Fourth XPU Customer With $10B+ in Orders, FY26 AI Growth Re-Rated Higher, $110B Consolidated Backlog: The Q2 Hold-Gating Concerns Have Been Retired — Upgrading to Outperform

Rating Action: Walking the Q2 Initiation to Today

We initiated coverage on Broadcom at Hold (constructive bias) in our Q2 FY2025 note. The Hold framing was a deliberate price-discipline call rather than an execution critique: Q2 was already a textbook print, the VMware integration was already converting acquired SG&A into structural margin, and AI semis were already compounding in the +50–60% YoY range. But the equity at Q2 had already discounted a multi-year continuation of the AI compounder at a healthy multiple, the XPU customer base was narrow (three customers, with four prospects engaged but not yet revenue), and the non-AI semis floor was taking longer to firm than we would have expected at that stage of the cycle.

We specified four explicit upgrade triggers in the Q2 initiation. Walking them to where Q3 leaves us:

• Q2 trigger (a): "A fourth XPU customer transitions from prospect to revenue with FY26 contribution." Q3 outcome: fired explicitly. A prospect — one that Hock described last quarter as engaged but not yet a customer — released production orders during Q3 and has secured "over $10 billion of orders of AI racks based on our XPUs," with the delivery window clustered in fiscal Q3 FY26. The customer is now characterized as "qualified" alongside the original three. This is not a probabilistic trigger any longer; it is in the order book, with timing.
• Q2 trigger (b): "Q3/Q4 prints validate the +60% AI growth rate sustaining into early FY26 with no single-customer concentration anomalies." Q3 outcome: fired explicitly, and stronger than the original framing. AI semis grew +63% YoY in Q3 (above the +60% framing), Q4 guide is +66% YoY, and management explicitly re-rated the FY26 trajectory upward: where last quarter Hock had said the FY25 growth rate (+50–60%) would "mirror" into FY26, this quarter he said FY26 will be "a fairly material improvement" with the growth rate accelerating. The three original customers continue to gain share inside their own compute footprints (Hock: as software stacks stabilize, customers "have the confidence to keep using a higher and higher percentage of their compute footprint in their own XPUs"). Concentration risk has been reduced (four customers, not three) and visibility extended (record $110B backlog, "much more AI" semiconductor mix).
• Q2 trigger (c): "Non-AI semis prints a credible cyclical inflection across two consecutive quarters." Q3 outcome: not fired. Non-AI semis revenue was flat sequentially at $4.0B; Q4 guide of ~$4.6B is up low-double-digits sequentially but mostly seasonality. Hock explicitly characterized the recovery as U-shaped rather than V-shaped, with meaningful cyclical inflection unlikely before "late twenty-six." Broadband is the only sub-segment showing genuine cyclical strength; everything else is being held up by seasonality. We acknowledge this trigger has not fired. We also note that as AI scales, non-AI semis is now ~25% of total revenue and shrinking as a share — it was a more material thesis variable at Q2 than it is today.
• Q2 trigger (d): "Drawdown of 15–20% without thesis impairment." Q3 outcome: not applicable. The stock has been broadly higher since the Q2 print — approaching a +75% YTD gain heading into the print, with a further +9.4% on the post-print session. We are upgrading despite the absence of a drawdown setup, on the grounds that the fundamental upgrade triggers have fired more cleanly and earlier than we expected.

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 (Today — Upgrading to Outperform) — Two of the three fundamental upgrade triggers fired in a single quarter: a fourth XPU customer with $10B+ in qualified production orders, and an explicit upward re-rate of the FY26 AI growth trajectory ("material improvement" / "accelerate") supported by a record $110B consolidated backlog. Cyclical non-AI recovery has not arrived but matters less to the model now that AI is >30% of revenue and growing. Add the management-continuity positive (Hock through 2030), and the case for Outperform is substantially stronger than the case for waiting another quarter.

