Broadcom's two moats and one margin line

Two businesses, one margin line

Broadcom is two companies under one ticker: semiconductor solutions at $20.8 billion of third-quarter revenue, 70% of the total, and infrastructure software — the acquired VMware, CA and Symantec estates — at $8.8 billion, the other 30% ^[2]. What binds them is the margin line: non-GAAP operating margin is guided at approximately 66% of revenue for Q4, "flat from a year ago," in the CFO's words, while consolidated revenue is guided up 93% ^[2]. Nearly doubling the top line at a held margin is the whole quarter in one sentence — and, as the rest of this piece argues, the sentence the market had to price.

The XPU customer map

The AI story is custom silicon, and this quarter the programs behind it were named in public detail. The 10-K lists "custom accelerators or XPUs" first among the AI semiconductor solutions, built for "hyperscalers and companies with AI frontier models" ^[1]; the earnings call put customers and capacity numbers on that sentence ^[3], and Meta's own newsroom carries the multi-generation MTIA co-development in its own words ^[4]:

 XPU / custom-accelerator programs, as stated on the record
 ──────────────────────────────────────────────────────────
 Google      TPU co-design; Ironwood + TPU 8i shipments to
             accelerate; "tens of billions of dollars of
             processors to Google each year for the next
             several years"                              [3]
 Anthropic   5 GW of TPU 8i deployment targeted for 2027,
             line of sight to a further 10 GW             [3]
 OpenAI      custom "Jalapeno" chip; 1.3 GW deployment
             eyed for 2027, line of sight beyond 5 GW
             across generations one and two; a third
             generation in planning                      [3]
 Meta        MTIA co-development across multiple
             generations; production shipments of the
             inference-optimized accelerator expected    [3][4]

The quarter's disclosed arithmetic sits under that map: AI semiconductor revenue of $16.7 billion grew 221% year over year and 54% quarter over quarter, and the company expects it "to accelerate to $21.7 billion" in Q4 ^[2]. A custom accelerator is co-designed into one customer's data center over a multi-year program — the design win itself is the switching cost, and it cuts both ways: the customer cannot easily leave, and Broadcom cannot easily replace the customer.

Both sides of the AI network

The less-told half is that Broadcom also sells the network everyone's accelerators hang on. The 10-K describes the franchise as "connectivity at scale, within and across AI server racks and across AI data center sites" ^[1] — a rack built around a Broadcom XPU and a rack built around a rival's accelerator can both be wired with Broadcom switching, NICs and optics. The filing's own competition section carries the counterweight in unusually plain language: competition is expected to increase "as new companies, including some of our customers, enter the markets" ^[1]. The customers designing XPUs with Broadcom today are the hyperscalers with the scale to design without it tomorrow.

Concentration, the Broadcom version

Sales to the top five end customers, through all channels, were approximately 40% of net revenue in each of fiscal 2025 and 2024, with 48% of revenue moving through distributors ^[1]. Forty percent across five buyers is real concentration — the filing says the loss of any one "could have a material adverse effect" ^[1] — but it is a different animal from the ninety-percent, ten-customer book one rung down the interconnect chain ^[7]. The filing does not name the five, and this article will not guess them.

The two-horizon guide

The report beat the consensus on both lines: EPS of $3.32 against $3.24 expected and revenue of $29.59 billion against $29.36 billion, per LSEG ^[3] — the same beat this site's own record shows against its frozen estimate ^[6]. The guide is where the story splits in two:

 horizon              company statement          street expectation
 ───────────────────────────────────────────────────────────────────
 Q4 FY2026 revenue    ~$34.8B guided [2]         $35.03B (LSEG) [3]   below
 FY2027 AI revenue    ~$115B targeted [3]        —                    stated on the call
 FY2028 AI revenue    ~$230B targeted [3]        —                    stated on the call
 FY2028 EPS           over $30 targeted [3]      $25.86 (LSEG) [3]    above

The near horizon printed below the street; the far horizon printed well above it. Both numbers are company statements, not this site's forecasts — the staircase they describe is steep enough that the AI line would roughly double twice from a $16.7-billion quarter ^[2][3].

What the tape did with it

Into the print, the stock was quiet: a September 1 close of $369.68 and a September 2 close of $367.24, down 0.7% on report day ^[5]. The after-hours session then traded the two horizons against each other: early coverage was headlined on the disappointing Q4 revenue guide, and the same CNBC report, updated after the call detail on Google, Anthropic and OpenAI, recorded that "shares edged up in extended trading" ^[3]. The regular-session verdict had not printed when this article was published — the reaction row on this site's earnings record will carry the measured answer once the September 3 close is in ^[6].

Context for whatever that answer is: coming into the report, Broadcom shares had gained about 6% in 2026 against a 12% rise in the S&P 500, on a market value near $1.8 trillion — a stock up more than sixfold since the end of 2022 ^[3].

The transmission mechanism

Two mechanisms, both visible in the cited record. First, the horizon contest: when the next quarter guides below consensus while the two-year statement guides far above it, the price becomes a vote on which horizon the market discounts — the after-hours path from guide-disappointment headlines to "edged up" tracked exactly that vote being recast as the call supplied program-level detail ^[3]. Second, the margin line: a 66% operating margin held "flat from a year ago" through a near-doubling of revenue ^[2] means the earnings staircase is built from volume, not from operating leverage — the same pattern this site's Credo deep dive found a rung lower, where a beat plus a flat margin guide met a fifth-of-value repricing at a far more demanding entry price ^[7]. Reading those two cases together: this earnings wave is pricing AI infrastructure growth as real and margin expansion as finished. That is a description of what the cited record and the tape did, not a view on what either does next.

What the record will answer next

Four things already named, worth re-reading next quarter: whether AI semiconductor revenue keeps outgrowing the semiconductor segment that contains it ^[2]; whether the FY2027 waypoint toward the $115-billion statement appears in the Q4 guide's composition ^[3]; whether the top-five concentration line moves off 40% in the next 10-K ^[1]; and whether the sentence about customers entering the markets acquires named examples ^[1].