Marvell Technology
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
Marvell is the number-two custom AI silicon house behind Broadcom and the share leader in high-speed optical DSPs; the Google warrant deal ties a hyperscaler to it for years, but custom sockets are won and lost each generation and four or five customers carry the business.
Marvell's edge is owning the electrical-to-optical boundary and co-designing custom chips with hyperscalers — real, but a narrower franchise than Broadcom's:
- Optical DSP Share Leader: Marvell, via the Inphi acquisition, holds roughly 60% of the 400G-and-above PAM4 DSP market against Broadcom's 30%-plus, and its 3nm Ara was the first 1.6T DSP to sample. Almost every pluggable optical module in an AI cluster needs one, which makes Marvell a toll on the copper-to-optics transition rather than a bet on one accelerator.
- Custom Silicon, Now Contracted: Custom XPUs and attach chips are co-designed over multiple generations, and the August 2026 Google agreement spans inference accelerators, NICs, storage and memory-interface controllers around the TPU. Google's warrant vests one tranche per $500M of custom revenue through FY2033, up to $120B. But Broadcom remains the larger custom partner across more hyperscalers, and Marvell has lost generations before.
- Concentration Is the Flip Side: A handful of hyperscalers drive most of data-center revenue, so a single program shifting to Broadcom, MediaTek, Alchip or an in-house team moves the whole company. The moat is depth with each customer, not breadth across many.
Moat Verdict
Marvell is a direct AI beneficiary: its strongest moat is the multi-generation custom-silicon embedding the Google warrant now contracts, backed by intact optical-DSP IP, bundling and talent. The vulnerability is concentration — Broadcom and in-house teams contest every socket — so the moat is as durable as the next design win.
70.9 resilient · 65.0 vulnerable · 80/20 = 69.8 · = 70
Open a moat to read its note.
Marvell sells silicon to system and hyperscaler engineers; there is no end-user interface to learn.
SerDes, PAM4 DSP and coherent-optics IP built since the Inphi acquisition is vendor-owned and hard to reproduce: Marvell holds roughly 60% of 400G-and-above PAM4 DSPs. Broadcom has the same capability at 30%-plus share, so this is a lead, not a monopoly.
Marvell does not derive advantage from controlling access to public data.
Leading-edge mixed-signal, SerDes and custom-ASIC teams at 3nm and 2nm are scarce, but Broadcom, MediaTek, Alchip and the hyperscalers' own groups hire from the same pool.
Marvell sells DSPs with matching TIAs and drivers, plus switches, NICs and custom attach chips, so optical module makers and hyperscalers can source a whole interconnect path from one vendor. Each part still faces a separate rival, so the bundle helps share more than it locks customers in.
Custom-silicon design data belongs to each hyperscaler customer; Marvell does not build or monetise a proprietary dataset.
No certification or licence regime ties customers to Marvell's silicon.
Interconnect chips follow industry standards; more customers do not make Marvell's parts more valuable to each other.
Custom chips are co-designed with each hyperscaler across generations, and the August 2026 Google agreement contractualises it: the warrant vests one of 240 tranches per $500M of Google custom revenue through FY2033, up to $120B. That is a named, multi-year embedding in a customer's silicon roadmap, though concentrated in a few programs.
Marvell's chips are components; no downstream system defers to them as a source of truth.
Leading-node mask sets and a multi-billion-dollar R&D budget spread across optics and custom programs keep small entrants out, but Broadcom runs a larger silicon and custom business, so Marvell has scale without a cost lead.
A B2B component vendor; buyers choose on specification and roadmap, not on a brand premium.
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
Marvell is the number-two custom AI silicon house behind Broadcom and the share leader in high-speed optical DSPs; the Google warrant deal ties a hyperscaler to it for years, but custom sockets are won and lost each generation and four or five customers carry the business.
Growth Score
Q2 FY2027 (reported August 27) printed record revenue of $2.739B (+37% YoY, +13% QoQ) and non-GAAP EPS of $0.94 (+40%). Data center was $2.17B, 79% of revenue and +46% YoY. Non-GAAP gross margin was 58.9%, down 50bp YoY as custom silicon grows in the mix. Q3 is guided to $3.15B ±5% (+15% QoQ, >50% YoY) and $1.05–$1.15 of non-GAAP EPS. Management raised FY2027 to about $12B (~45% growth) and FY2028 to $18B (~50% growth), with custom revenue expected to more than double in FY2028 and accelerate again in FY2029. On August 18 Marvell issued Google a warrant for up to 58.97M shares at $206.58, vesting one tranche per $500M of Google custom revenue through FY2033. Marvell also closed the Celestial AI (photonic fabric) and XConn acquisitions for next-generation scale-up interconnect.
Valuation Score
At $264.21 (September 30 close, ~$233B market cap on ~882M shares), Marvell sits between the $160 bear and $280 base, about 87% of the way from bear to base, for a valuation score of 68. That is ~38× the ~$6.9 FY2028 consensus EPS: the price already assumes most of the $18B outlook lands.
The Interconnect Specialist
Marvell's edge is owning the electrical-to-optical boundary and co-designing custom chips with hyperscalers — real, but a narrower franchise than Broadcom's:
- Optical DSP Share Leader: Marvell, via the Inphi acquisition, holds roughly 60% of the 400G-and-above PAM4 DSP market against Broadcom's 30%-plus, and its 3nm Ara was the first 1.6T DSP to sample. Almost every pluggable optical module in an AI cluster needs one, which makes Marvell a toll on the copper-to-optics transition rather than a bet on one accelerator.