What gets us to a more aggressive Outperform (or wider price target): (i) a fifth XPU customer transitions from prospect to revenue (Hock noted "we may get see one more perhaps as a prospect"); (ii) the FY26 AI revenue framing translates to a specific number that is materially above current Street modeling, with the $10B Q3-FY26 delivery acting as a single-quarter anchor; (iii) a non-AI cyclical inflection finally arrives in fiscal H1 FY26, layered on top of the AI ramp; (iv) further VCF advanced-services attach (security, DR, AI workloads) accelerates the software margin trajectory.

What would put us back on Hold: (i) any one of the four XPU customers signals a 2026 capex pause or design-cycle delay; (ii) the $10B Q3-FY26 delivery slips, suggesting yield, packaging, or supply-chain slippage at TSMC or a back-end partner; (iii) a meaningful tightening of US export controls that impairs a non-trivial share of the addressable AI semiconductor TAM (this risk is real but unbounded; we monitor without modeling); (iv) software gross or operating margin reverses, suggesting the integration leverage is exhausted earlier than the framing suggests.

Results vs. Consensus

Q3 FY2025 cleared consensus on every line, with the magnitude of the AI-semis beat being the most relevant single number. Revenue beat by ~$0.15B; non-GAAP EPS beat by ~$0.03; gross margin printed 78.4% vs. a guide that was originally lower; and adjusted EBITDA margin of 67% was 100bps above the 66% guide.

Quality of Beat

The beat is broad-based and high-quality on three dimensions. First, it's powered by AI semis (+63% YoY, the tenth consecutive quarter of robust growth) rather than by software working-capital or one-time items — AI semis is the segment with the most asymmetric forward optionality, and it's accelerating. Second, software contributed cleanly with $8.4B of total contract value booked in-quarter and operating margin compressing through to 77% as the VMware integration completes its second-phase value-extraction (we'll come back to this). Third, the operating leverage was visible at the consolidated level: operating margin expanded 20bps sequentially to 65.5% even as gross margin was down 100bps sequentially on revenue mix, and adjusted EBITDA margin came in 100bps above the 66% guide.

The one item we'd flag as cosmetic-only: gross margin of 78.4% benefited from higher software mix and product mix within semis, both of which will reverse next quarter (Q4 guide is for ~70bps of GM compression on a richer XPU mix and seasonal wireless). This is the structural mix headwind we flagged at Q2 and continue to underwrite. It's a feature of the AI ramp, not a bug.

Segment Performance

Semiconductor Solutions: AI Acceleration to +63%, XPUs to 65% of AI Mix

Semiconductor revenue of $9.17B was up 26% YoY (an acceleration from Q2's pace) and represented 57% of total revenue. Operating margin of 57% expanded 130bps YoY and was flat sequentially. Inside this:

AI semis decomposition. Within the $5.2B AI revenue, XPUs accelerated to 65% of mix (vs. ~60% prior). That's roughly $3.4B of XPU revenue and $1.8B of AI networking. Both lines grew strongly, but XPUs are now growing materially faster than networking and Hock explicitly guided that XPUs will continue to outgrow networking through FY26 as the new fourth customer ramps and the original three increase their XPU compute share. The networking-as-a-share-of-AI ratio will decline — not because networking weakens (Tomahawk 6, Tomahawk Ultra, and Jericho 4 are all new product cycles into FY26) but because XPUs grow faster.

Non-AI semis is still walking sideways. Q3 non-AI revenue of $4.0B was flat sequentially. Sub-segment color: broadband strong sequentially, enterprise networking and server-storage down sequentially, wireless and industrial flat. The Q4 guide assumes broadband, server-storage, and wireless all step up while enterprise networking remains down — a mix that is dominantly seasonal, not cyclical. The forward bookings in non-AI are up >20% YoY (Hock: "nothing like AI bookings, but 23% is still pretty good"), which is supportive but not yet inflective. We continue to model non-AI as a slow U-shaped recovery into late FY26 / early FY27. This is the single piece of the Q2 thesis that has not improved in our favor; we've sized it down accordingly in our forward model.

