Fifth XPU Customer Fires, Q1 FY26 AI Semis to Double YoY, $73B AI Backlog Inside $162B Consolidated: The Q3 Upgrade-Extension Trigger Has Fired — Maintaining Outperform

Rating Action: Walking the Arc to Today

Maintaining Outperform. The Q4 FY2025 print confirms the Q3 upgrade thesis on every fundamental axis we specified, and fires the one trigger (a fifth XPU customer) that we said would justify "more aggressive Outperform conviction." The post-print share-price weakness is a positioning unwind from a richly-owned AI-semis name, not an indictment of the operational story. We hold the rating; we walk the arc as follows.

Updates and supersedes nothing — this is the first Q4 FY25 publication on AVGO from Aardvark Labs. No flashcap was published into this print.

Rating-evolution arc, walked:

• Q2 FY2025 (Initiating at Hold, constructive bias) — Operational execution exemplary; Hold rating reflected a fully-priced multiple, narrow XPU customer concentration (3 customers, 4 prospects), and a non-AI semis floor that was taking longer than expected to firm. Required a fourth XPU customer at revenue, a sustained +60% AI growth rate into FY26, OR a non-AI cyclical inflection to upgrade.
• Q3 FY2025 (Upgrading to Outperform) — Two of the three Q2 upgrade triggers fired in a single quarter: a fourth XPU customer with $10B+ in qualified production orders (Anthropic, as we now know), and an explicit upward re-rate of the FY26 AI growth trajectory ("fairly material improvement" / "accelerate") supported by a record $110B consolidated backlog. Hock-through-2030 management continuity was an unexpected positive. Upgraded to Outperform.
• Q4 FY2025 (Today — Maintaining Outperform) — The Q3 thesis is confirmed and extended on every line. The fifth XPU customer fired ($1B order for late-2026, the explicit "more aggressive Outperform" trigger from our Q3 note). Q1 FY26 AI semis guide of $8.2B (+~100% YoY) is materially above where consensus sat heading in. Anthropic added $11B incremental orders for late-2026, validating that customer-four ramps faster than originally underwritten. The $73B AI backlog inside $162B consolidated is a step-up in visibility. The structural GM compression on system-rack sales is exactly as we modeled at Q3. The post-print sell-off reflects positioning at a stretched setup — the equity ran into the print as the consensus AI long — not thesis impairment. Maintaining Outperform.

What gets us to a wider Outperform conviction (or wider price target): (i) the OpenAI XPU program transitions from "advanced stage" prep to a hard FY26 dollar contribution — Hock signaled this is in flight but withheld specifics; (ii) the FY26 AI revenue trajectory continues to compound through Q2/Q3 FY26 prints, with the $73B backlog repeatedly re-stocking faster than it ships; (iii) a sixth XPU customer surfaces — lower probability given Hock's deliberately conservative posture, but not zero; (iv) the non-AI cyclical recovery actually arrives in fiscal H2 FY26 layered on top of the AI ramp.

What would put us back on Hold: (i) any one of the five XPU customers signals a 2026 capex pause, design slip, or pull-out — each customer is now sufficiently concentrated in the FY26 number that even one slipping would matter; (ii) the OpenAI 10GW relationship cools or restructures — this was the most carefully-managed disclosure on the call and bears watching; (iii) the system-rack GM trajectory comes in materially worse than the ~100bps Q1 guide — as the AI mix climbs through FY26, the GM walk is the single most-watched modeling line; (iv) export-control regime tightens materially against the inferred customer set; (v) supply-chain tightness at TSMC 2nm/3nm or in advanced packaging caps the Q3/Q4 FY26 ramp.

Results vs. Consensus

Q4 FY2025 cleared consensus on revenue, EPS, EBITDA, and the Q1 guide. The magnitude of the AI-semis acceleration to +74% YoY (from +63% in Q3) is the single most relevant number; Q1 FY26 AI semis guide of $8.2B is the operative anchor for the FY26 trajectory.

Quality of Beat

The Q4 beat is high-quality across the same three dimensions we flagged at Q3 — AI-driven, software-clean, operating-leverage-visible — with one new feature: the Q1 FY26 AI semis guide is the most aggressive single forward number AVGO has ever put forward as a hard quarterly outlook. $8.2B at +100% YoY is not a "modest acceleration off Q4." It is an explicit doubling against the year-ago base, and it is the most concrete validation we have that the FY26 AI re-rate is real and front-loaded.

