Layer 1 · MRVL — Marvell Technology
One-line thesis
Marvell is the second-largest custom AI ASIC design house, the principal silicon partner on AWS Trainium 3, and a Microsoft Maia design win speculatively priced in — and the May 27 print into a stock that is +135% in 10 weeks is the asymmetric binary of the entire AI semiconductor cohort.
What Marvell physically does
Marvell occupies the design-services niche immediately adjacent to Broadcom — co-designing custom AI accelerators for hyperscaler and large-enterprise customers, plus a substantial business in electro-optics (the silicon photonics and PAM4 DSP chips that drive 800G and 1.6T pluggable transceivers across every AI data-center optical link). The flagship AI program is AWS Trainium — Marvell is the principal design partner on Trainium 2 (in production) and Trainium 3 (ramping H2 2026 for the Anthropic Rainier supercluster build-out). The Microsoft Maia 100 / 200 / 300 series is rumoured to use Marvell on at least one generation of design service. The Google Axion ARM CPU also has Marvell-touched IP though that's a smaller line.
The custom-silicon design-services business at Marvell is structured around five technical pillars: high-speed SerDes (the IO links that move data on and off the chip — typically 112G or 224G per lane at AI generation), advanced packaging integration (CoWoS-S and CoWoS-L interface design at TSMC), high-bandwidth memory interface IP (the controllers that connect the accelerator die to its adjacent HBM stacks), networking-on-chip fabric design (the internal data-movement structure that connects compute cores to memory), and the physical implementation expertise (place-and-route at 3nm and 2nm nodes, signoff for advanced packaging). Of these, the SerDes IP and the advanced-packaging integration are the principal moats — both are multi-decade IP libraries that hyperscaler internal teams cannot easily replicate.
In addition to custom AI silicon, Marvell makes the optical DSPs that sit inside Lumentum and Coherent transceivers — the 100G / 200G PAM4 chips that drive 800G transceivers, and the next-generation 200G-per-lane silicon that enables 1.6T pluggables for AI cluster spine connectivity. The electro-optics franchise was built principally through the 2021 acquisition of Inphi (the merchant PAM4 DSP leader) and the 2022 acquisition of Innovium (cloud-grade switching silicon). Together these acquisitions gave Marvell a $2-3B annual revenue line in optical and networking silicon that operates structurally above the historical Marvell margin profile and is pulled by the same AI data-center capex cycle that drives the custom-silicon design-services revenue. This positions Marvell as a Layer-1 silicon company that simultaneously captures Layer-4 (optical connectivity) economics.
The data-center storage controller business — once Marvell's bread and butter — has compressed to a single-digit revenue contributor as the AI custom-silicon line has scaled. The enterprise networking line (the Prestera switching silicon for data-center top-of-rack switches) is the third-largest segment but is also shrinking as a percentage of revenue. The shift in the mix is the principal driver of the equity re-rate: Marvell five years ago was a storage / enterprise networking company; Marvell today is an AI custom-silicon and electro-optics company.
For an AI-stack investor: AWS Trainium volume is the swing variable. If Anthropic's Rainier cluster (1m+ Trainium 3 chips through 2027) builds out on AWS's current timeline, Marvell's AI revenue line approximately doubles between FY26 and FY28. The May 16 AMD cross-investment disclosure ($6.5M for an equity stake) is the second-order validation — AMD is signalling that Marvell's optical-DSP technology is part of the MI400-generation networking roadmap, and the strategic interpretation matters more than the dollars.
The financial print
Q4 FY26 (reported March 6 2026, quarter ended Feb) printed a beat across rev and EPS with the AI custom silicon and electro-optics businesses both contributing, and the guide was raised on AWS Trainium 3 ramp visibility. The market reaction has been dramatic: the stock is +135% over the trailing three months, +22% in the trailing month, and just touched an all-time high in mid-May. Goldman raised its PT to $125 from $100 on May 15 — and the stock at $181 is now well above the average sell-side PT, which sits closer to $135-145. The buy-side is ahead of the sell-side by a wide margin.
The fundamental backdrop into Q1 FY27: data-center revenue at Q4 ran approximately $1.4B (+78% YoY), with custom AI silicon contributing roughly $700-800M of that line and electro-optics contributing approximately $400-500M. Storage controllers and enterprise networking were the residual contribution, both declining in absolute dollar terms as the AI lines have scaled. Gross margin came in at 62.5% non-GAAP — pressured from the prior-year 65% level by mix-shift toward higher-volume lower-margin custom-silicon products. Operating margin held at approximately 28% non-GAAP, with the company maintaining its long-term targets in the 32-35% range as scale economics improve through 2027.
Q1 FY27 prints May 27 after-hours. Consensus sits at approximately $2.1B revenue and roughly $0.85 of EPS, but with a +135% rally into the print the bar is clearly higher than consensus — anything less than a sizeable beat-and-raise risks a sell-the-news reaction. The May 21 / May 27 timing window matters because of the cross-vendor signal embedded in the May 16 disclosure that AMD invested $6.5M in Marvell — small dollars, but the optical-link cross-validation matters strategically.
The positioning into the print is the principal tactical risk. Marvell has been one of the top three best-performing US semiconductor names of 2026 year-to-date; institutional positioning is materially long versus historical baselines; the consensus PT cohort is well below spot. This is the classic crowded-long setup into an earnings event — the binary is between (a) a clean beat-and-raise that produces $200-210 and a fresh leg higher, or (b) an in-line print that produces a 15-25% pullback regardless of the underlying fundamentals.
