Marvell Technology
Fabless IP-and-silicon: designs chips (custom XPU/ASIC co-design, optical DSPs, Ethernet switches, DCI, storage/networking), outsources fab to TSMC; revenue from product sales plus NRE/engineering on multi-year custom-silicon programs. ~76% of revenue is data center (Q1 FY27).
The thesis on this name
State of AI Compute
Fork-agnostic DSP/ASIC + custom-silicon rent that wins under LPO, CPO, or optical-I/O, PLUS the only public scale-up optical-I/O asset (Celestial), at a forward P/S below its own median despite accelerating interconnect (+70%) — BUT the verify is sharply split: in the ASIC shift-point it LOST Trainium3/4 to Alchip (a realized, 6-month-old, priced negative)…
State of AI Compute
Long MRVL over 12-18 months on an analytical/structural edge: the market is conflating Marvell's commoditizing custom-ASIC leg (Trainium3/4 lost to Alchip, priced) with its durable, mispriced optical-I/O + interconnect rent — guided >70% YoY interconnect growth FY27, the only public scale-up-optical asset (Celestial A…
State of Nvidia
Co-design + open-fabric supplier; with Broadcom ~95% of the custom-silicon co-design market. Switch silicon slips late — dates the NVLink m…
Earnings, margins, COGS & capex
Marvell is an AI-data-center pure-play in fact now: data center is 76% of revenue and growing ~27% YoY (Q1 FY27), with the comms/other leg (+29% YoY) re-accelerating off a cyclical trough. The market's question is mix: management raised the FY27 interconnect (optical) growth guide to >70% YoY from 50%, and the FY28 total-revenue target to $16.5B, while the custom-ASIC leg faces pricing pressure and the Trainium3 design loss. Non-GAAP gross margin (58.9% Q1 FY27) drifts down as lower-margin custom silicon scales, but operating leverage holds non-GAAP operating margin at ~35%. FCF (~$1.67B TTM) funds a $5B buyback authorization plus a $1B ASR; the dividend is a token $0.24/yr.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~41¢ is cost of goods and ~24¢ operating expense, leaving ~35¢ of operating profit (~30¢ net).
Revenue trend
Margins
down — was 59.0% in Q4 FY26; custom-ASIC mix dilutes
flat/down vs 51.7% Q4 FY26
up — operating leverage on revenue scale
up off prior losses; wide GAAP/non-GAAP gap from SBC + amortization
up
up with revenue and light capex
COGS structure
COGS is dominated by TSMC wafer + advanced-packaging (CoWoS/chiplet) costs, HBM/substrate and test/assembly, plus acquisition-related inventory step-up amortization. Custom-ASIC programs carry materially lower gross margin than optical-DSP/IP product, so rising custom-silicon mix is the structural drag on the 58.9% non-GAAP GM (Q1 FY27). Memory/substrate inflation and CoWoS allocation are the swing variables.
Capex
Light, fabless: $155.7M in Q1 FY27 (~6.4% of rev), ~$391M TTM (~4.8%). Funds test/lab equipment, design tools (EDA), tape-out masks, and IP for advanced nodes (3nm/2nm) — not fabs. Heavy node + mask costs are expensed/amortized rather than capitalized, so reported capex understates true program investment.
Latest earnings
Beat — revenue $18M above guidance midpoint; non-GAAP EPS $0.80 vs guide; stock rose on the raised FY27/FY28 outlook (fact)
Q2 FY27 revenue $2.7B ±5% (+35% YoY), non-GAAP EPS $0.93 ±$0.05; FY27 interconnect growth raised to >70% YoY; FY28 revenue target raised to $16.5B (fact)
- Data center revenue
- $1.832B, 76% of total, +27% YoY (fact, Q1 FY27)
- Comms & other revenue
- $585.1M, 24% of total, +29% YoY (fact, Q1 FY27)
- Non-GAAP gross margin
- 58.9% (fact, Q1 FY27)
- Capital returned
- $200M buyback in Q1 FY27; ~$2.2B returned in FY26 (fact)
Growth drivers
- Optical interconnect ramp — 1.6T PAM4 (Ara, first 3nm 1.6T DSP) + 800G; management raised FY27 interconnect growth to >70% YoY from 50% (fact)
- Custom XPU/ASIC programs — a new tier-1 XPU ramps in FY28 plus 10+ 'XPU-attach' programs reaching volume; custom silicon to more than double in FY28 (fact)
- 51.2T Ethernet scale-out switches (Teralynx) + datacenter interconnect (coherent DCI) for AI-cluster scale-out
- Hyperscaler capex super-cycle — Amazon/Microsoft/Google/Meta AI infrastructure buildout pulls both legs
- Nvidia NVLink Fusion partnership + $2B Nvidia investment (Mar 2026) integrating Marvell custom/networking into Nvidia fabric (fact)
- M&A bolt-ons in photonics — Celestial AI, XConn, Polariton, Inphi base — deepening scale-up optics and photonic-fabric IP
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2026-03-11. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’26
Bull & bear
The market is conflating Marvell's commoditizing, re-competed custom-ASIC leg with its durable, mispriced optical-I/O and interconnect franchise. The interconnect rent is structural and accelerating — management raised FY27 interconnect growth to >70% YoY (from 50%) — and at ~70% PAM4 DSP share Marvell collects toll on every AI cluster regardless of which XPU wins.
