The other custom-silicon king — and the one NVIDIA decided to tax instead of fight. From a $59 low to all-time highs on the AI-factory buildout.
Marvell is the #2 merchant designer of the custom AI chips (XPUs) hyperscalers use to build their own accelerators — plus the optical and interconnect “plumbing” that moves data between them. The data center is now 76% of revenue, management guides FY26’s $8.2B toward ~$11B this year and ~$15B next, and NVIDIA just put in $2B to plug Marvell silicon into NVLink Fusion. Jensen Huang reportedly called it the “next trillion-dollar company.” The catch: the stock round-tripped from a $59 low to ~$220 all-time highs, trades ~70x trailing earnings, carries a 2.3 beta, and competes with a bigger, fatter-margin Broadcom while leaning on a handful of concentrated hyperscaler programs. Below: the two engines, the custom-silicon math, the NVIDIA twist, the numbers, and the honest risk.
Jensen crashed the stage and made connectivity the headline of the show. The thesis got the biggest possible megaphone — four days after a record quarter.
On June 2 in Taipei, CEO Matt Murphy delivered Marvell’s first-ever Computex keynote — “The Future of AI Scaling Depends on Connectivity” — and NVIDIA’s Jensen Huang appeared as a surprise guest for about ten minutes, the most-talked-about moment of the hour. Important framing for traders: this was a strategy and positioning event, not a product launch — no new chip was unveiled and no fresh financial guidance was given. What it delivered was the connectivity narrative and the NVIDIA alliance, broadcast to the entire industry on the marquee AI stage of the year.
The Most Important Parts
| Moment | What Was Said |
|---|---|
| The core thesis | Connectivity is the next AI bottleneck — as compute outpaces networking, the next phase of AI scaling hinges on moving data across accelerators, racks, campuses and data centers at higher bandwidth, lower latency and lower power |
| “Switzerland of the industry” | Murphy framed Marvell as neutral and uniquely connectivity-weighted (vs pure compute): it works with everybody — every hyperscaler and NVIDIA alike |
| Jensen on stage | Validated the NVLink Fusion partnership + NVIDIA’s $2B investment; framed NVLink Fusion as “fusing” NVIDIA’s and Marvell’s platforms, letting cloud providers pair custom / semi-custom chips with NVIDIA hardware into a heterogeneous data center |
| The agent angle | Jensen: AI agents are disaggregated by nature, which drives connectivity demand; NVIDIA’s Vera Rubin platform (announced in production at the show) is built to run agents |
| “Next trillion-dollar company” | Jensen repeated his endorsement of Marvell — now from the Computex stage |
| Optics moves to the rack | Murphy’s analogy: like when 10Gb was cutting-edge and optics was a telecom-only technology, the same wave is now bringing optics into the rack — the copper→optical shift underpinning the optical franchise (§ 04) |
The picks-and-shovels of the AI factory. Custom silicon plus the optical plumbing that connects it.
Marvell Technology is a data-infrastructure semiconductor company that has reinvented itself into one of the two great merchant arms-dealers of the AI buildout. It does two things that matter: it co-designs custom AI accelerators (XPUs) for hyperscalers who want their own chips instead of buying only NVIDIA GPUs, and it builds the high-speed interconnect — optical DSPs, silicon photonics, switches — that moves data between all those chips. The data center is now ~76% of revenue, and the stock has gone from a $59 low to ~$220 all-time highs on the strength of it.
| Item | Detail |
|---|---|
| Ticker | NASDAQ: MRVL · Santa Clara, CA · founded 1995 · CEO Matt Murphy |
| Price | ~$220 · all-time high · 52-wk $59–$225 |
| Market Cap | ~$185B · ~875M shares |
| Revenue | FY26 $8.2B → FY27e ~$11B (+30%) → FY28e ~$15B |
| Mix | Data center ~76% · rest = enterprise networking, carrier, consumer, auto (cyclical) |
| Valuation | ~70x TTM EPS · ~46x forward · 2.3 beta |
| The rival | Broadcom (AVGO) — bigger, higher-margin, owns Google TPU |
| The endorsement | NVIDIA invested $2B · Jensen reportedly called it the “next trillion-dollar company” |
Custom compute and connectivity. Both ride the same AI-factory capex wave.
