Record Revenue $1.895B (+63% YoY), Data Center +76%, Non-GAAP EPS $0.62 (+158%), 3nm Wafer + Advanced Packaging Capacity Secured for Custom XPU Follow-On, NVLink Fusion Partnership, Auto Ethernet Sale to Infineon for $2.5B Cash — Initiating at Hold

Coverage Context — Why Marvell, Why Now

Marvell has emerged over the past 24 months as one of the two principal merchant-silicon partners for hyperscale custom AI accelerators (the other being Broadcom). The thesis under Matt Murphy's tenure has been: pivot from a diversified analog/communications company into a data-center-first AI silicon platform. The pivot's signature moves include the 2019 Avera (custom design services) and Inphi (PAM DSPs / optical interconnect) acquisitions, the 2022 Innovium (cloud switching) acquisition, the divestment of Wi-Fi in 2024, and now the May 2025 Auto Ethernet divestiture to Infineon. The result: roughly 80%+ of Marvell's R&D and roughly three-quarters of company revenue now in data center. This Q1 FY26 print is the first quarterly disclosure post the Trump-administration tariff resolutions, post Auto Ethernet announcement, and just three weeks ahead of Murphy's June 17 "Custom Silicon Investor Event." We initiate coverage at Hold now because the company is at an inflection — fundamentals are strong but the multi-year visibility framework requires the June event to be properly anchored. Subsequent quarters in this backfill (Q2 FY26 → Q4 FY26) will track how that visibility evolves.

Results vs. Consensus

Q1 Scorecard

Year-over-Year Comparison

Quality-of-Beat Callout

Revenue Assessment

Revenue +63% YoY to $1.895B was driven primarily by data center (+76%) with enterprise networking + carrier infrastructure contributing meaningful sequential acceleration (+14% combined). The data center mix is now 76% of HoldCo (vs ~70% Q1 FY25), and management explicitly flagged that AI now represents the majority of data center revenue — implying AI is roughly 40%+ of total HoldCo at this point, with a near-term path to crossing 50% of HoldCo. Enterprise networking + carrier infrastructure recovery is the second concurrent positive: this is the fifth consecutive quarter of sequential recovery in those combined end markets, with the Q2 guide implying the recovery continues. Consumer (gaming) seasonal decline is expected; auto/industrial decline driven by lumpy industrial orders rather than structural softness.

Margin Assessment

Non-GAAP gross margin 59.8% — slightly below the 60.5% Q4 FY25 print and below the Street's 60.0% expectation. The compression vs the prior quarter is mix-driven (custom silicon is meaningfully below the HoldCo gross margin average). Per CFO Meintjes on the Q2 setup: "When we look forward to the second half, clearly, we're optimistic about custom demand. And that interplay between that and the rest of the business is really gonna drive what gross margin will be." The framing implies gross margin probably ranges 59%–60% through the year as custom mix expands. Non-GAAP op margin +870bp YoY on operating expense discipline (opex slightly below guide). The path to the long-term op margin target (high-30s range) is visible as revenue scales.

EPS / Cash Assessment

Non-GAAP EPS $0.62 (+158% YoY) reflects the operating-leverage flow-through plus a modest contribution from the 10% non-GAAP tax rate and continued share-count reduction from the buyback program. Cash flow from operations $333M (+108% YoY); inventory $1.07B (+$42M sequentially to support growth). Q1 buyback $340M — a meaningful step-up from $200M Q4 FY25. Total debt $4.2B; gross debt-to-EBITDA 1.8x; net debt-to-EBITDA 1.42x. Both ratios continuing to improve as EBITDA scales. Cash + equivalents $886M. The capital structure is being managed well; the Auto Ethernet $2.5B proceeds (closing in calendar 2025) will fundamentally expand capital allocation flexibility.

Segment / Product Detail

Data Center — $1.44B (+76% YoY) — The Core Engine

Custom XPU Programs — The Pillar

Marvell's custom AI silicon business is now the company's largest revenue growth driver. The lead AWS Trainium-class XPU program (which began zero-defect first-silicon production in calendar 2024) is now at high-volume production. Critically, Murphy was explicit that the follow-on generation is progressing: "We have now secured three-nanometer wafer and advanced packaging capacity and expect to start production in calendar 2026." The architecture team is already engaged on the generation after that — a "multi-generational nature of these engagements." Plus: the second hyperscaler XPU program announced at AI Day '24 is in joint development; architecture engagement on its follow-on already underway.