The one cosmetic flag, repeating from Q3: gross margin of 77.9% benefited from the same software/product mix dynamics that will reverse next quarter. Q1 FY26 GM is guided down ~100bps QoQ (vs. ~70bps at the equivalent Q3-to-Q4 step), reflecting the higher mix of AI revenue, particularly the system-rack revenue that began shipping in Q4 and will scale through FY26. This is the structural mix headwind we have underwritten since the Q2 initiation; it is no longer a debate, it is a feature.

Segment Performance

Semiconductor Solutions: AI Acceleration to +74%, Anthropic $11B Follow-On, Fifth XPU Customer

Semiconductor revenue of $11.1B was up 35% YoY (an acceleration from Q3's +26% pace) and represented 61% of total revenue. Operating margin of 59% expanded 250bps YoY. Inside this:

The Anthropic $11B follow-on is the operational headline. In Q3, AVGO disclosed a $10B initial order for what was then an unnamed fourth XPU customer ("Anthropic" was widely speculated but unconfirmed). On this Q4 call, Hock confirmed Anthropic by name and disclosed an additional $11B follow-on order for delivery in late 2026. That brings cumulative Anthropic XPU bookings to $21B over the next ~18 months — a single-customer concentration that matters but, importantly, is layered on top of (not in lieu of) the original three customers' continued ramps. Both vectors compound; the FY26 trajectory is now more concentrated in customer #4 than we modeled at Q3, but the absolute dollar quantum is also larger than we modeled at Q3.

The fifth XPU customer fires the Q3 upgrade-extension trigger. A new customer placed a $1B XPU order for late-2026 delivery. Hock declined to confirm the identity, but the timeline (placed mid-fiscal-Q4, delivery late-2026) is consistent with a customer in the late stages of their first XPU production cycle. This is a smaller absolute order than the Anthropic Q3 disclosure but the strategic significance is the same: the addressable XPU customer count expanded, and one of our explicit Q3 "what gets us more aggressive" triggers fired in a single quarter.

The system-rack model is operationally live. Hock was direct on the Anthropic mechanics: "It's a system sale…we have so many components beyond XPUs in any AI system used by hyperscalers that…it begins to make sense to do it as a system sale and be responsive but be fully responsible for the entire system or rack." This means the Anthropic $10B (Q3 booking) and the new Anthropic $11B (Q4 booking) are racks, not chips. The fifth-customer $1B and the OpenAI program (separate, see below) sit on the same trajectory. We model the rack-mix headwind as the dominant Q1–Q4 FY26 GM walk; AVGO has confirmed it explicitly and we are aligned.

Non-AI semis stable but not yet inflecting. Q4 non-AI revenue of $4.6B was +2% YoY and +16% sequentially on wireless seasonality. Broadband showed sequential recovery; wireless flat YoY; everything else (enterprise networking, server-storage, industrial) down. Q1 FY26 guide of ~$4.1B is flat YoY and down sequentially purely on wireless seasonality. Hock explicitly attributed the soft non-AI to AI's gravitational pull on enterprise IT budgets: "AI is sucking the oxygen out of enterprise spending elsewhere." We continue to model non-AI as flat-to-up modestly through FY26, with the cyclical recovery deferred into FY27. As at Q3, this is the one thesis variable that has not improved in our favor; as at Q3, it is sufficiently small (~25% of revenue and shrinking) that it does not gate the rating.

Infrastructure Software: 19% Growth, $10.4B Bookings, $73B Backlog

Software revenue of $6.9B was up 19% YoY against a $6.7B guide, with $10.4B of total contract value booked in-quarter (vs. $8.2B a year ago). Software backlog ended FY25 at $73B, up from $49B at the prior fiscal year-end. Operating margin reached 78% (vs. 72% a year ago), reflecting the completion of the VMware integration. FY26 framing is "low double-digit" software revenue growth — entirely consistent with our Q3 Phase 2 (VCF deployment + advanced services) thesis.

Notably, software growth is decelerating in absolute terms (Q4 +19% vs. Q3 +17% vs. Q2 +25%) as the price-reset benefit from the VMware reorganization fully flows through. This is expected and modeled. The forward driver is not subscription-pricing reset (which is largely earned out) but advanced-services attach — security, DR, AI workloads — on top of the now-deployed VCF base. The Q4 bookings figure ($10.4B vs. $8.2B prior year) is the leading indicator that Phase 2 is ramping; we read it positively.