Strip away the segments and Marvell is really two businesses bolted to one secular trend. Engine one is custom silicon — designing bespoke XPUs and the chips that surround them for cloud giants. Engine two is interconnect — the optical and electrical components that let tens of thousands of accelerators behave like one computer. The legacy pieces (enterprise networking, carrier, consumer, automotive) still exist, but they’re cyclical and shrinking as a share of the story; data center is the engine the market pays for.
| Segment | What It Is | Read |
|---|---|---|
| Data center (~76%) | Custom XPUs, optical DSPs, silicon photonics, switches (Teralynx), CXL, retimers, DCI | The growth engine · AI-driven · all the upside lives here |
| Enterprise networking | Ethernet switching/PHYs for enterprise gear | Cyclical · recovering off trough |
| Carrier infrastructure | Wired/wireless infra silicon for telcos | Cyclical · slower-growth |
| Consumer / auto-industrial | Storage controllers, auto Ethernet, etc. | Lumpy · non-core to the AI thesis |
The hyperscalers want their own chips. Marvell is one of only two places they can go.
The biggest structural tailwind in semis right now is hyperscalers designing their own AI accelerators to reduce dependence on NVIDIA’s pricing and supply. They don’t do it alone — they co-design with a merchant partner, and that market is essentially a duopoly: Broadcom and Marvell. At its investor day Marvell raised its data-center TAM to $94B by 2028 (from $75B) and laid out a credible path to ~20% share — roughly $18.8B of data-center revenue, up from ~$4B in 2024.
| Metric | Detail |
|---|---|
| Design sockets | 18 multigenerational custom wins — 5 XPU designs + 13 XPU-attach (retimers, CXL, CPO, co-processors) |
| Customers | AWS, Microsoft, Google, Meta — plus emerging names (xAI, Tesla Dojo, sovereign AI) beyond the “big four” |
| Pipeline | 50+ additional opportunities · ~$75B potential lifetime revenue |
| Custom XPU TAM | ~$40.8B by 2028 · ~47% CAGR |
| XPU revenue | ~$1.5B in FY26 · guided to roughly double by FY28 |
| Share target | ~20% of data-center TAM by 2028, up from <5% a few years ago |
NVIDIA decided to tax Marvell, not fight it. And the optical franchise may be the real crown jewel.
In April 2026, NVIDIA invested $2 billion in Marvell and struck a partnership around NVLink Fusion — the rack-scale fabric that lets third-party silicon plug into NVIDIA’s proprietary interconnect. The clever read: Marvell’s custom XPUs displace NVIDIA GPUs, so rather than fight that, NVIDIA turned it into a toll booth — every custom chip Marvell designs now still pulls NVIDIA platform revenue. The stock jumped ~13% on the news. It’s validation and a hedge in one move.
The Optical / Interconnect Franchise
Marvell’s interconnect business — anchored by the Inphi acquisition — is arguably its most defensible asset, and it sits dead-center in your photonics thesis. As AI clusters scale, the bottleneck shifts from compute to moving data, and the industry is migrating from copper to optical. Marvell sells the picks for that shift:
| Product | Why It Matters |
|---|---|
| Optical DSPs | PAM4 & coherent/coherent-lite DSPs — the brains of optical modules · category leader |
| Silicon photonics / CPO | Co-packaged optics moving optical right next to the switch/XPU die |
| Teralynx T100 | Industry-first 102.4 Tbps switch for AI/cloud — the announcement drove a ~24% single-move pop |
| Celestial AI (acquiring) | Photonic Fabric for optical scale-up · ramps to a ~$500M run-rate by Q4 FY28, ~$1B by Q4 FY29 |
| LPO / AEC / PCIe retimers | The connective tissue of the rack — high-attach, recurring content |
A record quarter into all-time highs. Growth is real; so is the multiple.
Marvell reported Q1 FY2027 on May 27, 2026, and it was clean: record revenue, a beat, and raised outlooks. The question isn’t whether the business is growing — it plainly is — it’s whether ~70x trailing earnings already pays for several years of that growth in advance.
| Metric | Q1 FY27 / Guidance |
|---|---|
| Revenue | $2.418B · +28% YoY · record |
| Non-GAAP EPS | $0.80 · beat $0.79 · +29% YoY |
| Data center | ~76% of revenue · the growth driver |
| Operating cash flow | Record $639M |
| Q2 FY27 guide | Revenue ~$2.598B · EPS ~$0.91 (continued sequential growth) |
| Gross margin | ~51% blended — custom silicon dilutes vs the higher-margin legacy/optical |
| Valuation | ~70x TTM EPS · ~46x forward · ~$185B cap |
Trade ideas like this, before they hit the timeline.