Assessment: The 3nm capacity-secured language is the most important signal of the quarter for the Trainium exclusivity debate. Throughout March-May 2025, multiple Asia supply-chain reports suggested Marvell may not have won the Trainium 3 follow-on or could be sharing it with a competitor. Murphy's direct contradiction — "the fact is that these Asia supply chain sources have an incomplete view. We run a very tight ship here at Marvell. We don't share customer confidential information" — supplements the operational evidence (capacity already booked, production starting CY26). We treat the 3nm-secured statement as credible at face value but acknowledge the multi-vendor scenario Murphy hinted at ("given the volumes that have materialized for XPUs, it's certainly possible and likely that customers and our customer may be pursuing multiple paths") as a residual risk to be sized. Net: revenue continuity on the lead program is high-confidence; multi-year growth depends on incremental wins beyond just this one socket — which is what the June 17 event will detail.

NVLink Fusion Partnership — Validation of Custom-vs-Merchant Complementarity

"Marvell custom silicon with NVLink Fusion provides our customers with an accelerated path to custom scale-up solutions, offering greater flexibility and choice in developing next-generation AI infrastructure. This further validates the proliferation of custom XPUs as a strong complement to merchant solutions."
— Matt Murphy, CEO

NVIDIA opened its NVLink ecosystem to enable custom XPU integration; Marvell is the silicon partner. The chiplet-based delivery model means Marvell does the XPU design with the customer, then NVLink chiplets handle the scale-up network interface. The strategic message: NVIDIA — the merchant-XPU king — is acknowledging that custom is complementary, not just competitive. This is a quietly significant industry framing shift.

Vault — Multi-Die Packaging Platform in Production

Earlier in the quarter Marvell announced its Vault multi-die packaging platform — qualified and in production supporting a customer-specific XPU. Differentiated interposer technology with lower power, higher yield, lower product cost than traditional silicon interposers. Strategic positioning ahead of the industry shift to multi-die XPU architectures.

Electro-Optics / PAM DSPs — The Other Pillar

At Optical Fiber Conference (OFC) 2025, Marvell showcased:

• First 400 gig per lane PAM technology — path to 3.2T optical interconnects
• 6.4T light engine consolidating components for pluggable and CPO applications
• 1.6T linear drive pluggable optical modules using Marvell silicon photonics
• First 3nm 1.6T PAM4 DSP with 200 gig per lane electrical/optical interfaces — 20%+ module power reduction
• Next-gen 800 gig DCI modules supporting transmission up to 1,000km
• Production-ready 1.6T AEC DSPs
• Coherent light DSPs for emerging distributed campus DCI (up to 20km)
• PCIe Gen 6 and Gen 7 SerDes

Optical business growing throughout the year; 1.6T ramping at five-nanometer; three-nanometer 1.6T product on a much stronger ramp into next year. 800 gig still the volume bulk through 2025-2026.

Assessment: Marvell's electro-optics leadership is well-defended. First to market at 800G, first at 1.6T 5nm, first at 1.6T 3nm. The 400 gig per lane demonstration positions for 3.2T transition. Optics is the second pillar of the data center story alongside custom; it carries higher gross margins than custom, providing the offset to custom-mix gross margin pressure.

Enterprise Networking + Carrier Infrastructure — Combined +14% QoQ

Both end markets are in cyclical recovery. Q2 guide: combined +mid-single-digit sequential. The fifth consecutive quarter of sequential growth signals the recovery is structurally durable, not a one-quarter pop.

Assessment: The communications recovery is the underappreciated tailwind to the FY26 story. Off a $900M annualized trough in Q1 FY25, the business is now at a ~$1.3B annualized run rate and trending toward $2B+ over time. This is incremental EPS contribution as the businesses normalize from inventory destocking.

Consumer + Auto / Industrial — Smaller, Largely Seasonal

Consumer Q1 $63M (-29% QoQ) — seasonal gaming softness. Q2 expected to rebound ~50% QoQ on seasonality. Auto/industrial Q1 $76M (-12% QoQ) — driven by industrial lumpiness; auto Ethernet still growing sequentially before the divestiture. The Q2 guide for auto/industrial flat sequential reflects continuing puts and takes.