- Custom Silicon, Now Contracted: Custom XPUs and attach chips are co-designed over multiple generations, and the August 2026 Google agreement spans inference accelerators, NICs, storage and memory-interface controllers around the TPU. Google's warrant vests one tranche per $500M of custom revenue through FY2033, up to $120B. But Broadcom remains the larger custom partner across more hyperscalers, and Marvell has lost generations before.
- Concentration Is the Flip Side: A handful of hyperscalers drive most of data-center revenue, so a single program shifting to Broadcom, MediaTek, Alchip or an in-house team moves the whole company. The moat is depth with each customer, not breadth across many.
Moat Verdict
Marvell is a direct AI beneficiary: its strongest moat is the multi-generation custom-silicon embedding the Google warrant now contracts, backed by intact optical-DSP IP, bundling and talent. The vulnerability is concentration — Broadcom and in-house teams contest every socket — so the moat is as durable as the next design win.
70.9 resilient · 65.0 vulnerable · 80/20 = 69.8 · = 70
Open a moat to read its note.
Marvell sells silicon to system and hyperscaler engineers; there is no end-user interface to learn.
SerDes, PAM4 DSP and coherent-optics IP built since the Inphi acquisition is vendor-owned and hard to reproduce: Marvell holds roughly 60% of 400G-and-above PAM4 DSPs. Broadcom has the same capability at 30%-plus share, so this is a lead, not a monopoly.
Marvell does not derive advantage from controlling access to public data.
Leading-edge mixed-signal, SerDes and custom-ASIC teams at 3nm and 2nm are scarce, but Broadcom, MediaTek, Alchip and the hyperscalers' own groups hire from the same pool.
Marvell sells DSPs with matching TIAs and drivers, plus switches, NICs and custom attach chips, so optical module makers and hyperscalers can source a whole interconnect path from one vendor. Each part still faces a separate rival, so the bundle helps share more than it locks customers in.
Custom-silicon design data belongs to each hyperscaler customer; Marvell does not build or monetise a proprietary dataset.
No certification or licence regime ties customers to Marvell's silicon.
Interconnect chips follow industry standards; more customers do not make Marvell's parts more valuable to each other.
Custom chips are co-designed with each hyperscaler across generations, and the August 2026 Google agreement contractualises it: the warrant vests one of 240 tranches per $500M of Google custom revenue through FY2033, up to $120B. That is a named, multi-year embedding in a customer's silicon roadmap, though concentrated in a few programs.
Marvell's chips are components; no downstream system defers to them as a source of truth.
Leading-node mask sets and a multi-billion-dollar R&D budget spread across optics and custom programs keep small entrants out, but Broadcom runs a larger silicon and custom business, so Marvell has scale without a cost lead.
A B2B component vendor; buyers choose on specification and roadmap, not on a brand premium.
Growth Analysis
Growth Drivers
Key Risk
Marvell's growth rests on a few hyperscaler programs. If a next-generation custom accelerator goes to Broadcom, MediaTek, Alchip or an in-house team, or if hyperscaler AI capex pauses in 2027–28, the FY2028 $18B outlook breaks. Falsifiable: FY2028 revenue below $15B, or custom revenue failing to double in FY2028, would mean the custom thesis is not on track. The Google warrant also dilutes holders by up to ~59M shares as revenue milestones vest.
Score Derivation
90.0 base + 2.7 trajectory − 10 risk = 83
Base 90 (25–35% CAGR, midpoint 30%) + 2.7 trajectory (custom silicon and electro-optics accelerating; enterprise, carrier and auto/industrial stable) + 0 margin (stable: non-GAAP gross margin 58.9%, down 50bp YoY as lower-margin custom grows, while operating leverage lifts EPS faster than revenue) − 10 high risk (four or five hyperscalers drive most of data-center revenue; one lost custom generation, as with past AWS programs, cuts the growth path sharply; same capex-cycle risk as CRDO) = 83
Price Scenarios (12–24 Months)
Valuation Analysis
The base of $280 is ~40× FY2028 consensus EPS of ~$6.9, or ~30× a ~$9 FY2029 if custom accelerates as guided — a premium multiple for a growth rate few semiconductor names match. The bear uses the low FY2028 estimate (~$5.30) at ~30×. Add on weakness toward the $200s; the warrant's $206.58 strike is a useful marker of where Google was willing to commit. ~$280.
Valuation Multiples
| Forward P/E (FY2028) | ~38× |
| Run-rate P/E | ~60× |
| PEG Ratio | ~1.3× |
| Price / Sales (FY2028) | ~13× |
Marvell is priced as a growth compounder, not a cyclical chip supplier. A PEG near 1.3× is reasonable only if the FY2028 outlook and the custom doubling both land; a lost program would leave a ~38× multiple on lower earnings.
Approximate figures as of September 30, 2026.
Where We Are vs Targets
Loading live price…
A custom program slips or moves to a rival and AI capex digests in 2027; FY2028 lands well short of $18B and the multiple compresses.
- A next-generation hyperscaler accelerator goes to Broadcom, MediaTek or an in-house team
- 1.6T optics ramp meets price competition from Broadcom and Chinese DSP entrants
- FY2028 EPS lands near the $5.30 low estimate and the stock trades at ~30×, giving about $160
FY2028 lands near $18B, custom revenue doubles as guided, and the market keeps paying a growth multiple.
- Q3 and Q4 FY2027 land near guidance, keeping FY2027 at ~$12B
- Google programs vest the first warrant tranches, confirming the custom ramp
- FY2028 EPS near $6.9 at ~40×, giving about $280
Custom accelerates again in FY2029, Celestial AI wins scale-up optical sockets, and Marvell is valued alongside Broadcom.
- A new hyperscaler custom XPU win beyond Google and the existing programs
- Celestial AI photonic fabric designed into a scale-up AI system
- FY2028 EPS near the $8.49 high estimate at ~47×, giving about $400