Infrastructure Software: 77% Operating Margin, VCF 9 Released, Top-10K >90% Penetrated

Software revenue of $6.79B was up 17% YoY against a guide of $6.7B, with $8.4B of total contract value booked in-quarter. Gross margin of 93% expanded 300bps YoY. Operating margin reached 77% — a 1,000bps improvement vs. 67% a year ago, reflecting the completion of the VMware integration. On a $6.79B run-rate at 77% operating margin, software is now generating roughly $5.2B of quarterly operating profit (~$21B annualized), against a transaction price that is now well-known and progressively earned-out by these results.

The product news inside the segment is more interesting than the financials this quarter. After two years of engineering work involving 5,000+ developers, Broadcom released VCF 9.0, a fully integrated cloud platform that runs both VMs and modern containers on prem or in cloud, and is pitched as a real alternative to public-cloud workloads (including AI workloads). Hock confirmed that "way over 90%" of VMware's top-10,000 accounts have now purchased VCF licenses, but explicitly drew a distinction between licensed and deployed:

"We have sold them on it, and they bought licenses to deploy — doesn't mean they are fully deployed. Here comes the other part of our work, which is to take this 10,000 customers, or big chunk of them, who have taken — who have bought the vision of a private cloud on prem — and working with them to enable them to deploy and operate it successfully on their infrastructure on prem. That's the — the hard work over the next two years that we see happening."
— Hock Tan, CEO

The framing here matters: Phase 1 (sell-VCF) is done; Phase 2 (deploy-VCF + sell advanced services like security, DR, and AI-workload) is the next two years of the software thesis. Phase 3 (next 20–30K mid-market accounts) is more uncertain and Hock was deliberately measured about the TCO and skill-set hurdles for that cohort. We model Phase 2 as the dominant software margin driver into FY26 with optionality on Phase 3. Hock also explicitly disclaimed that the VMware footprint expansion would translate into more hardware sales for AVGO — "as they virtualize their data centers, we consciously accept the fact that we are commoditizing the underlying hardware in the data center." We respect the discipline of that statement; software and silicon are run as separate franchises and that's the right structure.

Key Topics & Management Commentary

The Fourth XPU Customer: $10B in Production Orders, Q3 FY26 Delivery

This is the headline disclosure of the call and the single biggest fundamental change since our Q2 initiation. Hock characterized it as follows:

"As we have previously mentioned, we have been working with other prospects on their own AI accelerators. Last quarter, one of these prospects released production orders to Broadcom, and we have accordingly characterized them as a qualified customer for XPUs. And in fact, has secured over $10 billion of orders of AI racks based on our XPUs. And reflecting this, we now expect the outlook for fiscal 2026 AI revenue to improve significantly from what we had indicated last quarter."
— Hock Tan, CEO

Asked specifically when the $10B ships, Hock was unusually precise: "likely to be Q3 of our fiscal twenty-six…starts and ends in Q3." That's an unusually narrow delivery window for an order of that size, which suggests the customer's data-center build-out is on a hard timeline. The implied revenue concentration in fiscal Q3 FY26 is meaningful and creates a single-quarter anchor for the FY26 trajectory.

Hock also directly addressed the source of the FY26 upgrade in response to a Deutsche Bank question:

"It's both. But to a large extent, it's the fourth quarter customer that we now add on to our roster, which we will ship pretty strongly in 2026, I should say. So combination of increasing the volumes from existing three customers — and we moved through that very progressively and steadily — and the addition of a fourth customer with immediate and fairly substantial demand. Really…changes our thinking of what '26 would be starting to look like."
— Hock Tan, CEO

The framing matters: the FY26 upgrade is not just the fourth customer being layered on top of the original three. It's the original three continuing to gain share inside their own compute footprints (XPU mix climbing within their existing AI infrastructure as software stacks stabilize) plus the new $10B from customer #4. Both vectors compound.