- Interconnect is the real asset: FY27 optical/interconnect growth raised to >70% YoY from 50% (fact), driven by 1.6T (Ara, first 3nm 1.6T DSP) and 800G — high-margin, share-led, and socket-agnostic to the GPU/ASIC war
- Two-leg AI play with FY28 revenue guided to $16.5B (fact); custom silicon more than doubles in FY28 on a new tier-1 XPU + 10+ XPU-attach programs — even after the Trainium3 loss the pipeline grows
- Capex-light, FCF-rich model (~$1.67B TTM FCF, ~35% non-GAAP op margin) funding a $5B buyback + $1B ASR — compounding per-share even at scale
- Duopoly economics: Broadcom + Marvell control ~95% of custom-ASIC co-design; only two vendors can sell the full interconnect stack (DSP + 51.2T switch + DCI + SerDes)
- Nvidia's $2B investment + NVLink Fusion validate Marvell as the interconnect partner of record inside the dominant AI fabric (fact, Mar 2026)
Both legs face credible margin and share pressure at a valuation (~61x forward earnings, ~$242B cap) that prices flawless execution. Broadcom out-scales the ASIC business, Alchip/GUC undercut it on design wins (Trainium3 already lost), and Broadcom's own 800G/1.6T DSPs attack the optical 'rent' the bulls call durable.
- Custom-ASIC incumbency is not sticky — Marvell designed Trainium2 but lost Trainium3 to Alchip on a monolithic-die/cost basis (fact); each generation is re-competed, capping pricing power
- Broadcom holds ~70%+ of custom-ASIC design services and is now shipping competitive 800G/1.6T DSPs — squeezing Marvell on both legs simultaneously
- Mix shift to custom silicon structurally dilutes gross margin (non-GAAP GM 58.9%, down from 59.0%); the 'high-margin interconnect rent' shrinks as a share of the mix
- Valuation leaves no cushion: ~61x forward P/E, GAAP EPS only $0.04 in Q1 FY27 — any AI-capex digestion or program slip de-rates the multiple hard
- Hyperscaler concentration + in-sourcing risk — customers can pull design IP in-house or dual-source, and 76% of revenue rides a few buyers' capex cycles
What it is worth
Forward P/E + EV on the FY28 $16.5B revenue target, cross-checked vs AVGO. Trades ~61x NTM EPS / ~$242B cap (fact, Jun 2026). Bull case capitalizes the durable, high-margin interconnect 'rent' separately from the lower-margin, contestable custom-ASIC leg (sum-of-the-parts).
$110–160
AI-capex digestion or further ASIC share loss de-rates the ~61x multiple toward a cyclical-semi range (Street low $110)
~$260–300
FY28 target broadly met but custom-ASIC margin dilution + Broadcom DSP entry cap the multiple
$350–385 (Stifel $350
KeyBanc $385, fact) — FY28 $16.5B hit, interconnect rent holds, multiple sustained on duopoly optics
Estimate, not advice: priced for flawless AI-capex execution — the thesis edge is whether the optical leg's >70% FY27 growth (fact) is durable enough to offset ASIC commoditization/Trainium3-type losses.