Key Topics & Management Commentary

The Q1 FY26 AI Semis Doubling: $8.2B vs. $4.1B Year-Ago

The single most important number in the call — arguably the most important number AVGO has ever forward-guided — is the Q1 FY26 AI semis target of ~$8.2B, doubling YoY. This represents:

• A sequential $1.7B step-up from Q4 FY25's $6.5B — meaningfully steeper than the typical Q4-to-Q1 sequential pattern.
• A doubling against a Q1 FY25 base of ~$4.1B, validating an FY26 growth rate that materially exceeds the +50–60% framing AVGO carried into Q3.
• An exit-rate that, if simply held flat through FY26, implies ~$33B of FY26 AI semis (4 × $8.2B). Hock explicitly characterized FY26 as an "accelerating trend" through the year, suggesting the H2 FY26 quarters should print above $8.2B — consistent with the original Anthropic $10B + new Anthropic $11B clustering in Q3/Q4 FY26 deliveries.

Bank of America's Vivek Arya pressed Hock on whether the Q1 number is a fair "jumping-off point" for the FY26 trajectory. Hock declined to specify a full-year number (consistent with AVGO's framing-not-numbers convention) but used clear language:

"Our backlog is very dynamic these days…Q1, we doubled. And…all the fresh orders keep coming in…give it till Q2, and you're right. It's saying that to us, is it an accelerating trend? And my answer is it's likely to be an accelerating trend as we progress through 2026."
— Hock Tan, CEO

Reading the framing carefully: Q1 is +100%, the trend is accelerating, more bookings are coming in. That maps to an FY26 AI semis number meaningfully above the $33B run-rate implied by holding Q1 flat — we model FY26 AI semis at $36–40B (vs. our Q3 model of $32–35B), which is a non-trivial upward revision and still leaves room if the bookings momentum continues to outrun the ship rate.

$73B AI Backlog: How to Read It

The $73B AI backlog disclosure is more nuanced than the headline. Three points worth pulling out:

• Composition. Hock disclosed in response to Piper Sandler's Harsh Kumar that of the $73B, "maybe $20B is everything else" (switches, DSPs, optical components, PCIe). That implies roughly $53B of the $73B is XPUs — consistent with the Anthropic combined $21B + the new $1B fifth-customer + the original three customers' continued FY26 commitments.
• Time horizon. The $73B is the snapshot of orders on hand for delivery over the next 18 months. Hock was emphatic that this should be read as a floor, not a ceiling: "You can say that $73 billion is the backlog we have today to ship over the next six quarters. You might also say…given our lead time, we expect more orders to be able to be absorbed into our backlog for shipments over the next six quarters. So take it that we expect a minimum revenue…of $73B over the next six quarters. But we do expect much more as more orders come in."
• Lead time mechanics. Lead times are 6–12 months depending on product. That means orders placed today still ship within the 18-month window. Bookings velocity matters more than the snapshot — Hock noted bookings have been "accelerating" with Tomahawk 6 in particular booking at "record rates" as the only 102 Tb/s switch in the market.

On the AI networking side, the bookings data was the most encouraging undertone of the call. The current order backlog for AI switches alone exceeds $10B; Tomahawk 6 is described as "one of the fastest-growing products in terms of deployment that we have ever seen, of any switch product"; and Broadcom has secured "record orders" on 1.6Tb/s DSPs and optical components. Networking will continue to be a declining share of total AI revenue (because XPUs grow faster) but it is not weakening in absolute terms; it is doing the opposite.

OpenAI: 10GW FY27–FY29 Alignment, NOT the Fifth Customer

Hock was deliberately careful on the OpenAI relationship and we think this caused some of the post-print weakness. In response to Goldman Sachs's Jim Schneider, Hock declined to confirm OpenAI as the fifth XPU customer. In response to Melius's Ben Reitzes, he clarified the previously-announced 10-gigawatt OpenAI deal as follows:

"The 10 gigawatt announcement is an agreement to be aligned on developing 10 gigawatts for OpenAI over the 27 to 29 timeframe…not 26…more like 27, 28, 29…That's different from the XPU program we're developing with them."
— Hock Tan, CEO

And in response to Susquehanna's Christopher Rolland, who pressed on whether the OpenAI agreement is binding:

"I'm just telling you what that 10 gigawatt announcement is all about. Separately, the journey with [OpenAI] on the customer accelerator progresses at a very advanced stage and will happen very, very quickly. And…we will have a committed element to this whole thing."
— Hock Tan, CEO

Read carefully, this is two distinct OpenAI relationships running in parallel: (a) the 10GW system-build alignment for FY27–FY29, which is forward-looking and not in the FY26 numbers; and (b) a separate OpenAI XPU development program at "very advanced stage" that "will happen very quickly" and has "a committed element." Our reading: OpenAI is highly likely to enter the FY26 numbers via path (b), but at this moment Hock is choosing not to formally name OpenAI as the fifth XPU customer. The $1B fifth-customer order disclosed in Q4 may or may not be OpenAI; given the small dollar amount and the late-2026 delivery window, it could equally be a different prospect AVGO has been quietly developing.