The FY26 Re-Rate: Quantification

Asked directly by Bank of America to quantify the FY26 AI revenue, Hock declined to give a specific number but was deliberately precise about the direction:

"When I — last quarter when I said, hey, the trend of growth of '26 will mirror that of '25, which is 50, 60% year on year — that's really all I said. I didn't quote a band. We're seeing the growth rate accelerate, as opposed to just remain steady at that 50, 60%. We are expecting and seeing 2026 to accelerate more than the growth rate we see in '25."
— Hock Tan, CEO

This is materially different language from Q2. At Q2, the framing was a steady-state +50–60%. At Q3, the framing is accelerating above +50–60% — meaning we should be modeling the FY26 AI growth rate north of +60%, not at +60%. On a Q4 FY25 AI semis exit-rate of $6.2B (i.e., ~$25B annualized), accelerating AI growth implies an FY26 AI semis number that is comfortably above $35B and could plausibly approach $40B at the upper end. That is a material upward re-rate of the FY26 algorithm.

Networking: Three Layers (Scale-Up, Scale-Out, Scale-Across) and Why Mix Declines

Broadcom is now positioned across all three layers of AI networking, with new product launches in each:

• Scale-up (intra-rack). Tomahawk Ultra is the new Ethernet alternative to NVLink, supporting up to 512 compute nodes (vs. NVLink's 72 GPUs at 28.8 Tb/s). Hock dismissed proprietary scale-up protocols as unnecessary — "tweak the switches to make the latency super good. Better than NVLink, better than InfiniBand. Less than 250 nanoseconds. Easily." Whether that's bravado or operationally true matters less than the fact that Ethernet is open-source and Broadcom is the lead vendor, with explicit XPU-Ethernet compatibility shipped at the customer interface.
• Scale-out (across racks within data center). Tomahawk 6 was launched in June with 102 Tb/s switching, flattening what was previously a three-tier network into a two-tier topology. Lower latency, lower power.
• Scale-across (across data-center sites). The newly-launched Jericho 4 (51.2 Tb/s, deep-buffering, intelligent congestion control) targets clusters above 200,000 compute nodes that span multiple data centers up to ~100km apart. This product extends a category Broadcom has been quietly shipping into for two years via the Jericho 3 generation; it is not a green-field product launch but the next-generation extension of an installed base.

The strategic insight Hock offered — and one that we think is being under-appreciated — is that the scale-across use case is being forced by power and land constraints on single-data-center footprints. Above ~100,000 GPUs, customers are putting compute in multiple sites within a regional radius and wiring them together as a logical cluster. That's a category Ethernet-based deep-buffer routers (which Broadcom has been making for telco backbone for 20 years) are uniquely well-suited for. It's a non-trivial moat extension into FY26 and beyond.

Counter-intuitively, networking will be a declining share of AVGO's AI revenue mix into FY26. That's not weakness — it's the XPU layer growing faster, and the customer mix is the right one to be over-indexed to (the four customers driving the XPU growth are buying their networking from Broadcom anyway, but the marginal networking-only customer mix is shrinking as a share of the total).

Backlog at $110B: New Disclosure

Hock disclosed for the first time a record consolidated company backlog of $110B. He declined to break it out fully, but offered three guides: at least 50% is semiconductors; that semiconductor backlog is "much more AI than non-AI"; and software adds steadily on top. On a back-of-envelope basis, that's at least ~$55B of semiconductor backlog, weighted toward AI. Against an FY25 AI semis run-rate that exits at ~$25B annualized, the implied AI backlog visibility extends well into FY26 and beyond. This is meaningfully better visibility than we had at Q2 initiation, when no backlog figure of this magnitude was disclosed.

Hock Through 2030: Succession Overhang Removed

Toward the end of his prepared remarks, Hock disclosed that he and the board have agreed he will remain CEO through at least 2030. This is a meaningful fundamental development that we did not expect. Hock has been the strategic architect of AVGO from the LSI/Avago/Broadcom/CA/Symantec/VMware acquisition arc; his continuity through the FY26–FY28 AI ramp removes a non-trivial succession-related discount from the multiple. We treat this as a small but real positive on the rating decision.

