Broadcom’s print says the AI ASIC cycle is moving from narrative to contracted revenue

The actionable read from this quarter is that Broadcom’s AI business has crossed the line from optionality into scheduled capacity consumption, while the rest of the company is behaving more defensively than the headline multiple debate implies. What was priced in was a clean AI beat, given the Street already expected $17,465.9 million of revenue and $1.87 of EPS. What actually surprised was the scale and specificity of demand conversion: revenue came in at $18,015.0 million, a +3.1% surprise, EPS came in at $1.95, a +4.3% surprise, and management framed the next leg around backlog rather than around customer interest. Hock Tan’s most important line was not the record-quarter language, but the delivery commitment: “We expect this $73 billion in AI backlog to be delivered over the next eighteen months.” That sentence changes the burden of proof. The short case now has to argue either that the backlog fails to ship, or that the economics deteriorate as it ships. The quarter does not support either view yet.

The distinction matters because the printed numbers were good, but not sufficient by themselves to justify a re-rating if the market treats them as another AI-semiconductor momentum quarter. Q4 FY2025 revenue of $18,015.0 million rose +12.9% sequentially and +28.2% year over year, with gross margin at 68.0% and diluted EPS of $1.74 on the company’s quarterly history basis. The Street-comparison basis shows actual revenue of $18,015.0 million versus estimate $17,465.9 million and actual EPS of $1.95 versus estimate $1.87, so the beat was real but not the main event. The market likely expected upside because Q3 FY2025 had already shown revenue of $15,952.0 million, +6.3% sequentially and +22.0% year over year, with AI demand visibly pulling the consolidated model higher. The variant view is that Q4 moved the debate from “how fast is AI growing?” to “how much of the next eighteen months is already spoken for?” That is a more durable setup than a one-quarter revenue surprise.

The financial trajectory reinforces that this is not just volume without margin discipline, although the next quarter introduces a margin question that investors should not ignore. Gross margin in Q4 FY2025 was 68.0%, matching Q1 FY2025 at 68.0% and Q2 FY2025 at 68.0%, and above Q3 FY2025 at 67.1%. That matters because the company absorbed a mix shift toward AI semiconductors without giving back the post-VMware margin recovery visible after Q4 FY2024 gross margin of 64.1%. The conflict comes from the forward quarterly history, where Q1 FY2026 revenue is listed at $19,311.0 million with gross margin of 65.6%, followed by Q2 FY2026 revenue of $22,187.0 million and gross margin of 67.2%. The numbers say revenue acceleration is not automatically gross-margin expansion in every quarter. A PM should therefore underwrite the backlog as revenue visibility, not as a straight-line margin lever, until the company proves that the Q1 FY2026 gross margin of 65.6% was temporary rather than the price of scaling AI systems.

That margin caveat is manageable because the segment mix is doing the heavy lifting in dollars. On the company’s own reported basis, Hock Tan said, “In our fiscal 2025, consolidated revenue grew 24% year over year, to a record $64 billion, driven by AI semiconductors and VMware.” The point of quoting that line is the attribution: management tied the full-year step-up to two engines, not to a broad cyclical rebound. For fiscal 2025, AI revenue grew 65% year over year to $20 billion, semiconductor revenue reached $37 billion for the year, and infrastructure software revenue grew 26% year on year to $27 billion. In Q4, semiconductors were $11.1 billion with year on year growth accelerating to 35%, while infrastructure software revenue was $6.9 billion, up 19% year on year and above the company’s outlook of 6.7. This is the core of the thesis: Broadcom is not being pulled by one narrow accelerator product cycle. It is monetizing custom AI silicon, switching, and VMware at the same time, which is why consolidated Q4 revenue reached $18 billion on the company’s call basis and $18,015.0 million on the Street-comparison basis.

The AI composition also matters because the market often treats Broadcom as a second-derivative AI name behind merchant GPU leaders, when this print argues it is becoming an infrastructure allocation name in its own right. AI semiconductor revenue in Q4 was $6.5 billion, up 74% year on year, and the company guided Q1 AI semiconductor revenue to $8.2 billion, up approximately 100% year on year. Hock Tan also put a hard number on switching demand: “Our current order backlog for AI switches exceeds $10 billion as our latest 102 terabyte terabit per second Tomahawk six switch, the first and only one of its capability out there, continues to book at record rates.” The quote earns its place because it shows the AI backlog is not only custom compute; networking is becoming a contracted bottleneck. That has two implications for valuation. First, ASIC revenue is not the only AI lever. Second, the switch backlog reduces the probability that AI growth falls away immediately after a customer-specific ASIC ramp.