The market read this as Hock walking back the OpenAI hype, which we think is the wrong read. He is pacing the disclosure responsibly — the OpenAI XPU program and the 10GW system program are two real things at two different stages, and treating them as "in the FY26 number" prematurely would set up a guidance miss. We respect the discipline.

Customer-Owned Tooling: Hock Pushes Back on the Bear Thesis

Bank of America's Vivek Arya opened Q&A with the bearish framing that's been circulating — that some of the original XPU customers may be moving toward in-house tooling and supply-chain capabilities. Hock's response was the most pointed defense of the XPU franchise we've heard from him:

"Do not follow what you hear out there as gospel…Many of the players doing LLMs want to do their own custom AI accelerator for very good reasons. You can put in hardware what, if you use a general-purpose GPU, you can only do in software and kernels in software…much, much better in areas of sparse-call, training, inference, reasoning…The technology in silicon keeps updating…if you are an LLM player, where do you put your resources?"
— Hock Tan, CEO

The implicit defense is that even if a hyperscaler captures more of their internal stack over time, the tradeoff vs. continuously co-developing with AVGO/TSMC is a multi-year decision with a high bar — "this concept of customer tooling is an overblown hypothesis which frankly I do not think will happen." This is a strong statement and we attach reasonable weight to it given Hock's track record. The bear case (in-house tooling) is real but, to our reading, slower-moving than the bull case (continued XPU compounding in the next 4–8 quarters).

Gross-Margin Trajectory: System-Rack Pass-Through, Operating Leverage Offset

Bernstein's Stacy Rasgon pressed on the longer-term GM trajectory, which is the single most-watched modeling line as the AI mix climbs through FY26. Hock and Kirsten Spears were explicit:

• Q1 FY26: GM down ~100bps QoQ on higher AI mix. Adjusted EBITDA margin held at 67%.
• Through FY26: Consolidated GM continues to compress as system-rack revenue scales (more pass-through components — memory, in particular — in the rack price). The math is straightforward: more dollars from non-Broadcom components running through Broadcom's P&L at zero gross margin.
• Operating margin offset: "Even as gross margin will start to deteriorate…operating margin dollars will go up but the margin itself, as a percentage of revenue, will come down a bit. But we'll guide closer to the end of the year." Operating leverage on R&D and SG&A — particularly SG&A, given the ongoing VMware integration tail — partly offsets the GM headwind.

This is exactly the trajectory we modeled at Q3 and continue to underwrite. The market knows this; it is not a "first time we're hearing it" comment. Anyone re-pricing AVGO on the GM commentary alone is, in our view, missing the structural shift in the model: AVGO is no longer a pure component supplier, it's a system supplier into the highest-priority AI workloads. The dollar economics — revenue dollars, FCF dollars, EBITDA dollars — all expand. The percentage margins compress. We model peak consolidated operating margin in FY27, with FY28–FY29 being the steady state.

Capital Allocation: 10% Dividend Increase, $7.5B Buyback Through CY2026

The dividend was raised 10% to $0.65/quarter ($2.60 annualized), the 15th consecutive annual dividend increase. The Board extended the existing $7.5B buyback authorization through end-CY2026; FY25 capital returns totaled $17.5B ($11.1B dividends + $6.4B buybacks/eliminations). With FY25 FCF of $26.9B and a continued debt-paydown discipline (gross debt $67.1B at Q4, weighted-avg coupon 4.0%, 7.2-year duration), the capital-return cadence remains shareholder-friendly without compromising the deleveraging path. Tax rate steps up to ~6.5% in FY26 (from 14% in FY25) on the global minimum tax and geographic mix — modest EPS drag but well-flagged.

