Marvell Technology

Phase A — Understand the business

Lens 1 · Company Overview

Marvell is a fabless designer of data-infrastructure semiconductors — analog, mixed-signal and digital IP integrated into custom ASICs, interconnect/optical DSPs, Ethernet switches & PHYs, storage controllers and DPUs. It reports as one segment ("design, development and sale of integrated circuits") split across two end markets: Data center and Communications & other.

The business has been radically reshaped into an AI-infrastructure pure-play in three years:

• FY2026 net revenue $8,194.6M, +42.1% YoY (FY2025 $5,767.3M; FY2024 $5,507.7M).
• Data center = $6,100.3M, 74% of revenue in FY2026 — up from 72% (FY2025) and just 40% (FY2024). AI-related revenue was ~50% of data-center revenue in FY2025, up from <10% in FY2023.
• Communications & other = $2,094.3M, 26% (+31% YoY as carrier/enterprise inventory normalized).

Business model. Two engines: (1) custom ASICs ("XPUs") co-designed for individual hyperscalers — built on Marvell IP (224G SerDes, ARM compute, security, custom HBM, silicon photonics, advanced packaging/CoWoS), executing 5nm/3nm now and 2nm/1.4nm (A14) next; and (2) a merchant electro-optics/interconnect franchise (PAM & coherent DSPs, TIAs, drivers, silicon photonics, AEC/LPO chipsets, PCIe retimers) that is the connective tissue of AI clusters. The custom business is the growth driver; the optics franchise is the durable annuity.

Contract structure. Design-win driven, not take-or-pay. Marvell places firm foundry orders up to 26 weeks out and capacity reservations up to 52 weeks, and signs multi-year capacity-reservation agreements with foundries/substrate partners. Concentration is severe (see Lens 3 / Lens 13). It has also issued customer warrants (4.2M shares @ $87.77, 7-yr; 1.0M shares @ $87.00) that vest on the customer hitting revenue milestones and are booked as a reduction to revenue — i.e. it is paying a hyperscaler in equity to lock in volume.

Lens 2 · Supply Chain

Upstream → Marvell → end customer, named at each link:

• EDA / IP: Synopsys, Cadence (tools); Arm (compute cores licensed into ASICs/DPUs).
• Foundry (the chokepoint): "Most of our products are manufactured by third-party foundries located in Taiwan" — overwhelmingly TSMC for leading-edge (5/3/2/1.4nm), with secondary sources in China, Germany, South Korea, Singapore, the U.S.. Single-source by design: "most of our products are not manufactured at more than one foundry… designed to be manufactured in a specific process at only one of these foundries" — transition to an alternate foundry would be slow and costly. This is the hardest chokepoint in the whole chain.
• Advanced packaging: TSMC CoWoS (also InFo, EMIB) — the same CoWoS line that gates NVIDIA/AMD; HBM integration is a packaging-capacity bet.
• Memory: custom HBM partnerships (Micron / Samsung / SK hynix ecosystem) for in-package memory.
• OSAT (assembly/test/packaging): subcontractors in China, Malaysia, Singapore, Taiwan, Canada — also single-sourced in places.
• Distribution: a single distributor (Distributor A = 37% of FY2026 revenue) fronts a large share of volume — almost certainly the channel through which custom/hyperscaler programs flow.
• End customers: hyperscalers — AWS (Trainium), Microsoft (Maia), Meta, Google (in talks), plus carrier/enterprise OEMs for comms.

Geographic chokepoint. Sales shipped to Asia were 77% of FY2026 revenue (75% FY2025; 70% FY2024); in Q1 FY2027 China was 44% of shipment destination (37% PY) and Taiwan 21%. Marvell flags that most China shipments are to non-China customers with China-based contract manufacturing — a mitigant, but the export-control surface is real and growing.

Lens 3 · Competitive Advantages (moats)

The moat is a duopoly position plus a vertical-IP stack that is genuinely hard to replicate — but it is narrower than bulls imply.