SWOT
Strengths
- #1/#2 optical-DSP franchise (~70% PAM4 DSP share, fact) — Inphi-derived 800G/1.6T (Ara, first 3nm 1.6T) is the durable, high-margin 'interconnect rent' the bull thesis prizes
- One of only two credible custom-ASIC co-design houses (with Broadcom) — anchored by AWS Trainium2 + Microsoft Maia; ~18 cloud design wins
- Fabless, capex-light (~5% of revenue) with ~20% FCF margin and ~35% non-GAAP operating margin (fact, Q1 FY27)
- Broad data-infrastructure IP stack — DSP + switching (51.2T) + DCI + SerDes — lets Marvell sell the whole AI-cluster interconnect, not one chip
- Nvidia $2B investment + NVLink Fusion partnership embeds Marvell in the incumbent's roadmap (fact, Mar 2026)
Weaknesses
- Customer concentration — a handful of hyperscalers drive the data-center 76% of revenue; one program slip moves the model
- Custom-ASIC gross margin dilutes the corporate non-GAAP GM (58.9%, sliding from 59.0%) as that mix grows
- Wide GAAP/non-GAAP gap — Q1 FY27 GAAP EPS $0.04 vs non-GAAP $0.80 — heavy SBC and acquisition amortization
- Net-debt balance sheet (~$1.1–1.4B net debt) from the $10B Inphi deal and serial M&A, unlike net-cash peers
- Design wins are re-competed each generation — Trainium3 loss to Alchip shows incumbency is not sticky at the socket level (fact)
Opportunities
- 1.6T optical ramp + emerging co-packaged/linear-drive optics as AI clusters scale beyond copper reach
- FY28 $16.5B revenue target — custom silicon more than doubling plus 10+ XPU-attach programs to volume (fact)
- New tier-1 XPU program ramping FY28 diversifies beyond AWS/Microsoft
- Photonic-fabric / scale-up optics via Celestial AI + Polariton — TAM expansion into chip-to-chip optical I/O
- Comms/enterprise-networking + carrier recovery off a cyclical trough (+29% YoY in Q1 FY27, fact) adds a second growth leg
Threats
- Broadcom — ~70%+ of custom-ASIC design services and now shipping competitive 800G/1.6T DSPs, attacking both Marvell legs
- Alchip / GUC (Taiwan) winning hyperscaler sockets on cost/monolithic-die designs — Trainium3 already lost (fact)
- Hyperscalers in-sourcing more design IP, compressing the co-design value Marvell captures per program
- AI-capex digestion / cycle turn — extreme valuation (~61x fwd P/E) leaves no margin for a demand air-pocket
- Optical-DSP disruption — linear-drive/LPO and co-packaged optics could bypass standalone DSP sockets over time
Moats, dependencies & bottlenecks
Moats
~70% PAM4 DSP) leads each node transition (first 3nm 1.6T), but Broadcom's 800G/1.6T entry and LPO/CPO threaten long-term The core of the bull thesis: socket-agnostic 'interconnect rent' on every AI cluster
re-competed each generation; Trainium3 lost to Alchip (fact) Real but contestable; Broadcom has ~70%+ of design services
Full interconnect stack (DSP + 51.2T Ethernet switch + coherent DCI + SerDes) few vendors sell the whole AI-cluster scale-out/scale-up fabric Cross-sell and architectural lock-in across an AI pod
chiplet/CoWoS) table stakes shared with Broadcom/Alchip/GUC Necessary, not differentiating; the chiplet bet lost the Trainium3 bakeoff
storage, automotive) Diversification ballast, low share of the AI-driven thesis
Dependencies
Sole leading-edge foundry for 3nm/2nm; CoWoS allocation gates every AI program. Single point of failure
76% of revenue is data center, concentrated in a few buyers; any capex digestion or in-sourcing hits hard
Programs re-competed each generation; Trainium3 already lost to Alchip (fact)
Design-flow dependency; Nvidia-Synopsys loop and rising license costs
SK Hynix/Samsung/Micron HBM and substrate availability/pricing swings ASIC COGS
Entire thesis rides the AI-capex super-cycle; semiconductors are historically cyclical
Advantages
- Optical-DSP share leadership (~70% PAM4) — toll-taker on AI-cluster bandwidth regardless of which XPU/GPU wins
- Only Marvell + Broadcom can sell the full interconnect stack (DSP + 51.2T switch + DCI + SerDes) at scale
- Capex-light fabless model — ~5% capex/revenue, ~20% FCF margin, ~35% non-GAAP operating margin (fact)