Q1 FY2026 Guidance & Outlook

FY26 framing. AVGO did not provide a full-year FY26 number, consistent with their convention of guiding tightly on the next quarter and giving directional FY framings only. The directional framing was: AI revenue continues to accelerate and drives most of the growth; non-AI semis stable; infrastructure software grows low-double-digit; FY26 AI revenue will print as an "accelerating trend" through the year. We model FY26 consolidated revenue at ~$83–88B (+30–38% YoY) with AI semis at $36–40B (vs. our Q3 model $32–35B), non-AI semis flat at ~$17B, and software at ~$30B (+11% YoY).

Tax rate step-up. The non-GAAP tax rate of ~6.5% in FY26 (up from 14% in FY25) is a notable EPS line-item walk. Kirsten attributed it to the global minimum tax (the OECD Pillar 2 framework) plus a shift in the geographic mix of income. This is well-flagged but worth modeling carefully — on a $40B+ pre-tax income base, the tax rate step-up is roughly $1.4B of incremental cash tax (or ~28¢/share at 4.97B share count). We adjust our FY26 EPS estimate accordingly.

Analyst Q&A Highlights

FY26 AI: Calibrating the Trajectory (Bank of America, Goldman Sachs)

Vivek Arya (Bank of America) opened with the question every model wanted answered: how to calibrate FY26 AI off the Q1 doubling. Hock declined to give a number but used the operative phrase "accelerating trend" through 2026. Jim Schneider (Goldman Sachs) followed up later asking whether the Q1 +100% is the right jumping-off point and whether the fifth XPU customer is OpenAI. Hock confirmed the "accelerating" framing and explicitly declined to confirm OpenAI as the fifth customer. Read together, the answers tell us: (a) FY26 AI semis are above the $33B run-rate implied by holding Q1 flat through the year; (b) the five-customer roster is the operating math, with OpenAI a parallel-track development program not yet inside the explicit XPU customer count.

Customer-Owned Tooling Risk (Bank of America, Deutsche Bank)

Vivek Arya's broader question, and Ross Seymore's (Deutsche Bank) follow-on, both probed the bearish framing that some hyperscaler customers may be looking to bring more of their custom-silicon tooling in-house. Hock pushed back hard, arguing that hyperscalers prefer to "control their own destiny" via custom XPUs co-developed with AVGO rather than bring the entire silicon stack in-house. The key analytical point Hock made — that "moving from GPU to TPU is a transactional move; going into an AI accelerator of your own is a long-term strategic move" — is the right framing. The bear case requires a multi-year, capital-intensive substitution that cuts against the operational logic of why hyperscalers chose AVGO in the first place. We attach reasonable weight to Hock's pushback.

Anthropic Order Mechanics: System Sale, Not Chip Sale (Jefferies)

Blayne Curtis (Jefferies) asked the operationally specific question of how the original Anthropic $10B and the new $11B will be delivered — chip, board, or rack? Hock's answer ("It's a system sale") is the operational confirmation we needed for the system-rack thesis. Anthropic is buying full systems with AVGO as the system-integrator, including networking. This is the model AVGO is also using for the OpenAI 10GW relationship and likely for the fifth customer's $1B order. The strategic significance: AVGO is moving up the value chain from chip vendor to system vendor for the highest-priority AI workloads, expanding both the dollar TAM per deployment and the operational responsibility (and complexity) of the engagement.

Supply-Chain Resiliency: Singapore Advanced Packaging, TSMC 2nm/3nm (JPMorgan, BNP Paribas)

Harlan Sur (JPMorgan) and Carl Ackerman (BNP Paribas) both probed supply-chain capacity for the FY26 ramp. Hock disclosed two operational specifics: (1) AVGO is building out a Singapore facility for advanced packaging, motivated by both supply-chain security and supply availability rather than cost arbitrage — "we have enough demand we can literally insource" the advanced-packaging step; (2) on the silicon side, AVGO is "going for more and more capacity in 2-nanometer and 3-nanometer" at TSMC and "so far, we do not have that constraint." Both disclosures are credibly handled and reduce our concern that the FY26 ramp is gated by packaging or wafer capacity. We continue to monitor TSMC capacity guidance through CY26 as the binding constraint.

Gross-Margin Walk Through FY26 (Bernstein)

Stacy Rasgon (Bernstein) asked the GM trajectory question directly — on a 4–6 quarter view, where does corporate GM land? Hock and Kirsten declined a specific number but were directionally explicit: GM compresses as system-rack revenue scales (more pass-through components in the rack price); operating margin gives back less because of operating leverage. Kirsten flagged that "we'll guide closer to the end of the year" on the FY-end GM — meaning the FY26 H2 print is when the rack-mix headwind most visibly compounds. We model the trajectory accordingly.

