately $115 billion for fiscal 2027 and outlined roughly $230 billion for fiscal 2028, with order visibility extending into 2028. The market sold the print anyway, which sharpens the point rather than softening it: the total fourth-quarter guide against the street's compilation, not the AI outlook, drove the reaction.

The conditions this print bears on, quoted from our conditions page as they stood August 31, 2026, all HOLDING:

Broadcom's custom-accelerator line is one of the cleanest reads on that first condition available anywhere in the disclosure calendar: it is AI compute revenue reported by a supplier whose customers are the profitable hyperscale buyers themselves. Six days ago NVIDIA reported data-center revenue up 117% year over year and management said customer forecasts point to demand roughly doubling, per the company's August 26 report and call. Tonight's print either rhymes with that or it does not.

How we score the outcomes

We wrote the following decision tree before the print, so the reader can check our scoring against our own pre-commitment. Opinions throughout, labeled as such:

Tonight's actuals land in the first row: AI revenue above the $16.0 billion guide, and a fourth-quarter AI guide that steps up sequentially. Applying the read we committed to before the print: condition 6 is holding and reinforced. Demand from the profitable buyers is still outrunning price deflation at the custom-silicon layer, and conditions 7 and 13 are untouched. One thing the tree did not anticipate: the size of the step. A $21.7 billion fourth-quarter AI guide is not a sequential nudge, it is an acceleration of the acceleration, and it rhymes loudly with NVIDIA's August commentary. The call added something the tree did not anticipate: a sixth core custom-chip customer base with order visibility management says extends into 2028, and Anthropic described as on track to become the largest XPU customer next year, per coverage of the call. And the market added something else the tree explicitly did anticipate: the stock fell about 5 percent after hours because the total fourth-quarter guide of $34.8 billion missed the street's $35.03 billion. That is the tree's fourth row in action. The shortfall against consensus sits outside the AI line, the conditions are about AI demand, and we say that plainly even though the stock moved hard.

The Honest Negatives

First, margins. Fourth-quarter non-GAAP operating margin is guided to about 66 percent of revenue, per the release, a point below the prior quarter's roughly 67 percent framing. That is the arithmetic of custom silicon scaling: XPUs carry lower margin than the merchant networking they sit beside, so the fastest-growing line dilutes the mix even while it builds the moat. Bulls will call a point of margin cheap for 236 percent AI growth. They are probably right, and it is still the direction to watch.

Second, concentration. The AI line remains a bet on the delivery schedules of a handful of hyperscale customers, and a $21.7 billion quarterly guide leans harder on each of them. Custom silicon is a contract business: extraordinary visibility until a single customer's roadmap changes, then a cliff where a merchant vendor would have a slope. Management itself flagged margin pressure on the call, per post-earnings coverage, which is the version of this concern stated in the company's own voice.

Third, positioning. The stock consolidated between roughly $365 and $377 into the print with options pricing a move near 8.7 percent, so a large reaction in either direction was pre-paid. The reaction was down about 5 percent in extended trading, with some coverage citing declines beyond 6 percent, because the street's compiled Q4 estimate was $35.03 billion and the guide was $34.8 billion. Corrected September 6: tonight did answer the question the multiple rides on. The fiscal 2027 AI outlook became a bigger number, approximately $115 billion, with roughly $230 billion outlined for fiscal 2028. Our original version said management let the order book argue instead of the guidance line; the transcript shows they did both, and the stock fell anyway on the total-revenue guide. That is a sterner verdict on positioning than the one we originally printed.

What Would Change Our Read

The next dated evidence: the October and November hyperscaler capex guides, which bear directly on condition 7; NVIDIA's next quarterly report, which is the other clean read on the demand ratio; and Broadcom's own fourth-quarter report in December, where the $21.7 billion guide either converts to revenue or does not.

The qualitative signal that would connect this name to condition 5, "Guarantee contagion": customer guarantee or vendor-financing structures appearing in custom-silicon agreements. Contract manufacturing paid for by the buyer's balance sheet is healthy; silicon paid for by the seller's is the pattern the condition exists to catch.

What would make us say tonight's read was wrong, stated as concretely as the conditions themselves: if it emerges that the fourth-quarter AI guide leans on pull-forwards or one-time shipment timing rather than run-rate demand, then "reinforced" overstates the evidence and we will grade it down in public, the same way this note graded it up.