• Structural position: Marvell + Broadcom control ~95% of the custom-AI-ASIC co-design market; Broadcom ~70%, Marvell ~20–25%. Being the credible #2 is itself a moat — a hyperscaler that wants a second source to Broadcom has essentially one option.
• IP scale / process leadership: 224G SerDes, custom HBM, silicon photonics, advanced packaging, and a multi-node roadmap (5→3→2→1.4nm) that few merchant players can match. 18 active custom-silicon programs in hand. Switching cost is real: an XPU is co-designed over years and locked to Marvell's SerDes/packaging.
• Optics franchise = the durable edge: Marvell is the leader in electro-optics DSPs (PAM/coherent), the connective layer every AI cluster needs regardless of whose accelerator wins. The Celestial AI Photonic Fabric acquisition (see Lens 5) extends this into scale-up optical interconnect, the next bottleneck as XPU links move from copper to optics.
• The NVIDIA validation (decisive, 2026): NVIDIA took a $2.0B convertible-preferred stake and pulled Marvell's custom XPUs + optics into NVLink Fusion and AI-RAN. This is a moat-widener: the dominant accelerator vendor is now an investor and ecosystem partner, blunting the "hyperscalers/NVIDIA will route around Marvell" bear thesis.

Bargaining power — asymmetric and against Marvell. Its customers (AWS, MSFT, Google) are ~$2T+ companies that fund their own silicon teams and can dual-source; its key supplier (TSMC) is a monopoly at the leading edge. Marvell is structurally squeezed from both ends — its leverage is design-win lock-in and IP, not commercial power. The customer warrants (paying hyperscalers in equity) are the tell.

Lens 4 · Segments

Marvell discloses end-market and geography, not P&L-by-segment (one reportable segment), so segment operating income is n/a — not disclosed at segment level.

By end market:

Trend & cause: a violent mix inversion. Data center went 40% → 76% of revenue in two years — accelerating (FY2026 DC +46% YoY; Q1 FY2027 DC +27% YoY but guided to re-accelerate). Comms collapsed in FY2025 (carrier inventory glut) then recovered +31% in FY2026; it is now a cyclical sidecar to the AI story.

By geography (Q1 FY2027 shipment destination): China 44%, Taiwan 21%, U.S. 7%, Other 28%. By customer type (Q1 FY2027): Distributors 51% / Direct 49% — distributor mix rising as custom programs scale through the channel.

---

Phase B — Measure performance

Lens 5 · Earnings Result (Q1 FY2027, quarter ended 2026-05-02)

The headline GAAP number is a trap; the operating quarter was strong.

• Net revenue $2,417.8M, +27.6% YoY (PY $1,895.3M); +$18M above the prior guide midpoint. A modest beat, not a blowout.
• GAAP gross profit $1,260.8M = 52.1% GM (PY 50.2%); non-GAAP GM 58.9%.
• GAAP operating income $339.4M, +25.4% (PY $270.6M). Operating leverage is muted because R&D = $652.3M (27% of revenue) and SG&A $258.4M — Marvell is spending heavily to hold its node/optics lead.
• GAAP net income $34.5M / $0.04 diluted — a cliff from PY $177.9M / $0.20. The entire collapse is below the line: "Other expense, net" swung to −$203.3M (PY −$6.0M), driven by a $331.8M non-cash increase in the fair value of the Celestial earnout contingent-consideration liability, partly offset by an $81.1M unrealized gain on the forward stock-purchase hedge. The liability grew because the deal is going well (higher MRVL stock + higher revenue-milestone probability) — a perverse GAAP penalty for success.
• Non-GAAP net income $718.0M / $0.80 diluted — the number the market trades on.
• Record operating cash flow $638.8M.
• Balance sheet: cash $3,843.6M; goodwill jumped $11,062M → $13,884M and acquired intangibles $1,755M → $2,562M (the Celestial + XConn closings); long-term debt $4,961M (well-laddered senior notes, no maturities until FY2029); APIC +$3.9B (NVIDIA preferred $2.0B + acquisition stock). Inventory roughly flat ($1,400.9M) vs revenue up — clean.
• Guidance: Q2 FY2027 = $2.7B at midpoint (+35% YoY); management guided revenue to accelerate each quarter through FY2027 (full-year ~$11.5B, +~40%) and FY2028 to ~$16.5B. The tone turned decisively bullish vs the cautious FY2026 exit.
• Market reaction: stock +24% to ~$166 on the print. (The bigger move came later — see Lens 8.)