Key Topics & Management Commentary

1. The "Asia Supply Chain Noise" Around Trainium Follow-On — Murphy's Direct Response

The opening Q&A pushed directly on the multi-week investor concern that Marvell may not be the exclusive supplier on the AWS Trainium 3 follow-on at 3nm. Multiple Asia-based supply-chain reports through March-May 2025 had suggested a competitor (Alchip/MediaTek) was also engaged. Murphy's response was unusually direct:

"The fact is that these Asia supply chain sources have an incomplete view. We run a very tight ship here at Marvell. We don't share customer confidential information. We treat that with the highest priority. So they simply have no idea what we're doing for our customer. So let me state the facts as they exist today. We're the incumbent shipping the current generation of this AI XPU, and we, as I've detailed, we've had a very successful and rapid ramp on this program. From a zero, which is first-time success, to high-volume production. … This next-generation program has continued to move forward, and this quarter, we have secured three-nanometer wafer and advanced packaging capacity. And that's for 2026, where we expect to start production. … given the volumes that have materialized for XPUs, it's certainly possible and likely that customers and our customer may be pursuing multiple paths to meet their requirements."
— Matt Murphy, CEO

Assessment: Three things matter in this response: (1) Murphy explicitly confirms the lead-XPU follow-on is on the Marvell roadmap with 3nm capacity already booked — a concrete operational fact. (2) Murphy explicitly does NOT confirm exclusivity. The phrase "customers and our customer may be pursuing multiple paths" is the key concession — i.e., AWS may also use a competitor for some volume. This is a significant framing shift. (3) Murphy's broader point — "we expect our revenue to continue to grow on a multi-year, multi-generational basis with this customer" — is consistent with the revenue ramp being durable even in a multi-vendor scenario, because XPU volumes are growing significantly faster than any single vendor's capacity. Net: the bear narrative ("Marvell loses Trainium 3") is materially overstated; the realistic outcome is multi-vendor sourcing with Marvell maintaining a substantial share. We treat this as constructive but not fully resolving the debate — which is why the June 17 event matters.

2. The June 17 Custom Silicon Investor Event — The Coming Catalyst

"We are hosting a dedicated forum on June 17 for investors to gain deeper insight into Marvell's unique position in the custom silicon market. … the agenda will include a market-focused section with updates on the expanding market opportunity for custom silicon, a robust design win pipeline, and the growing role of custom silicon in AI infrastructure. We will also share our progress towards the market share goals we set at our AI Day event last year, along with our vision for significant growth in the years ahead."
— Matt Murphy, CEO

At AI Day 2024 (April 2024), Marvell laid out: $75B 2028 data center TAM, 15% Marvell market share goal, $2.5B 2025 AI revenue commitment, multi-customer custom design wins. The June 17 event will refresh and expand this framework.

Assessment: The June 17 event is the rating catalyst for the next 4-6 weeks. The market is positioned for some specific data points: (a) updated TAM (above $75B), (b) explicit design-win count beyond the current "couple" framing, (c) revenue trajectory through 2028, (d) market share target update. If the event delivers materially expanded framing, the stock has 20-30% upside from current $67 levels. If the event is a "more of the same" repackaging, the stock could give back 10-15%. We initiate at Hold partly to preserve flexibility to upgrade post-event.

3. Auto Ethernet Divestiture to Infineon — $2.5B Cash

"During the first quarter, we announced the sale of our automotive Ethernet business to Infineon in an all-cash transaction valued at $2.5 billion. … The closing of this transaction, which we expect within calendar 2025, will provide us with additional flexibility in our capital allocation strategy."
— Matt Murphy, CEO

The Auto Ethernet business (~$200M annual revenue) sold to Infineon at a double-digit revenue multiple — explicitly above Marvell's own ~9x revenue multiple. This is an exemplary strategic move: monetize a non-core sub-scale business at a peak valuation, redeploy proceeds into the core data-center platform.

Assessment: Net positive on multiple levels: (1) strategic focus tightens to data center; (2) $2.5B cash inflow ($1.5B+ net of tax/transaction costs) materially expands capital flexibility; (3) signals to investors that the management team will be aggressive about portfolio rationalization. We expect the proceeds to fund continued accelerated buybacks (Q1 already stepped up to $340M from $200M Q4) and potential tuck-in acquisitions to expand the data center platform (CXL, photonics, etc.).

4. NVLink Fusion Partnership with NVIDIA

Marvell custom silicon will integrate NVIDIA's NVLink Fusion chiplet to enable custom XPUs to interface with NVIDIA's rack-scale solutions. Delivered via chiplet integration — Marvell does the XPU design, NVLink chiplet handles scale-up network interface.

"That partnership … demonstrates how well, one, you know, we're deeply engaged in the ecosystem, Tore, with key partners. And the second, I think it's some validation as well that a market standpoint, there is a complementary role for custom even as acknowledged by NVIDIA. … And we just view it as another key piece to the whole portfolio that we offer to customers end to end."
— Matt Murphy, CEO

Assessment: Strategically meaningful. NVIDIA is signaling that the merchant-vs-custom narrative is too binary — the future is hybrid. Marvell becomes the silicon partner for customers who want NVIDIA's scale-up network but custom XPU compute. This is genuinely new optionality.