- First-mover on each node (first 3nm 1.6T DSP, 'Ara') keeps a generation lead in the highest-value tier
- Nvidia $2B investment + NVLink Fusion designation as interconnect partner of record (fact, Mar 2026)
Weaknesses
- Custom-ASIC incumbency is contestable — Trainium3 lost to Alchip; design wins re-competed each generation (fact)
- Gross-margin dilution from rising custom-silicon mix (non-GAAP GM 58.9%, down from 59.0%)
- Wide GAAP/non-GAAP gap — Q1 FY27 GAAP EPS $0.04 vs non-GAAP $0.80 on heavy SBC + amortization
- Net-debt balance sheet (~$1.1–1.4B) from $10B Inphi + serial M&A, unlike net-cash AVGO/NVDA
- Extreme valuation (~61x fwd P/E) leaves no margin for error on a cyclical, concentrated revenue base
- Customer + foundry concentration — a few hyperscalers and a single foundry (TSMC) control the outcome
Bottlenecks
- TSMC advanced-node + CoWoS/advanced-packaging capacity allocation — caps how fast both ASIC and optical can ramp
- HBM and substrate supply for custom-ASIC programs — gates accelerator shipments
- Custom-ASIC mix dilutes corporate gross margin (non-GAAP GM 58.9%, sliding) — a structural ceiling on profitability as that leg grows
- Design-win cadence — each hyperscaler generation is a bakeoff; losing one (Trainium3) removes a multi-year revenue annuity
- Engineering/NRE capacity — custom programs are people-intensive; pipeline of 10+ XPU-attach programs strains design resources
Top signals & trends
Top signals
Core of the bull thesis — the mispriced, durable rent is accelerating (fact, Q1 FY27 call)
Validates the bear's 'ASIC is commoditizing/re-competed' view; watch for Trainium4 and Maia-next outcomes (fact)
Pipeline of new tier-1 XPU + 10+ XPU-attach offsets the Trainium3 loss (fact)
Watch whether custom-ASIC mix pulls GM below ~58% — tests the 'high-margin rent' claim
Optical share defense is the whole thesis; any DSP share loss to Broadcom is the key tell
76% of revenue rides their capex; the macro swing factor for the cycle
Trends
Directly feeds the interconnect leg the bull thesis is built on
Expands Marvell's ASIC TAM but each socket is re-competed (Alchip/GUC/Broadcom)
Near-term tailwind for DSP/optics; LPO/CPO could eventually bypass standalone DSP sockets
Attacks Marvell's highest-margin moat on both legs
Allocation favors scale players but caps ramp and pressures COGS
Second leg re-accelerating (+29% YoY Q1 FY27), reduces pure data-center dependence
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Sole leading-edge foundry (3nm/2nm) + CoWoS advanced packaging for all ASIC/optical programs
EDA design tools + IP for chip design flow
EDA tools + IP, second design-flow dependency
OSAT — assembly/test/packaging partner
HBM + memory/substrate supply for custom-ASIC programs (Micron US-listed)
Lasers/optical components paired with Marvell DSPs in transceiver modules (Coherent = COHR)
Largest custom-ASIC customer — Trainium2 in volume; Trainium3 lost to Alchip (fact)
Maia AI-accelerator custom-silicon program
Axion Arm CPU work; in talks on custom AI inference silicon
Custom DPU / data-processing silicon program
NVLink Fusion partner integrating Marvell custom/networking; also an investor (fact)
Networking OEMs + optical-module vendors (Arista = ANET) buy Marvell DSPs/switches
The structural rival on BOTH legs — ~70%+ of custom-ASIC design services (Google TPU, Meta MTIA, OpenAI) and now shipping competitive 800G/1.6T DSPs
Merchant-GPU + NVLink/Spectrum networking competes for AI-cluster interconnect; also a $2B Marvell investor/partner via NVLink Fusion — coopetition
Taiwan ASIC design house (TPE:3661) — won AWS Trainium3 from Marvell on a monolithic-die/cost design; the immediate custom-silicon share threat
Taiwan ASIC/TSMC-aligned design house (TPE:3443) — competes for hyperscaler custom-silicon sockets
AEC/active-copper + SerDes/DSP for AI clusters — competes at the edge of Marvell's interconnect (scale-up copper vs optical)
Connectivity (retimers, PCIe/CXL, AEC) for AI fabric — overlaps Marvell's scale-up interconnect