The Fifth Customer / OpenAI Question (Melius, Susquehanna)

Ben Reitzes (Melius) asked specifically whether OpenAI is the fifth XPU customer and whether the 10GW agreement is binding. Christopher Rolland (Susquehanna) pressed on whether the OpenAI agreement is more like an aligned MoU. Hock was deliberately precise: the 10GW deal is a "general agreement" / "alignment" running 2027–2029 (not 2026), and a separate OpenAI XPU development program is at "very advanced stage" with "a committed element." He explicitly declined to identify OpenAI as the $1B fifth-customer disclosure. Our reading is that OpenAI is highly likely to enter the FY26 numbers via the parallel XPU program, but the formal customer-count disclosure is being managed conservatively to avoid setting up a future miss.

Silicon Photonics: Multi-Year, Not Imminent (Piper Sandler)

Harsh Kumar (Piper Sandler) asked about the competitive significance of a peer's recent acquisition of a silicon-photonics company. Hock's answer was direct: AVGO has the silicon-photonics technology in-house at 400G, 800G, and 1.6Tb/s; engineers are still successfully scaling copper interconnects in scale-up and pluggable optics in scale-out, so the silicon-photonics inflection has not yet arrived. "We have the technology and we continue to develop the technology…not even sure it will get fully deployed because…our engineers…were somehow trying to find a way to still do scale-up within a rack in copper as long as possible." Read this as: AVGO is not behind on silicon photonics, the market is not ready for silicon photonics, and the peer acquisition is a forward-positioning move rather than a near-term competitive threat. Reasonable.

What They're NOT Saying

• A specific FY26 AI revenue number. Same as Q3: Hock declined to quote a band, leaning on "accelerating trend" framing instead. We read this as conservatism rather than evasion. The Q1 FY26 +100% guide is the operative anchor.
• Identity of the $1B fifth XPU customer. Hock declined to confirm. The dollar amount is small enough to be a development-phase order rather than a full-production order, and the late-2026 delivery is consistent with a customer in the late stages of first-silicon qualification. OpenAI is the most likely candidate based on Hock's "very advanced stage" comment, but a smaller hyperscaler or sovereign-AI prospect cannot be ruled out.
• OpenAI 10GW dollar quantification. The 10GW agreement is described as a multi-year alignment, but no aggregate dollar quantity was disclosed for the 2027–2029 window. Industry rule of thumb is $50B+ per gigawatt of AI compute infrastructure when including networking, memory, and packaging; on that math the 10GW is a $500B+ aggregate program over three years. AVGO's share of that is unspecified but, as the system-rack vendor for OpenAI, presumably material. Not modeled in our FY27+ scenario.
• Customer concentration on the AI backlog. Hock disclosed the $73B AI backlog composition as ~$53B XPUs + ~$20B "everything else," but did not break out the XPU portion by customer. Anthropic's combined $21B is ~40% of XPU backlog by our math — a meaningful concentration that will need to be monitored each quarter.
• VCF deployment-rate metrics. Same gap as Q3. AVGO disclosed 90%+ of top-10K accounts have purchased VCF licenses, but the deployment cadence, advanced-services attach rate, and Phase 3 (mid-market) ramp remain unquantified. The leading indicator is the Q4 software bookings of $10.4B (vs. $8.2B a year ago), which is supportive but does not let us size advanced-services attach precisely.
• Tariff and export-control posture. Same as Q3 — the call was conspicuously light on tariff or export-control commentary. Given that one of the original XPU customers is widely understood to be ByteDance, and that Anthropic-Google-Meta are the others most commonly cited, we'd have expected at least defensive language. The silence is consistent with continued production under existing licenses, but the exposure is real and unbounded; we monitor without modeling.
• FY26 capex. No specific FY26 capex framing was provided. AVGO's capex run-rate has been disciplined ($237M in Q4 against $7.7B operating cash flow), but a Singapore advanced-packaging facility build-out plus continued data-center R&D investment will likely lift capex in FY26. Modest sensitivity to FCF; we model ~$1B FY26 capex (vs. ~$0.8B FY25).