5. AI Now Majority of Data Center — Path to HoldCo Majority

"As I mentioned during our March earnings call, AI now represents the majority of our data center revenue, and we expect the relative proportion of AI-related revenue to grow further in the coming years, driven in large part by our custom silicon business. … we see a path here in the not too distant future where AI is not only the majority of the data center end market, but it becomes the majority of whole co."
— Matt Murphy, CEO

If AI is currently >50% of data center, and data center is 76% of HoldCo, then AI is roughly 40%+ of HoldCo today. The "path to AI as majority of HoldCo" implies AI crosses 50% of HoldCo within 2-3 quarters at current growth trajectories.

Assessment: The mix shift to AI-majority HoldCo will materially change Marvell's investor narrative and likely the multiple over the next 12 months.

6. Custom Business Gross Margin Dynamics

"When we look at our overall, you know, custom business as a whole, we've been able to manage the gross margins, you know, even with these ramps in the range we've talked about. But within custom, there is a range. And it's like the old, you know, the old adage in semiconductors. You know, the higher the volume, you're gonna see lower gross margins, but you're going to see a lot more operating income."
— Matt Murphy, CEO

Custom inherently lower gross margin than HoldCo average. Within custom, XPU programs are lower margin than XPU-attach or peripheral programs. Operating margin remains accretive due to the operating-leverage on scale.

Assessment: The trade-off is well-understood and well-managed. Investors should focus on operating-dollar growth, not gross-margin %, for the custom business.

7. 1.6T DSP Ramp — Five-Nanometer Now, Three-Nanometer in Pipeline

Marvell commenced shipments on 1.6T at five-nanometer. Strong demand for three-nanometer. 800G continues to dominate volume through this year (1.6T ramps significantly stronger in FY27).

Assessment: 1.6T leadership well-established. The 3nm transition is the natural progression — power-efficiency benefits drive customer pull, supporting 3nm becoming the dominant node by end of FY27.

8. Capital Allocation — Buyback Step-Up

Q1 buyback $340M vs Q4 FY25 $200M. The step-up signals management's confidence in valuation. Pre Auto Ethernet close (calendar 2025), buyback capacity expands materially.

Assessment: Capital return discipline + accelerating cadence is the right capital posture given the stock has dropped from $120+ early 2025 to $67 now. Management is taking advantage.

Analyst Q&A

3nm exclusivity and content direction on next-gen XPUs

Q: "Just given all the Asia supply chain noise, I guess I have two related questions. One is what is the direction of content in these next-generation programs? You know? Is it up, down, or flat? And number two, are you exclusive on these three-nanometer XPUs, or do you expect to share that program given that, you know, your competition also seems to be convinced that they have the program?"
— Vivek Arya, Bank of America

A: "The fact is that these Asia supply chain sources have an incomplete view. We run a very tight ship here at Marvell. … This next-generation program has continued to move forward, and this quarter, we have secured three-nanometer wafer and advanced packaging capacity. … given the volumes that have materialized for XPUs, it's certainly possible and likely that customers and our customer may be pursuing multiple paths to meet their requirements. … irrespective of any of these relationships, we expect our revenue to continue to grow on a multiyear, multigenerational basis with this customer."
— Matt Murphy, CEO

Broader custom customer expansion capacity

Q: "One of your primary XPU competitors has talked about engagements at additional customers and not necessarily to front-run anything you may or may not say on June 17, but how do you see Marvell's ability to support a broader customer base beyond the initial, I guess, three/four that you have depending upon if we're talking XPUs or connectivity."
— Ross Seymore, Deutsche Bank

A: "We absolutely have the capacity engineering-wise to expand our portfolio and expand our engagements. … if you look at just the raw R&D spending, the investments we're making in next-generation technologies, and we're gonna cover all this at the AI day. We are extremely well positioned to support and are supporting multiple increased engagements across a number of programs. … there's been a really interesting emergence over the last year or so of the next wave of hyperscale kinda class customers that can benefit from our technology, relative to going custom."
— Matt Murphy, CEO

200G SerDes positioning + NVLink Fusion

Q: "There's a lot of debates out there about your SerDes technology, especially, you know, for 200 gig SerDes. … And maybe related to that, you know, how this relationship with NVIDIA is gonna, you know, work out."
— Tore Svanberg, Stifel