Q4 FY2025 Guidance & Outlook

Implied FY25 full-year (with Q4 guide). Q1 + Q2 + Q3 + Q4 implied = $14.92B + $15.00B + $15.95B + $17.40B = ~$63.3B in FY25 revenue, +21% YoY. AI semis FY25 implied = ~$3.0B + ~$4.4B + $5.2B + $6.2B = ~$18.8B, growing ~+59% YoY against prior-year AI of ~$11.8B. The AI-semis FY25 trajectory is the key reference point for the FY26 re-rate: if the FY26 AI growth rate accelerates from this +59% base, we land at ~$30B at the conservative end and approach $35–40B at the upper end.

Mix-shift gross margin. Q4 GM down ~70bps sequentially is consistent with the Q2-flagged structural compression we underwrote. The sequential drivers are higher XPU revenue mix (gross margin lower than the corporate average) plus seasonal wireless ramp (gross margin lower than the corporate average). Software mix ticks up sequentially but not enough to offset. Operating margin is partly insulated by the software OpEx leverage at 77% segment operating margin. We continue to model consolidated operating margin expanding modestly through FY26 before stabilizing as the software-margin lift exhausts.

Analyst Q&A Highlights

What Drove the FY26 Upgrade — Existing Three or Fourth Customer?

Ross Seymore (Deutsche Bank) opened the Q&A by asking whether the FY26 upgrade was driven by the new fourth customer, the existing three, or both. Hock answered that it's both, but with the fourth customer doing the heavier lift — "to a large extent, it's the fourth quarter customer that we now add on to our roster" — combined with the original three "increasing the volumes…very progressively and steadily." This dual-source framing matters because it means the FY26 trajectory is not entirely dependent on a single customer's ramp; it's the layered combination.

Quantifying the FY26 AI Number

Vivek Arya (Bank of America) pressed on quantification. Hock declined to give a specific number ("we're not supposed to be giving you a forecast for '26"), but used the operative phrase "fairly material improvement" and explicitly said the growth rate is accelerating rather than holding steady at +50–60%. He also confirmed that XPU mix continues to climb as a share of AI revenue and networking declines as a share — not because networking weakens but because XPUs outgrow it.

Backlog Composition

Stacy Rasgon (Bernstein) pressed on the $110B backlog. Hock declined to break it out but indicated at least 50% is semiconductors and that the semiconductor portion is "much more AI than non-AI." Software adds on a "steady basis." This level of detail is more than Broadcom has historically given on backlog and is, by itself, an upgrade to the FY26 visibility we had at Q2.

The Fifth Customer? Hock's Hedge

Ben Reitzes (Melius) and Jim Schneider (Goldman Sachs) both asked about prospects beyond the now-four qualified XPU customers (out of an initially-cited universe of seven). Hock was deliberately conservative on additional customers transitioning to revenue: "I rather not give you any more color on prospects…sometimes we stumble into production in fairly — in time frames that are fairly unexpected, surprisingly. Equally, it could get delayed." And on the question of whether more than seven prospects exist: "We may get see one more perhaps as a prospect. But again, we are very thoughtful and careful about even making that qualification. But right now, for sure, we have seven." We read this as: there is one additional prospect being explored, but the bar for qualifying as the fifth XPU customer remains intentionally high, consistent with how AVGO has run this discipline historically.

Non-AI Cyclical Posture

Harlan Sur (JPMorgan) probed whether the magnitude of the non-AI cyclical recovery was building toward a more meaningful upturn. Hock was direct that he expects U-shaped (not V-shaped) recovery, with broadband the only consistent uptrend, and "perhaps by late twenties — mid-twenty-six, late twenty-six, we'll start to see some meaningful recovery." Bookings are up 23% YoY in non-AI. We treat this as confirming our Q2 view that non-AI is a slow-grinding recovery, not yet inflective.