Unusual vs own history: the $331.8M earnout mark and the divestiture-distorted FY2026 comparison (the $1.8B Infineon gain inflated FY2026 GAAP net income to $2,670M) make GAAP nearly useless this year; non-GAAP and cash flow are the honest lenses.

Lens 6 · Earnings Calls (sentiment trend)

transcripts/ is empty, so this is ``-grounded across the last ~4 prints:

• Q3 FY2026 (Dec 2025) — defensive. The call landed amid the Trainium-loss scare (see Lens 8/13); management leaned on the optics franchise and the breadth of the 18-program pipeline. Tone: reassuring, on the back foot. Stock fell ~6–9% in the surrounding window.
• Q4 FY2026 (Mar 2026) — pivot. Announced the Celestial AI acquisition as a "transformational milestone" for scale-up; framed data-center growth for the next year as above prior expectations. The narrative flips from "are they losing share?" to "they're buying the next moat."
• Q1 FY2027 (May 27 2026) — euphoric. "Exceptional AI-related demand," upward FY revision, accelerating-every-quarter language, FY2028 ~$16.5B target volunteered.

Recurring phrases: "custom silicon," "scale-up," "electro-optics leadership," "accelerating," "design wins." What they stopped saying: the cautious "inventory normalization / digestion" framing that dominated the FY2025 comms-collapse calls. The arc over four quarters is caution → pivot → conviction, tracking the stock from ~$90 fear to ~$310 euphoria — sentiment and price are now tightly, and dangerously, correlated.

Lens 7 · Comps

Marvell + custom-silicon / AI-infrastructure peers. Multiples are ``, mid-June 2026, and the table is provenance-honest — where a clean figure isn't sourced it says so. MRVL's own multiples are extreme and must be read against the FY2026 non-GAAP base ($2.84) and forward consensus.

Read: MRVL trades at a clear premium even within the AI-infra basket. Analyst average PT $238.75 (S&P Global, "Strong Buy") sits ~25% BELOW the ~$310 spot — the stock has out-run even the sell-side bulls. That gap is the single most important valuation fact in this dossier.

Lens 8 · Stock-Price Catalysts (>5% moves, last ~18 months)

Marvell is a pure AI-narrative beta — it moves on hyperscaler design-win signals and macro AI-capex sentiment, not on its own quarterly beats per se. The five-year tape is dominated by AI:

• 2024–early 2025: repeated AI-capex updrafts as DC revenue compounded ~85–88% YoY; the custom-silicon story re-rated the multiple.
• Dec 2025 — the scare (−6 to −9%): reports that Marvell lost AWS Trainium 3/4 to Alchip (chiplet-vs-monolithic-die debate) + Microsoft–Broadcom design talks; Benchmark downgraded to Hold "with high conviction" of the loss. Stock troughed near ~$90.
• Mar 31 2026 — the NVIDIA deal (+10–20%): $2.0B convertible-preferred investment + NVLink Fusion / AI-RAN partnership.
• May 27 2026 — Q1 print (+24% to ~$166): raised FY27/FY28 outlook.
• Jun 2 2026 — "next trillion-dollar company" (+30%+ single session): Jensen Huang anointed MRVL on stage; record close $316.43 on Jun 4.
• Jun 5 2026 — S&P 500 inclusion announced (replacing Pool Corp, effective Jun 22 2026); forced index buying; stock ~$324.79 on Jun 18.
• Jun 15 2026 — new CFO Dan Durn (ex-Adobe, ex-Applied Materials) appointed.

What the pattern reveals: the market reacts hardest to single-customer design-win news (Trainium loss, NVIDIA tie-up) and narrative endorsements (Huang), more than to the income statement. That makes the name headline-fragile in both directions — the same lever (a hyperscaler decision) that 3x'd it can halve it. The stock round-tripped from a ~$90 design-loss scare to ~$320 euphoria in six months.