Join Discord →Why it can still compound from here.
- Duopoly in custom silicon — hyperscalers want their own chips and there are really only two merchant partners; Marvell is one of them.
- Visible, multi-year growth — $8.2B → ~$11B → ~$15B guided, 18 multigenerational sockets, ~$75B lifetime pipeline.
- Expanding TAM — data-center TAM raised to $94B by 2028; XPU TAM ~$40.8B at ~47% CAGR; 20% share target from <5%.
- Optical leadership — the Inphi-anchored interconnect franchise is the steadier, broad-based compounder levered to copper→optical, plus Celestial AI for scale-up photonics.
- The NVIDIA hedge — $2B investment + NVLink Fusion turns the GPU-vs-XPU tension into a partnership; Jensen’s “next trillion-dollar company” line is free marketing.
- Customer broadening — wins beyond the big four (xAI, Tesla Dojo, sovereign AI) start to dilute concentration risk over time.
- Real numbers — record revenue, EPS beats, record $639M operating cash flow · this is a profitable franchise, not a promise.
- Analyst conviction — ~44 analysts, overwhelmingly Strong Buy, with high targets toward $300.
Why all-time highs are a dangerous entry.
- Priced for perfection — ~70x trailing / ~46x forward bakes in years of growth and margin expansion; a single soft quarter can re-rate it hard.
- Customer concentration — the custom business leans on a handful of hyperscaler programs; losing or sharing one (an AWS or Microsoft socket) is a real, recurring fear — it’s what drove the 2025 drawdown to $59.
- Broadcom is bigger and fatter — AVGO runs ~77% gross margins vs Marvell’s ~51%, larger contracts, and the marquee Google TPU franchise · Marvell is the #2.
- Margin drag — custom silicon is structurally lower-margin; rapid XPU growth can dilute blended profitability even as revenue soars.
- Lumpy & cyclical — custom programs ramp and roll off in steps; the legacy segments add their own cyclicality · revenue can be jumpy quarter to quarter.
- High beta (2.3) — moves ~2x the market; in any AI-sentiment drawdown (recall the DeepSeek scare, the 2025 selloff) it falls faster than almost anything.
- Hyperscaler capex dependence — the entire thesis rests on cloud AI capex staying vertical; any digestion phase hits Marvell directly.
- Execution risk on ramps — the $15B FY28 number requires multiple design wins to ramp on time and at volume · slips compress the story.
Own the franchise, respect the entry. This is a buy-the-dip name, not a chase-the-high one.
MRVL is the rare AI-infra name that’s both a real business and a high-beta trade. The franchise deserves a place in an AI-infrastructure sleeve; the entry is the hard part, because buying a 2.3-beta stock at ~70x into all-time highs is how you end up underwater for a year even when the thesis is right. The discipline is to define your entries around weakness and your risk around the quarters.
The Catalysts to Watch
| Catalyst | Why It Matters |
|---|---|
| Quarterly prints (next ~Aug 2026) | Data-center growth rate + margins + guidance · the stock trades on the trajectory, not the beat |
| New XPU socket wins | Each multigenerational win extends the $75B pipeline · losses are the bear trigger |
| Hyperscaler capex commentary | AWS/Microsoft/Google/Meta capex guides are MRVL’s real demand signal |
| Optical ramps | Teralynx, CPO, Celestial AI — the steadier, higher-conviction growth leg |
| Broadcom read-across | AVGO results/commentary set the tone for the whole custom-silicon complex |
Great franchise, demanding price. Buy weakness, not the all-time high.
Marvell is the real thing — one of two merchant designers of the custom AI silicon hyperscalers are standardizing on, wrapped around the best optical/interconnect franchise in the group, with NVIDIA paying $2B to keep it inside the tent. The growth is visible and funded by actual record cash flow, not a story. But you’re being asked to pay ~70x trailing earnings for a 2.3-beta stock at all-time highs, against a bigger, fatter-margin Broadcom, on a thesis that leans on a concentrated set of hyperscaler programs and vertical cloud capex. None of that makes it a short — it makes it a franchise you accumulate on fear, not euphoria. The 52-week range from $59 to $225 is the whole lesson: the business compounds, but the stock hands you violent entry points if you’re patient. Own it for the connectivity-and-custom-silicon thesis; let the multiple and the next AI wobble give you the price.