Gross-Margin Walk for Q4

CJ Muse (Cantor Fitzgerald) asked the gross-margin question directly — how a 70bps GM decline reconciles with higher XPU and wireless mix offset by stronger software. Kirsten Spears confirmed the math: TPUs (XPUs) and wireless are typically higher-volume, lower-margin (wireless is seasonally Q4-heavy), while software margin steps up a bit. The walk implies semiconductor gross margins step down in Q4 on AI mix, software gross margin steps up, and the consolidated arithmetic delivers ~70bps of compression. This is structural and we underwrite it.

Scale-Up Ethernet vs. NVLink, UALink, Sue

Joshua Buchhalter (TD Cowen) and Christopher Rolland (Susquehanna) asked about the competitive positioning of Tomahawk Ultra against NVLink and UALink in scale-up. Hock's posture was consistent and confident: Ethernet is open-source, well-known to hyperscaler architects, latency can be tweaked sub-250ns when the use case demands, and proprietary protocols are an unnecessary complication. He explicitly framed AVGO's XPUs as having Ethernet-compatible interfaces shipped at the customer side — meaning even on scale-up, where NVLink has been the incumbent, Broadcom has the customer relationship. The hyperscaler-led shift toward Ethernet for scale-up is a meaningful tailwind we now assign more weight in the model.

VMware Phase 2: Top-10K Deployment, Mid-Market Optionality

Carl Ackerman (BNP Paribas) drilled into the VMware story. Hock's response did the most work for our software thesis: 90%+ of the top-10K accounts have purchased VCF, but deployment is the next two-year slog (which is where advanced-services attach and operating-margin expansion will come from). The next 20–30K mid-market accounts is exploratory; Hock was deliberately measured about the TCO and skill-set hurdles. We model Phase 2 (deploy + advanced services) as the dominant software contribution into FY26 and treat Phase 3 (mid-market) as call-option upside.

XPU Share within Existing Customers

Harsh Kumar (Piper Sandler) closed the Q&A with a useful question on whether AVGO's XPU share at the existing four customers will eventually exceed those customers' merchant GPU share. Hock answered yes, calling it a "logical conclusion" and a "multi-year journey…multigenerational" with each chip generation increasing customer confidence to deploy a higher percentage of their compute footprint on their own XPUs. This is the most explicit forward statement Hock has made about the share-gain trajectory within the existing customer base.

What They're NOT Saying

• Specific FY26 AI revenue number. Hock explicitly declined to quote a band. The "fairly material improvement" language is deliberate and gives him the flexibility to over-deliver. We read this as conservatism rather than evasion — AVGO has a strong track record of guiding tightly on quarterly numbers and giving directional FY framings only.
• Customer names for XPU customers 1–4. Industry consensus has long held that the original three are Google, Meta, and ByteDance. The Q3 disclosure of the fourth is unattributed; market speculation has variously suggested OpenAI or Apple, with OpenAI more commonly cited based on the build-out timeline. Hock will continue to refuse to confirm specific names. The risk in our model is the customer-concentration question — we can size the dollars but not stress-test the per-customer dependence directly. We treat the four-customer disclosure as a meaningful concentration improvement vs. three, but not as eliminating concentration risk entirely.
• Networking dollar revenue inside AI. Hock gave the XPU/networking mix ratio (65%/35%) but not the absolute networking dollars. With AI semis at $5.2B and XPUs at 65%, we infer ~$1.8B of AI networking, but management did not validate this directly. In FY26, with networking declining as a share of AI mix but still growing in absolute dollars, the absolute networking line becomes harder to model precisely.
• VCF deployment-rate metrics. Hock disclosed that 90%+ of top-10K accounts have purchased VCF but explicitly differentiated this from deployed. He did not disclose what percentage of purchased licenses are deployed today, nor did he provide a deployment-cadence metric. This is the leading indicator for advanced-services attach (security, DR, AI workloads) and will matter materially to the FY26–FY27 software margin trajectory.
• Tariff and export-control exposure. The call was conspicuously light on tariff or export-control commentary. Given that one of the three original XPU customers is widely understood to be ByteDance (which would be subject to potential export-license requirements), we'd have expected at least defensive language. The silence may reflect the fact that the relevant production is occurring under existing licenses; or it may be that management is intentionally avoiding setting up a guidance-cut narrative. Either way, we monitor without modeling.
• Capital allocation / buyback re-engagement. Q3 returned $2.8B in dividends but no buyback discussion. AVGO has historically been disciplined about buybacks during M&A digestion (the post-VMware deleveraging is the relevant context), and we'd expect buyback re-engagement to feature in the Q4 / FY26 capital-allocation discussion. The silence here is consistent with continued debt paydown discipline; it's a positive on credit quality but a small drag on near-term EPS optics.