The non-AI semiconductor line is the reason this is still a disciplined thesis rather than a blanket bullish read on every Broadcom end market. Q4 non-AI semiconductor revenue was $4.6 billion, up 2% year on year and up 16% sequentially based on favorable wireless seasonality, but Q1 non-AI semiconductor revenue is forecast at approximately $4.1 billion, flat from a year ago and down sequentially due to wireless seasonality. Those numbers tell investors not to pay for a synchronized semiconductor recovery. They also help explain why Q1 consolidated revenue guidance of approximately $19.1 billion, up 28% year on year, depends on AI and software rather than on a cyclical snapback in wireless or legacy semis. The market may miss this nuance: a flat non-AI forecast of approximately $4.1 billion actually improves the quality of the AI read, because the consolidated guide is not being flattered by broad-based end-market recovery. If Q1 revenue reaches approximately $19.1 billion while non-AI semis are flat from a year ago, the incremental growth is exactly where investors want evidence.

Software adds a second contracted-revenue layer, but it also creates the cleanest place for skepticism next quarter. Infrastructure Software revenue was $6.9 billion in Q4, up 19% year on year and above outlook of 6.7, while Q1 infrastructure software revenue is expected to be approximately $6.8 billion, up 2% year on year. The sequential softness from $6.9 billion to approximately $6.8 billion is not alarming by itself, but the year-on-year deceleration from 19% to 2% is the clearest number that can challenge the “two-engine” thesis if it persists. The profitability, however, remains highly relevant: Kirsten Spears said operating expenses were $1.1 billion in the quarter, resulting in Infrastructure Software operating margin of 78%. With consolidated operating expenses of $2.1 billion, including $1.5 billion of research and development, the software segment is funding a large portion of the AI investment profile while still producing operating leverage. That is a different setup from a pure fabless AI ramp that must absorb every engineering and tape-out burden inside one segment.

Cash flow and capital return make the backlog more investable because they lower the financing risk of the eighteen-month delivery window. Q4 adjusted EBITDA was $12.12 billion, or 68% of revenue, above guidance of 67%, while free cash flow in the quarter was $7.5 billion and represented 41% of revenue. For fiscal 2025, free cash flow grew 39% year on year to $26.9 billion, and the company returned $17.5 billion of cash to shareholders through $11.1 billion of dividends and $6.4 billion in share repurchases and eliminations. Inventory ended the fourth quarter at $2.3 billion, up 4% sequentially. That inventory increase is modest relative to the backlog language, which supports the idea that Broadcom is not yet building balance-sheet risk ahead of unverifiable demand. The better interpretation is that capacity, not finished inventory, is the gating item. For suppliers, that points to sustained demand for wafers, test, IP, and optical components rather than a short-lived shipment spike.

The supply-chain read-through is therefore most constructive for the partners tied to AI ASIC fabrication, networking, and optical attach, rather than for broad distribution. Alphabet (Google), identified as a silicon implementation partner for TPU ASICs at ~$8B/yr, has the clearest customer-side validation from Broadcom’s AI revenue of $20 billion in fiscal 2025 and Q1 AI semiconductor revenue guidance of $8.2 billion. TSMC, as supplier for 3nm/5nm networking plus custom AI ASIC fabrication, is the most direct wafer-capacity beneficiary of a $73 billion AI backlog over the next eighteen months. Keysight benefits from test and measurement exposure as semiconductor revenue is forecast at approximately $12.3 billion in Q1, up 50% year on year. Cadence, Rambus, and CEVA have IP exposure to the custom ASIC design pipeline, while Coherent and Lumentum are tied to optical components for Broadcom CPO and Bailly CPO. GlobalFoundries is more of a specialty-node read-through than an AI leading-edge proxy, and Avnet is a distribution and logistics read-through rather than the main beneficiary of the AI backlog. The magnitude that matters across the chain is not Q4’s +3.1% revenue surprise; it is the $73 billion AI backlog and the AI switch backlog that exceeds $10 billion.

