Broadcom’s AI ASIC guide is the real beat, while the EPS print was a sideshow

Broadcom did not deliver a dramatic Q3 beat, and that is precisely why the print matters: the quarter itself was only modestly ahead, but the forward composition sharpened the bull case. What was priced in was a clean AI-led quarter with limited estimate risk, reflected in revenue of $15,952.0 million versus the Street at $15,826.1 million and EPS of $1.69 versus $1.66. What actually surprised was not the +0.8% revenue beat or the +1.8% EPS beat, but the company’s willingness to frame Q4 around a step-up in AI semiconductor revenue to approximately $6.2 billion and consolidated revenue to approximately $17.4 billion. The variant perception is that investors are still treating Broadcom as a diversified compounder with AI upside, when this print argues the earnings power is increasingly dictated by a narrower set of custom AI programs whose growth is already large enough to absorb seasonal non-AI cyclicality and still leave operating margin intact.

That distinction between a modest print and a more consequential guide is important because the headline quarter alone could be misread as merely in line. Hock Tan described the company’s own basis plainly: “In our fiscal Q3 2025, total revenue was a record $16 billion, up 22% year on year.” The Street-comparison basis was $15,952.0 million, and the surprise was only +0.8%, so the market’s first reaction should not be to pay for a big backward-looking beat. The better reading is that Broadcom validated the next leg of the model: Q3 semiconductor revenue was $9.2 billion, while AI semiconductor revenue was $5.2 billion, and management guided that AI piece to approximately $6.2 billion in Q4. That is the part of the call that changes the debate, because the AI franchise is no longer an offset to softness elsewhere; it is becoming the primary swing factor in total company revenue and mix.

The financial trajectory supports that view because revenue has broken out of the pre-VMware and early integration range while gross margin has not collapsed under the AI mix. The historical series shows revenue moving from $8,876.0 million in Q3 FY2023 to $15,952.0 million in Q3 FY2025, with gross margin still at 67.1% in the latest reported quarter. The market concern is easy to understand: Q3 gross margin was below the 68.0% seen in each of Q1 FY2025 and Q2 FY2025, and AI accelerators plus networking content can carry different margin structures than legacy franchises. But the margin debate is too narrow if it stops at gross margin. Kirsten Spears said the company absorbed a mix hit while still expanding below the line: “On a sequential basis, even as gross margin was down 100 basis points on revenue mix, operating margin increased 20 basis points sequentially to 65.5% on operating leverage.” That sentence matters because it commits to the mechanism: the AI ramp is not just revenue growth, it is revenue growth with enough scale to pay for the leading-edge R&D bill.

The capacity story explains the margin guide, because management is spending ahead of demand while keeping conversion high. Operating expenses increased 9% year on year to $951 million on higher R&D for leading-edge AI semiconductors, yet adjusted EBITDA was $10.7 billion, or 67% of revenue. Free cash flow was $7 billion and represented 44% of revenue, which gives Broadcom room to fund design activity, packaging commitments, and shareholder returns without asking investors to finance an AI buildout through deteriorating cash conversion. Inventory also supports the forward revenue argument rather than warning of channel stuffing: Spears said inventory was $2.2 billion, up 8% sequentially, “in anticipation of revenue growth next quarter.” The number to focus on is not the inventory build alone, but the linkage to Q4 revenue of approximately $17.4 billion. If inventory is rising into a guided revenue step rather than into a missed demand environment, the working-capital signal is confirmation, not contradiction.

The mix issue is also less binary than the market narrative of AI up, legacy down. Hock Tan’s Q4 language explicitly gives both halves of semis a role, saying, “In contrast, in Q4, driven by seasonality, we forecast non AI semiconductor revenue to grow low double digits sequentially to approximately $4.6 billion.” That matters because Q4 semiconductor revenue is guided to approximately $10.7 billion, and AI is expected to be $6.2 billion of that. The surprise is not just that AI remains large; it is that non-AI is not being left behind at the trough while AI carries the entire company. A low double digits sequential move in non-AI semis changes the risk profile for the quarter because a miss would now have to come from either a failed AI ramp or a seasonal recovery that does not appear, rather than from one highly visible AI bucket alone.

Infrastructure software gives the AI thesis a second stabilizer, and the market may still be under-crediting the VMware contribution because the semiconductor narrative dominates the multiple. Q3 infrastructure software revenue was $6.8 billion, up 17% year on year, and represented 43% of revenue. Management expects approximately $6.7 billion in Q4, up 15% year on year, which means software is not accelerating sequentially into the guide, but it is preserving the consolidated margin profile while semiconductor AI ramps. The important number is infrastructure software operating margin of approximately 77%. That margin level allows Broadcom to fund AI ASIC R&D while keeping consolidated profitability high, and it is why the 67.1% gross margin in Q3 should not be analyzed like a pure fabless AI margin print. Broadcom’s model is increasingly a barbell: custom AI silicon supplies the revenue acceleration, while VMware-heavy software protects operating economics.