Market Reaction

AVGO closed up approximately +9.4% on the post-print session (Sept. 5), with after-hours trading on the print itself reflecting the strong AI semi result and the FY26 re-rate. The +9.4% move is meaningful given that AVGO had already rallied substantially heading into the print (approaching +75% YTD) and was therefore not setting up with a low-bar investor base. The market read the print the same way we did: the fourth XPU customer with $10B+ in qualified production orders, plus the explicit FY26 AI growth rate acceleration, are the headline value-drivers. The +9.4% move is a high-quality, fundamentally-driven re-rate rather than a short-cover or technical bounce.

That the move is this large despite the equity already running into the print is, paradoxically, supportive of an Outperform rating. It means the market continues to find new ground to underwrite as the FY26 picture becomes more concrete. The principal risk to the upgrade is that incremental upside from here requires either further FY26 AI specifics (likely at the Q4 FY25 print or at the customary post-CES update) or further fundamental developments (a fifth XPU customer; a non-AI cyclical inflection). We think the asymmetry still favors longs.

Street Perspective

Debate: Is the FY26 AI Growth Rate "Accelerating" Above +60% Already in Consensus?

The bull case being made on the Street is that the Q3 disclosure of the fourth XPU customer plus the FY26 acceleration framing means consensus FY26 AI semis estimates are stale and need to walk up materially — potentially toward the $30–35B range from prior consensus closer to $25–28B. The bear case is that the $10B in fourth-customer orders is already at least partly in some sell-side models (the customer's existence has been telegraphed for several quarters) and that the "acceleration" framing may translate into a +65% growth rate rather than the +75–80% that some bulls now want to model. Our take: the truth is between, but closer to the bull case. We model FY26 AI semis at $32–35B (vs. our prior model of $28–30B), which is a meaningful upward revision but still leaves room for upside if the original three customers ramp XPU-share faster than we assume.

Debate: Is the Non-AI Floor a Drag, or Does It Stop Mattering?

The bull case being made on the Street is that as AI scales toward 40%+ of revenue, non-AI semis becomes a smaller share and its cyclical drag stops mattering for the consolidated trajectory. The bear case is that non-AI is still ~25% of revenue, and a slow U-shaped recovery means consolidated revenue growth continues to be capped relative to a scenario where non-AI joins the AI rally. Our take: the bull side is right for the FY26 consolidated trajectory but the bear side is right that the non-AI cyclical recovery, when it eventually arrives, will be a meaningful tailwind we have not yet underwritten. We model non-AI as flat-to-up modestly through FY26 and as a positive optionality layer beginning fiscal H1 FY27.

Debate: Is XPU GM a Multi-Year Drag on Consolidated GM?

The bull case being made on the Street is that XPU GM is structurally below the corporate average but growing at the highest incremental ROIC of any line in the model, so the consolidated operating margin keeps expanding via software-margin lift and operating leverage. The bear case is that as software margin saturates near 78–80%, the XPU mix headwind eventually starts to bite at the consolidated level. Our take: the bull case is right through FY26–early FY27; the bear case becomes more relevant in FY27–FY28 when the software-margin lift is largely earned out. We model consolidated operating margin expanding modestly through FY26, peaking in FY27, and stabilizing thereafter.