---

Phase C — Judge people & books

Lens 9 · Management

• Track record (strong). Matt Murphy, Chairman & CEO since July 2016 (~10-yr tenure; 22 prior years at Maxim Integrated). Under him Marvell's enterprise value grew >10x via aggressive portfolio reshaping — exited legacy (Wi-Fi, auto-ethernet) and pivoted hard into data-center/AI. FY2026 non-GAAP EPS $2.84, +81% YoY. This is one of the better large-cap-semi capital-allocation records of the decade.
• Capital allocation (mostly disciplined, now aggressive). FY2026: sold auto-ethernet to Infineon for $2.5B all-cash (Aug 2025, a clean $1.8B gain), then bought Celestial AI (~$3.5B) + XConn to buy the scale-up-optics moat. Returned $2.2B to holders in FY2026 — $2.0B buybacks (26.6M shares at avg ~$85 via a $1.0B ASR) + $205M dividends; $5.5B buyback authorization remaining. The buyback timing looks brilliant in hindsight — repurchasing at ~$85 against a stock now ~$310 (3.6x). Capex is light ($354M/yr, fabless) — capital goes to R&D, M&A and buybacks, not fabs.
• Skin in the game / comp. CEO 2026 comp ~$25.1M (−28% YoY), 95% performance-based (TSR + EPS-CAGR multipliers; FY2025 annual incentive paid 131% of target on revenue/margin beats). Structurally well-aligned; insider ownership modest (professional-manager profile, not founder).
• New CFO Dan Durn (ex-Adobe CFO, ex-Applied Materials CFO) effective Jun 15 2026 — a heavyweight hire that adds credibility for the scale-up to $16B+ revenue.
• Red flags (mild): the customer-warrant mechanism (paying hyperscalers in equity for volume) and the Celestial earnout (up to ~22–24M dilutive shares + a $300M forward-purchase hedge to manage stock-price exposure) add complexity and dilution. Archetype: professional operator-allocator, not founder — appropriate for this scale, but it means the moat is institutional, not visionary-dependent.

Lens 10 · Forensic Red Flags

Acting as a forensic analyst — and the honest finding is that the accounting is clean but unusually noisy this year.

• Earnout mark distorts GAAP (disclosed, non-cash). The $331.8M Celestial contingent-consideration fair-value charge runs through "Other expense, net" and will recur every quarter with stock-price/revenue-probability swings (Monte-Carlo, Level 3; balance grew $315.8M → $647.6M in one quarter). This is real GAAP volatility, not fraud — but it makes GAAP EPS unusable and pushes investors onto non-GAAP, which is exactly where SBC hides (below).
• Stock-based comp flatters non-GAAP. FY2026 SBC = $590.8M (~7% of revenue) and Celestial/XConn added ~$205M of post-acquisition SBC to amortize. The $0.80 non-GAAP quarter excludes a large, genuine cash-equivalent cost — the non-GAAP-to-GAAP bridge is the single thing to police on this name.
• Goodwill/intangibles balloon = impairment risk. Goodwill $13.9B (51% of $26.9B assets) + intangibles $2.6B. FY2025 already took $357.9M of impairments on prior acquisitions. If Celestial's scale-up optics under-deliver, expect a future write-down.
• Receivables factoring & warrants. ~$300M/qtr of receivables sold non-recourse (factoring) — normal but worth noting; customer warrants reduce reported revenue.
• Cash vs earnings: OCF ($638.8M Q1; $1.8B FY2026) comfortably exceeds the noisy GAAP net income — cash generation is real and improving, the reassuring counterweight to the GAAP optics. Inventory in line with revenue; no receivables blow-out.

Regulatory findings (required).