The customer read-through is most acute for Alphabet, because the data pack identifies Google as a silicon implementation partner for TPU ASICs at approximately $8B/yr. Broadcom’s AI semiconductor revenue of $5.2 billion in Q3 and Q4 guidance of $6.2 billion imply that the custom ASIC ecosystem remains supply constrained by execution rather than demand rhetoric. For Avnet, the read-through is narrower but still relevant: distribution and logistics benefit from a non-AI semiconductor recovery to approximately $4.6 billion in Q4, not just from AI programs that may bypass traditional channels. The supplier map is equally specific. TSMC is tied to 3nm/5nm networking and custom AI ASIC fabrication, so the AI guide points to continued leading-edge wafer demand, while Keysight has test-and-measurement exposure as AI semiconductor revenue scales. WIN Semiconductors has GaAs RF wafers at approximately 20% of WIN revenue, but the strongest Q4 read-through for WIN is not AI, it is the seasonal non-AI semiconductor rebound that management sized at approximately $4.6 billion.

The optical and IP suppliers also get a more nuanced signal than “AI good.” Coherent is tied to Broadcom CPO optical components, and Lumentum is tied to Broadcom Bailly CPO, so the AI semiconductor guide supports continued investment in networking architectures around custom accelerators rather than only compute ASICs. Cadence, Rambus, and CEVA are exposed through IP and design content, but the magnitude that matters is Broadcom’s R&D intensity: consolidated operating expenses included $1.5 billion of research and development, while semiconductor operating expenses increased 9% year on year to $951 million. That spending base is the evidence that Broadcom is not harvesting one generation of AI wins; it is funding successive leading-edge designs. GlobalFoundries has specialty-node exposure, but the print’s center of gravity is the 3nm/5nm custom AI chain rather than mature-node normalization.

Against peers, Broadcom screens less like a high-beta AI pure play and more like a margin-protected AI infrastructure supplier, which is exactly why the print can be mispriced. NVDA’s latest peer revenue growth of +85.2% and gross margin of 74.9% remain in a different category, so the right comparison is not whether Broadcom can match GPU acceleration. The relevant peer contrast is that Broadcom’s Q3 revenue grew +22.0% year on year with gross margin of 67.1%, while Alphabet posted +21.8% revenue growth and 62.4% gross margin. That comparison is not meant to equate the businesses; it highlights that Broadcom is producing hyperscaler-like top-line growth with a semiconductor-plus-software margin structure. If investors value Broadcom only as a diversified fabless company and discount the VMware margin floor, the print argues that discount is too punitive.

The call delivery backs up the thesis, although it also shows why management’s tone did not sound euphoric despite a better forward setup. The tone history shows Q3 FY2025 sentiment at 0.17, guidance_tone at 0.39, and uncertainty at 46.1. That is not a promotional call profile; guidance_tone was lower than Q1 FY2025 at 0.55 and Q2 FY2025 at 0.43, even as the company guided Q4 consolidated revenue to approximately $17.4 billion. The tension is useful. Management is giving investors specific revenue commitments while avoiding a tone spike that would suggest expectations are being pulled forward too aggressively. The delivery was especially notable because qa_evasiveness was -17.5, which argues the call was not hiding the mix debate behind vague AI language.

That measured tone also frames the one real risk in the quarter: AI growth is becoming so large that Broadcom’s margin optics will increasingly be judged by mix rather than scale. Q3 gross margin was 67.1%, and Spears acknowledged a 100 basis points sequential gross-margin headwind from revenue mix. If AI ramps faster than software or non-AI recovery, investors could see more quarters where gross margin pressure masks operating leverage. The offset is that Q3 operating income was $10.5 billion, up 32% from a year ago, and operating margin increased to 65.5%. Those numbers conflict only at the line-item level: gross margin says mix is a headwind, while operating margin says scale is overwhelming it. For now, the operating margin number is the one to underwrite because the business is converting AI demand into earnings before the gross-margin debate becomes visible in cash flow.

The balance sheet does not change the thesis, but it limits the downside case that AI investment will crowd out capital returns. Broadcom ended Q3 with $65.8 billion in fixed rate debt carrying a 3.9% weighted average coupon rate and 6.9 years to maturity. Floating-rate debt was only $500 million, with a 4.7% weighted average interest rate and 0.2 years to maturity. That liability structure matters because the company paid stockholders $2.8 billion of cash dividends in Q3 while still generating $7 billion of free cash flow. The print therefore does not ask PMs to choose between AI investment, VMware deleveraging, and cash returns. It shows all three can coexist as long as the Q4 revenue ramp materializes.

What to watch next is concrete. The thesis is confirmed if, on the next print, Q4 consolidated revenue lands near the approximately $17.4 billion guide, semiconductor revenue reaches approximately $10.7 billion, and AI semiconductor revenue is approximately $6.2 billion. The thesis is strengthened if gross margin stays near the 67.1% Q3 level while operating margin remains around 65.5%, because that would show mix pressure is still being offset by scale. It breaks if non-AI semiconductor revenue fails to recover toward approximately $4.6 billion, because that would make the Q4 guide too dependent on a single AI ramp. The date anchor is the next quarterly update after the 2025-09-04 call, and the numbers to underwrite are simple: $17.4 billion for total revenue, $6.2 billion for AI semis, and approximately $6.7 billion for infrastructure software. If those hold together, the market is still too focused on the modest Q3 beat and not focused enough on Broadcom’s transition into a custom AI revenue compounder with software-grade operating protection.