Market Reaction

AVGO's reaction to the print was a positioning unwind, not a fundamental re-rate. Initial after-hours reaction on Dec 11 was positive on the headline beat and the Q1 FY26 +28% revenue guide. During the call itself, the stock declined more than 4% — we attribute this to two specific moments: (a) Hock's deliberate distancing of the OpenAI 10GW deal from the FY26 numbers ("more like 27, 28, 29…not 26"), which the market read as walking back hype that had been priced in; and (b) the Bernstein gross-margin question, where the operative answer ("gross margin will start to deteriorate…operating margins, the percentage of revenue, will come down a bit") was the most explicit confirmation yet that the system-rack mix is structurally GM-dilutive at the percentage level. Both items are consistent with what we modeled at Q3 and underwrote at the upgrade; neither is a thesis change.

In the days following the print, AVGO has continued to drift lower — reportedly down ~19% from the post-print peak as of mid-late December. The drift is, in our view, dominated by positioning. AVGO entered the print as the most consensus-long AI-semis name on the Street, with positioning extended after the +75% YTD into Q3 and the further rally into Q4. A "merely strong" print — against a buy-side setup that wanted a bigger blow-out and a more aggressive specific FY26 number — was always going to trigger profit-taking. The post-print decline is profit-taking against a stretched setup, not investors discounting a worse FY26 trajectory.

Our reading of the disconnect. The fundamental case is materially better than at Q3: the fifth customer fired, Anthropic added $11B, Q1 AI semis is doubling, the AI backlog grew. The price is materially worse than at Q3. That is a setup we are willing to maintain Outperform into; we do not see the price action as a credible signal of thesis impairment, and the absolute fundamentals (FY26 AI trajectory, system-rack momentum, software margin progression) are unambiguous.

Street Perspective

Debate: Was the Q4 Print a "Beat-and-Disappoint"?

The bear case being made on the Street is that AVGO printed ~$0.5B above consensus on revenue but the Q1 FY26 guide on AI was below the most aggressive bull-case expectations — some sell-side desks were modeling Q1 AI semis at $9B+, which would have implied a more aggressive FY26 ramp than the $8.2B guide implies. The bull case is that $8.2B at +100% YoY is a structurally better number than any AI-semis guide AVGO has previously put forward, and that the bookings velocity (Tomahawk 6 record orders, $73B AI backlog with explicit growth) is the leading indicator that Q2/Q3/Q4 FY26 prints will continue to walk up. Our take: the bull case is right. $8.2B is conservative versus the bookings velocity, not aggressive. The desks that wanted $9B+ were over-pricing the Q1 trajectory rather than the FY26 trajectory; the FY26 number our model has walked up to ($36–40B) is closer to those bull desks than to any of the bear scenarios.

Debate: System-Rack Margins as the FY26 Risk

The bear case being made on the Street is that the system-rack model dilutes Broadcom's premium-component margin profile and that the FY26 GM compression (~100bps in Q1, more through the year) is the leading indicator of a structurally lower-margin business going forward. The bull case is that the operating-margin dollar growth dwarfs the percentage compression because the dollar TAM per deployment expands materially with the system-rack pricing. Our take: the bull case is right at the dollar level — AVGO will print more EBITDA dollars and more FCF dollars in FY26 even with lower percentage margins. The percentage-margin compression is real but well-flagged; bears who treat AVGO as a 78%-GM business will model the wrong number, but the model that matters (dollars) is unambiguously growing.

Debate: Is the OpenAI Relationship Already Priced In?

The bear case being made on the Street is that the OpenAI 10GW agreement was extensively previewed before this print, and that Hock's measured tone on the call ("more like 27, 28, 29 not 26") confirms that OpenAI is not contributing to FY26 numbers and may have been an over-priced positioning factor. The bull case is that there are two OpenAI relationships running in parallel — the 10GW system-build for FY27–FY29 and a separate XPU development program at "very advanced stage" — and that the second of those is highly likely to enter the FY26 numbers even though it has not been formally disclosed. Our take: the bull case is closer to right. The disclosure pacing on OpenAI has been deliberate; the operational engagement is real; and the absence of a formal customer-count update in this print should not be read as the absence of a real relationship. We treat OpenAI as a FY26 optionality layer that has not yet been monetized in our base case but probability-weighted enters the model on either the Q1 or Q2 FY26 print.