• SEC: No Litigation Releases and no AAERs naming Marvell Technology since 2021-06-21 (EDGAR EFTS LR + AAER search).
• 10-K Item 3 (Legal Proceedings): routine IP/commercial litigation in the ordinary course; no single matter flagged as material to financial condition in the FY2026 10-K.
• Non-SEC / web: no material FTC/DOJ/FDA/CFPB enforcement actions surfaced; the live regulatory exposure is U.S.–China export controls on advanced semiconductors (a policy/operating risk, not an enforcement finding) given 44% of shipments destine to China.
• Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER), web search, and 10-K Item 3 as of 2026-06-21. The accounting risk here is complexity and non-GAAP reliance, not enforcement.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (FY2027 / FY2028 / FY2029 — non-GAAP EPS)

Built bottom-up from the Q1 FY2027 actual ($0.80 non-GAAP) + management's own revenue guide. All outputs ``; share count ~895M diluted and rising on dilution/earnout.

Revenue anchors: FY2027 ~$11.5B (+40%); FY2028 ~$16.5B (+~44%, DC +~55%); FY2029 ~$20–21B (+~25%, normalization).

Base-case arithmetic: FY27 = Q1 $0.80 stepping to ~$1.00+ by Q4 on $2.7B→$3.2B quarters at 59–60% non-GAAP GM and ~32% non-GAAP op-margin → ~$3.60. FY28 = $16.5B × ~33% non-GAAP op-margin × (1 − ~14% tax) / ~900M sh ≈ ~$5.60. Operating leverage is the swing factor — R&D at 27% of revenue must fall as a % as revenue scales for the bull case to land.

Valuation cross-check: at ~$310, the stock is ~86x FY27e and ~55x FY28e base non-GAAP EPS. Even on the bull FY28 ($6.60), that's ~47x — rich for a company whose growth decelerates after FY28 and whose revenue is hyperscaler-concentrated. The price already discounts the base-to-bull FY28 outcome.

Brier forecast not logged (watchlist/breadth mode — no forecast.ts create per skill rules). If promoted to a thesis, the scoreable line would be: "MRVL FY2028 (ending ~Jan-2028) non-GAAP EPS ≥ $5.00, p≈0.55."

Lens 12 · Bull vs Bear

Bull case. Marvell is the only credible #2 in a custom-AI-silicon duopoly growing 45%/yr, with a widening optical moat (Celestial Photonic Fabric: 16 Tbps/chiplet, 10x a 1.6T port) that becomes essential as XPU links go copper→optical for scale-up. NVIDIA's $2.0B stake + NVLink Fusion integration converts the most dangerous potential adversary into a partner and distribution channel. Management has a >10x-EV track record and just timed $2.0B of buybacks at ~$85. If FY2028 hits ~$16.5B and the multiple holds, the bull math compounds. Jensen's "next trillion-dollar company" framing is a self-fulfilling narrative magnet now reinforced by S&P 500 forced buying.

Bear case (the 2–3 that could permanently impair or de-rate).

1. Hyperscaler in-sourcing / second-sourcing. The customers ARE the competitors — they fund their own silicon teams and proved in Dec 2025 they will route a flagship program (Trainium 3/4 to Alchip) away from Marvell. A repeat on Maia or a Google "no" permanently caps the growth multiple.
2. Concentration + warrants. ~76% DC revenue, Customer A 14%, Distributor A 37%, and Marvell is paying hyperscalers in equity to retain them — pricing power runs the wrong way. A single program pause hits revenue, margin (mix) and the multiple at once.
3. Valuation / expectations. At ~109x trailing and ~86x FY27e non-GAAP, the average analyst PT ($238.75) is 25% below spot. Any AI-capex normalization or a guide that merely meets (not beats-and-raises) likely triggers a 30–40% de-rate; the bear technical target is ~$128.

Pre-mortem (18 months out, thesis broke): AI-capex growth decelerated from the 2025–26 peak; one hyperscaler paused/re-architected an XPU; ASIC mix compressed gross margin faster than modeled; Celestial scale-up slipped a node and took a goodwill impairment; the S&P-inclusion forced-buying unwound. The stock round-trips toward $130–160 — exactly the kind of move it already did once in Dec 2025.