Model Implications

The Q3 print and accompanying FY26 re-rate prompt the following changes to our baseline model:

• FY25 revenue. Walking up to ~$63.3B (+21% YoY) from prior ~$62B as Q3 cleared and Q4 guide is above prior assumption.
• FY25 AI semis. ~$18.8B (+59% YoY). This is the operative reference for the FY26 re-rate.
• FY26 AI semis. Walked up to $32–35B (+70–85% YoY) from prior $28–30B (+50–60% YoY). The upper-end is anchored by the $10B fourth-customer Q3 FY26 delivery layered on top of the original three customers' continued share-gain inside their compute footprints.
• FY26 non-AI semis. Modeled flat-to-up modestly at $17–18B; we have removed the modest cyclical recovery we previously layered into fiscal H2 FY26 in light of Hock's "U-shaped, mid-to-late twenty-six" framing.
• FY26 software. Modeled at ~$28–30B (+10–12% YoY) with operating margin expanding to ~78–80% as Phase 2 (deployment + advanced services) attaches. VCF Phase 3 (mid-market) is treated as call-option upside, not modeled.
• FY26 consolidated revenue. ~$77–83B (+22–31% YoY) up from prior ~$75–80B. The dispersion in the range is dominantly the FY26 AI semis question.
• FY26 consolidated operating margin. Modeled at ~67–68% (modest expansion vs. FY25 ~66%) on continued software margin lift offsetting the structural XPU mix headwind on gross margin.
• FY26 non-GAAP EPS. Walking up to ~$8.50–9.50 from prior ~$7.75–8.75, driven by AI revenue acceleration plus continued capital return. Reasonable compounders at this growth rate trade at high-20s to low-30s forward P/E; on a midpoint $9 EPS, that maps to a fair-value range that supports an Outperform rating against current setup.
• Capital structure. Continued debt paydown (gross debt $66.3B at Q3 with weighted-average rate 3.9%) on ~$28B of FY25 free cash flow; we model net leverage approaching the sub-2x target by mid-FY26.

Valuation framing. AVGO at the post-print level trades at a forward P/E in the high-20s on our revised FY26 EPS estimate. For a franchise printing 20%+ revenue growth, expanding operating margins, $28B+ of free cash flow, and a record backlog with multi-year visibility, that multiple is defensible relative to the AI-compounder peer group and at the lower end of where the highest-quality AI compounders have traded through this cycle. The asymmetry on our updated FY26 EPS range still favors longs.

Asymmetric scenarios that could change our rating again: (i) a fifth XPU customer transitions from prospect to revenue with FY26 contribution → supportive of widening Outperform conviction; (ii) any of the four XPU customers signals a 2026 capex pause → downgrade pressure (size of the move depends on the customer); (iii) a non-AI cyclical inflection arrives in fiscal H1 FY26 → modest further upgrade pressure; (iv) export-control regime tightens materially → downside pressure but bounded.

Thesis Scorecard Post-Earnings

Overall: Five of five bull points are now confirmed (with three of them materially upgraded since Q2). Of the five bear points, one has improved meaningfully (concentration), two are confirmed and underwritten (XPU GM, valuation), one is live but bounded (export controls), and one remains unresolved (non-AI floor). The new fundamental positive (Hock through 2030) was not in the Q2 thesis and is incremental. The composition is now decisively bull-skewed.

Action: Upgrading to Outperform from Hold. The Q2 initiation specified that we needed to see a fourth XPU customer transition to revenue, the FY26 AI growth rate sustaining above +60%, OR a non-AI cyclical inflection to upgrade. The Q3 print delivered the first two of those triggers explicitly and concretely, and added a meaningful management-continuity positive on top. Two of three explicit upgrade triggers firing in a single quarter, with the magnitude of the FY26 re-rate larger than we modeled at Q2, is more than enough to walk the rating. We move to Outperform on AVGO with a 12-month horizon, with continued upside optionality from a fifth customer, the eventual non-AI cyclical recovery, and continued VCF advanced-services attach. We will revisit on the Q4 FY25 print.