The competitive comparison also argues that Broadcom deserves to be evaluated differently from both hyperscalers and high-growth merchant accelerator peers. In the peer table, NVDA posted revenue YoY of +85.2% with gross margin of 74.9%, which is still the benchmark for merchant AI acceleration economics. Broadcom’s Q4 FY2025 revenue YoY was +28.2% with gross margin of 68.0%, so the print does not say Broadcom is outgrowing the leading AI peer. The comparative point is subtler: Broadcom’s margin is closer to the software-heavy and platform-heavy peer set than to a traditional fabless supplier, while its AI growth is now backed by backlog rather than by generalized capex commentary. Alphabet (Google) in the peer table had revenue YoY of +21.8% and gross margin of 62.4%, while MSFT had revenue YoY of +18.3% and gross margin of 67.6%. Broadcom’s Q4 FY2025 gross margin of 68.0% sits above GOOGL’s 62.4% and near MSFT’s 67.6%, while its +28.2% revenue YoY is higher than both GOOGL’s +21.8% and MSFT’s +18.3%. That combination is why the company can trade as more than a component supplier if it keeps converting AI backlog at current margins.

Management tone supports the backlog thesis, but the call delivery was not uniformly cleaner. The tone history shows Q4 FY2025 sentiment at 0.26, up from Q3 FY2025 at 0.17, and prepared_sentiment at 0.68, up from Q3 FY2025 at 0.35. That fits the print: management had more to say and said it with more confidence in the prepared remarks. The complication is that guidance_tone slipped to 0.35 from Q3 FY2025 at 0.39, while uncertainty rose to 63.5 from 46.1 and qa_evasiveness moved to 2.7 from -17.5. The numbers conflict in a useful way. Prepared remarks were much more positive, but Q&A became more uncertain and more evasive. That pattern is consistent with a company willing to commit to backlog and near-term revenue, while avoiding too much granularity on customer mix, timing, or margin by program. Investors should treat the backlog as real, but should not assume management will give clean customer-level attribution.

That tone pattern becomes more constructive when viewed against later call history, because Q1 FY2026 and Q2 FY2026 show the uncertainty problem easing rather than worsening. Q1 FY2026 guidance_tone recovered to 0.52 from Q4 FY2025 at 0.35, tone_confidence improved to 0.52 from 0.45, and qa_evasiveness fell to -31.0 from 2.7. Q2 FY2026 then showed guidance_tone of 0.54, tone_confidence of 0.55, uncertainty of 43.1, and qa_evasiveness of -40.6. The call-over-call delta from Q1 FY2026 to Q2 FY2026 shows uncertainty down -18.4 and qa_evasiveness down -9.6, even as sentiment moved -0.02 and guidance_tone rose +0.02. This matters because it says the Q4 FY2025 evasiveness was more likely a function of sensitive backlog detail than a sign that management lacked visibility. If the next quarter keeps guidance_tone near 0.52 or 0.54 while uncertainty stays closer to 43.1 than 63.5, the market should become more comfortable underwriting the eighteen-month AI delivery schedule.

The risk to the thesis is not that Q4 was backward-looking; it is that investors may over-extrapolate the Q1 AI guide while ignoring the gross-margin and software deceleration markers. The clean confirmation path is specific. For Q1 FY2026, management has guided consolidated revenue to approximately $19.1 billion, semiconductor revenue to approximately $12.3 billion, AI semiconductor revenue to $8.2 billion, non-AI semiconductor revenue to approximately $4.1 billion, and infrastructure software revenue to approximately $6.8 billion. The quarterly history lists Q1 FY2026 revenue at $19,311.0 million and gross margin at 65.6%. Confirmation would be revenue at or above approximately $19.1 billion, AI semiconductor revenue at or above $8.2 billion, and evidence that gross margin can recover from 65.6% toward the 67.2% shown for Q2 FY2026 rather than settling below the Q4 FY2025 level of 68.0%. The thesis breaks if AI semiconductor revenue misses $8.2 billion, if non-AI weakness falls below approximately $4.1 billion and starts to contaminate consolidated growth, or if infrastructure software fails to hold approximately $6.8 billion after Q4’s $6.9 billion. The next date that matters is the Q1 FY2026 report, because that is where the market will learn whether the $73 billion AI backlog is translating into shipped revenue without sacrificing the margin structure that made this quarter investable.