Customer mix
AWS is the single largest customer at approximately 25-30% of revenue, with Trainium custom-silicon plus electro-optics revenue combined. Microsoft is rising — currently 10-15% of revenue, potentially the largest customer by FY28 if the rumoured Maia design wins land. The optical-DSP business is broadly distributed across the transceiver-module makers (Coherent, Lumentum, Innolight, Sumitomo Eudyna, Hisense Broadband) and is structurally less concentrated. Customer concentration in the AI custom-silicon line is currently the largest financial risk, and the principal reason Marvell trades at a discount to Broadcom on forward multiples.
Competitive context
Broadcom is the dominant competitor in custom AI silicon — larger team, longer design library, more packaging integration. Marvell's structural advantage is the electro-optics adjacency (Broadcom owns SerDes and DSP but Marvell owns more of the merchant optical-DSP market) and the AWS Trainium incumbency that is now multi-generational. Alchip and GUC are the Taiwanese ASIC houses that compete on the back-end of custom-silicon design but typically lack the front-end architectural co-design depth Broadcom and Marvell offer. In optical DSPs, the competition is Inphi (which Marvell acquired in 2020), MaxLinear (smaller-scale), and Broadcom's own optical-DSP line — Marvell maintains roughly 40-50% merchant share at the leading 100G/200G PAM4 generation.
The AMD cross-investment disclosure on May 16 is significant. AMD does not invest in semiconductor companies frivolously; the most likely interpretation is that AMD is validating Marvell's optical-DSP technology for MI400-generation interconnect and is publicly signalling commitment to a non-NVIDIA networking option. That single data point is worth a leg of the bull case.
Terminal risk
The terminal risk is the same as Broadcom but more acute: a single major customer pulling a design generation in-house. AWS has the internal Annapurna design team that originated Graviton, and the long-tail question of whether Trainium 4 or 5 stays at Marvell or moves to AWS-internal is the multi-year variable. If AWS pulls Trainium 4 in-house, Marvell's AI custom-silicon revenue takes a multi-quarter air-pocket that cannot be filled by Microsoft Maia ramp alone. The current sell-side bull cases mostly assume Trainium remains at Marvell through 2028; that assumption is doing a lot of work.
Bull case
The base case is FY27 revenue of $9.0-9.8B (FY26 prints around $7.0B), AI custom-silicon revenue at $3.5-4.0B (up from $2.2B in FY26), electro-optics holding $2.0B+, and non-GAAP EPS at $4.20-4.80 versus FY26 around $3.10. At 35x forward — generous but not unreasonable for the cohort — that prints the stock at $150-170, roughly in line with spot but with multi-year compounding.
The upside case is Microsoft Maia 200 design win confirmed in 2026, Trainium 3 ramp accelerating on Anthropic Rainier build, and a third hyperscaler design win (Meta on a future MTIA generation, or Apple inference) lifting the AI custom-silicon line above $5B by FY28. That case prints the stock at $230-260 in 2027 on FY28 EPS of $6.00-7.00.
The AWS Trainium 3 ramp is the principal volume driver. Anthropic's Project Rainier — disclosed in 2024 as a 1m+ Trainium-chip supercluster build across multiple AWS regions — is the largest single Trainium 3 commitment and the principal reason Marvell's AI custom-silicon revenue line is doubling between FY26 and FY28. AWS volume commentary through 2025-26 has been consistent: Trainium 3 production capacity is sold out for 2026, deployments are running ahead of plan, and Trainium 4 design work is already underway with Marvell on a continuing-engagement basis. The continuity of the Marvell-AWS relationship through multiple generations is the structural advantage that distinguishes Marvell from a pure design-services vendor.
Gap / bear case
Three concerns the market may be glossing over. First, the +135% in 10 weeks has compressed the risk-reward — consensus PT of $135 sits 25% below spot, meaning the marginal sell-side analyst would need to raise PTs above $200 for the stock to look reasonable on technicals. Second, the Microsoft Maia design win is rumour, not confirmed disclosure; if Marvell does not announce the win during 2026, the stock loses a third of its rally driver. Third, the gross margin trajectory — historically 65-66% non-GAAP — has been compressing as the lower-margin custom-silicon revenue grows; FY27 GM consensus is closer to 62-63%, which is a real headwind even if revenue compounds at 35%+.
Optionality
Microsoft Maia design-win confirmation during 2026 — the single biggest discrete option. Apple inference-silicon design-services partnership (rumoured for late 2026 / 2027). And the slow-burn option: 1.6T-class optical DSP volume ramp through 2027-28 as the 800G generation gives way and Marvell maintains share leadership at the leading-edge generation. Each of these is a separate $300-600M annual revenue layer that compounds the AI thesis.
The trade
MRVL — BUY 7.5/10. Entry: this is the tactically toughest entry in Tier A. Do not chase above $185. Wait for either (a) a clean beat-and-raise on May 27 with the stock breaking out above $200, in which case add on the breakout, or (b) a sell-the-news pullback to $150-160 post-print, which is the better R/R. Position size: 1.5-2.5% of NLV — half of NVDA / AVGO sizing because of the customer concentration and the recent rally. Stop: daily close below $145 (the early-March base before the rip). Catalyst date: May 27 print after-hours; any Microsoft Maia disclosure; Q2 FY27 print late August. Trim/exit triggers: AWS Trainium 3 ramp slipping into 2027 instead of H2 26; gross margin guide below 62%; AMD investment narrative reversed. Conviction: real AI custom-silicon franchise, real momentum, but a positioned-long set-up into the print. Treat this as the asymmetric-but-mispositioned trade — own it small, scale on better entry, do not anchor to the +135% rally as the new baseline.