Contrarian view (what the market is refusing to see): the bull narrative treats the NVIDIA partnership as pure upside, but it's double-edged — NVIDIA is now a 21.8M-share-convertible holder whose interests (NVLink Fusion ecosystem control) may not always align with Marvell maximizing merchant optics sales to NVIDIA's rivals (AMD/UALink camp). Marvell is being pulled into NVIDIA's orbit at the exact moment its hyperscaler customers are backing the competing open standard. That tension is unpriced.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case:

• The growth is rented, not owned. Revenue concentrates in a handful of hyperscaler XPU programs that are re-competed every generation. Dec 2025 already showed the failure mode: Marvell's chiplet Trainium 3 proposal lost to Alchip's monolithic die. There is no contractual moat — only the current design win.
• The most dangerous competitor bulls underrate is Broadcom, not a startup. Broadcom holds ~70% share, has deeper packaging relationships and a far larger balance sheet; it is already in talks to take Microsoft Maia business. Marvell is the price-taker in every negotiation.
• The accounting invites "look-through-the-noise" bullishness. GAAP EPS is $0.04; the bull case rests entirely on non-GAAP that adds back ~$590M+/yr of real SBC and a goodwill base ($13.9B) one disappointment away from impairment. The Celestial earnout is a recurring negative-convexity GAAP item tied to the very stock price bulls are levered to.
• Dilution is structural. NVIDIA preferred (up to 21.8M shares), Celestial earnout (up to ~22–24M shares), assumed option pools, customer warrants — the share count grinds higher precisely as the stock is richest.
• What must hold for ~$310: FY2028 ~$16.5B revenue, ~33% non-GAAP op-margins, NO hyperscaler defection, NO AI-capex air-pocket, AND the ~85x forward multiple persisting. If growth disappoints 20–30% (FY28 lands ~$12–13B, EPS ~$3.90), the stock at even 40x = ~$156 — a ~50% drawdown.
• The single permanent-impairment scenario: AI training/inference costs fall sharply (cheaper alternatives), hyperscalers cut custom-XPU budgets, and Marvell — without merchant-GPU breadth or Broadcom's diversification — is left with an over-built optics/ASIC cost base. Plausibility: low-to-moderate, but it is the asymmetric tail the price ignores.

Lens 14 · Management Questions (ordered by information value)

1. With Trainium 3/4 reportedly going to Alchip, what is your current share of each named hyperscaler's next-gen XPU sockets — and which of the 18 programs are production-ramping vs. design-phase?
2. What governance, IP, and channel rights did NVIDIA receive with the $2.0B preferred, and do any terms constrain your merchant optics/ASIC sales to NVIDIA's competitors (AMD, the UALink camp)?
3. Quantify the Maia relationship today given Microsoft–Broadcom talks: is Marvell the lead designer for the next Maia generation, yes or no?
4. What is the path to non-GAAP operating margin if R&D is currently 27% of revenue — at what revenue level does R&D intensity fall, and to what?
5. Walk through the Celestial earnout: under what revenue path do you issue the full ~24M shares, and how should we model the recurring fair-value mark in "Other expense"?
6. What is the realistic FY2028 gross-margin range as custom-ASIC mix rises, and how much does Celestial scale-up optics offset the ASIC margin drag?
7. How concentrated is Distributor A (37%), and which end customers/programs flow through it?
8. What is your committed TSMC leading-edge + CoWoS capacity allocation for FY2027–28, and how does it rank against NVIDIA/AMD claims on the same lines?
9. On the customer warrants — how much revenue (gross) are you forgoing via warrant amortization, and is this becoming a standard cost of winning hyperscaler volume?
10. How exposed is the business to U.S.–China export controls given 44% of shipments destine to China; what share is genuinely at policy risk?
11. What would have to be true for you to NOT take a goodwill impairment on Celestial/XConn within three years?
12. How do you think about buybacks at ~$310 vs. the ~$85 average of FY2026 repurchases — is the $5.5B authorization now on hold?
13. What is the inference-vs-training mix of your custom pipeline, and how does the Google inference discussion change it?
14. Beyond NVLink Fusion, what is your scale-up roadmap if the industry standardizes on UALink (which your own customers back)?
15. What single competitive development would most threaten the FY2028 plan, and what is your hedge?

---