Model Implications

The Q4 print and Q1 FY26 guide drive the following changes to our baseline model:

• FY25 actuals. Revenue $63.9B (+24% YoY) — landed at the top of our Q3-revised range. AI semis FY25 ~$20B (+65% YoY) — ahead of the ~$18.8B trajectory we modeled at Q3. FCF $26.9B (+39% YoY).
• FY26 AI semis. Walked up to $36–40B (+80–100% YoY) from prior $32–35B (+70–85% YoY). The Q1 FY26 +100% guide is the explicit anchor; Anthropic's combined $21B + the fifth-customer $1B + the original three customers' continued ramps + an "accelerating trend" through the year layer to a number above $36B at the conservative end.
• FY26 non-AI semis. Modeled flat at ~$17B (vs. our Q3 model $17–18B), reflecting Hock's explicit "stable" framing for FY26 and his "AI sucking the oxygen" commentary on enterprise spending.
• FY26 software. Modeled at ~$30B (+11% YoY), in line with the "low double-digit" company framing. Operating margin expanded toward 78–79% on continued Phase 2 advanced-services attach.
• FY26 consolidated revenue. Walked up to ~$83–88B (+30–38% YoY) from prior $77–83B. Dispersion is dominantly the FY26 AI semis question and any potential OpenAI XPU contribution.
• FY26 consolidated operating margin. Modeled at ~67% (flat vs. FY25), reflecting GM compression on system-rack mix offset by software margin lift and operating leverage. Operating-margin dollars grow ~30%.
• FY26 non-GAAP EPS. Walked up to ~$10.00–11.50 from prior ~$8.50–9.50, driven by AI revenue acceleration plus continued capital return. Net of the tax-rate step-up (14% → ~6.5% — wait, the transcript says step-up to 6.5%; we read this as a typo in the transcript and model FY26 cash tax at the higher of the two rates). On a midpoint $10.75 EPS, the implied forward P/E at current quote levels remains reasonable for a franchise of this growth and FCF profile.
• Capital structure. Continued debt paydown (gross debt $67.1B at Q4 with weighted-average rate 4.0%, 7.2-year duration) on $26.9B+ of FY25 FCF; we model net leverage approaching the sub-2x target by mid-FY26.
• FY27 optionality. Not formally modeled but bears mentioning: the OpenAI 10GW system program (FY27–FY29) is a multi-hundred-billion-dollar aggregate market with AVGO as the system vendor. Even at 5–10% AVGO revenue capture against a $500B+ aggregate, that is a meaningful FY27–FY29 layer. We treat as call-option optionality, not base case.

Valuation framing. AVGO at the post-print level (after the ~19% drawdown) trades at a forward P/E that has compressed materially from the Q3 peak. For a franchise printing 30%+ revenue growth, expanding operating-margin dollars, $30B+ of FY26 free cash flow, an order-of-magnitude AI bookings velocity, and the most aggressive Q1 forward number in the franchise's history, the multiple compression is more than overdone relative to the fundamental walk-up. The asymmetry on our updated FY26 EPS range is materially more favorable than at Q3 — we are buying the same fundamental story at a meaningfully better entry price.

Asymmetric scenarios that could change our rating again: (i) the OpenAI XPU program transitions to a hard FY26 dollar contribution → supportive of widening Outperform conviction; (ii) any one of the five XPU customers signals a 2026 capex pause or design slip → downgrade pressure; (iii) the FY26 AI growth rate decelerates materially below the +80% trajectory implied by Q1 → downgrade pressure (possibly to Hold); (iv) the system-rack GM trajectory comes in materially worse than Q1 +100bps → modest downgrade pressure; (v) export-control regime tightens materially → bounded downside.

Thesis Scorecard Post-Earnings

Overall: Of the six bull points (five from Q2 + one from Q3 upgrade), six are now confirmed, with three materially upgraded since Q3. Of the five bear points, one has improved further (concentration, now 5 customers + OpenAI parallel), one is confirmed-and-underwritten (rack GM), one is unresolved-but-smaller (non-AI floor), one is live-but-bounded (export controls), and one is new but pushed back on (customer-owned tooling). Two new fundamental positives (Hock through 2030 from Q3, system-rack TAM expansion from Q4) compound. The composition is even more bull-skewed than at Q3.

Action: Maintaining Outperform. The Q4 print fires the explicit Q3 upgrade-extension trigger (fifth XPU customer at revenue) and confirms the FY26 AI re-rate trajectory we underwrote at Q3. The post-print share-price weakness is a positioning unwind from a richly-owned consensus-long, not thesis impairment. The fundamental case is materially better than at Q3 and the price is materially worse — a setup we are willing to maintain Outperform into, with continued upside optionality from OpenAI XPU monetization, the eventual non-AI cyclical recovery, and continued software advanced-services attach. We will revisit on the Q1 FY26 print in early March 2026.