A: "On the first topic, and we'll actually be showcasing this at our 200 gig, it's performing extremely well. And we just showed off at OFC the first and only 400 gig per lane demonstrations with, you know, a very aggressive roadmap there. … On NVLink Fusion … I think it demonstrates how well, one, we're deeply engaged in the ecosystem … I think it's some validation as well that a market standpoint, there is a complementary role for custom even as acknowledged by NVIDIA."
— Matt Murphy, CEO

Data center AI revenue trajectory + gross margin

Q: "I think last quarter, you said that AI was about 55% of data center. I wonder if it was in the same range. I assume maybe custom silicon grew a little because the gross margin was a little bit lower. … obviously, you're running way above this $2.5 billion number this year for AI revenue."
— Timothy Arcuri, UBS

A: "We did say that at that time, that within data center AI had crossed and was now the majority of our data center revenue. … we see a path here in the not too distant future where AI is not only the majority of the data center end market, but it becomes the majority of whole co. … on the gross margin side of things, there's mix within the quarter. And clearly, the custom business we have does run at a fundamentally lower gross margin. So that is gonna modulate our gross margins on a quarterly basis."
— Matt Murphy, CEO

Second-half outlook + carrier/enterprise recovery

Q: "I'm sure you're not ready to provide that guidance right now. But could you give us some puts and takes on, you know, what your expectations may be? You expect the custom business to continue to grow in the second half of the year? You know, what do you expect with regard to, you know, the carrier enterprise businesses?"
— Chris Caso, Wolfe Research

A: "We're monitoring, you know, the macro and the various dynamics going on out there. So that's always a factor. But at the moment … we see the AI demand continuing, and our data center demand continuing. And then on top of that, we have a nice recovery, strong recovery, actually, we've been driving in our core business, in particular, enterprise networking and carrier. So we expect that to continue to recover and grow throughout the rest of the year as well. … right now, we expect growth kind of across the board through the year with, I think, what can be a really nice setup for fiscal 2027 with some of the growth drivers we've articulated before."
— Matt Murphy, CEO

On-prem data center drag + AI growth dynamics

Q: "When we talk to investors, Matt, we find that there's a desire on part of investors for the data center business to grow faster than, let's say, what you've been putting up mid-single in the April and then also the guidance. I know that in the last quarter, Matt, you had some moving parts with on-prem, etcetera. I was hoping that you could clarify the growth of your AI business versus some of the non-AI pieces like on-prem or some other things that might be moving around and just help us understand how fast maybe your AI business is growing."
— Harsh Kumar, Piper Sandler

A: "The way to think about it is helpful is to contextualize. You know, it's not a one-quarter snapshot. … if you look at the last few quarters, you know, we had a very dramatic step up from Q3 to Q4, you know, 20% plus in the data center business. It grew four. We're guiding it up mid-single. … look, a year ago, we were in a very strong position relative to our data center and how it was growing. So the comparables are also telling. … on-prem is pretty small at this point. So it has some effect, and it's always, you know, continue to be a little bit of a drag. But, you know, the overall data center revenue is driven primarily by AI."
— Matt Murphy, CEO

Optical + AWS networking products

Q: "You are guiding data center at mid-single digits into the out quarter, and you're saying custom silicon is growing quite strongly. Is there something going on with optical? Is that flattening out a bit? … And then in your announcement with Amazon, you talked about a bunch of other products other than custom silicon that you guys were engaging with over the next couple of years. PCIe cables, PCIe retimers, AC. Could you maybe try to size what that impact is right now?"
— Tom O'Malley, Barclays

A: "Those engagements are going very well. … And then with respect to how the optics business is doing, it's done quite well. And in fact, you know, we see that business growing throughout the year this year. So we see both growth from custom as well as optics. … we have, you know, in some of these new emerging categories, we do have AEC ramping this year. … as well as continued revenue growth in switching. So a lot of good things happening there, Tom."
— Matt Murphy, CEO

1.6T optical ramp + 3nm transition

Q: "The lead AI GPU vendor, I think, is getting ready to commence shipments of the next-generation platform. … typically, like, the optics start to ship ahead of the GPUs and wrap scale platforms. So do you anticipate a strong ramp of your 1.6 t solution starting now? More importantly, your five-nanometer solution, I think, is already qualified with this set with this customer. Are you guys gonna be mapping your three-nanometer DSP solution into this next-generation, next networking platform?"
— Harlan Sur, JPMorgan

A: "You're right about the sequencing. And so we have commenced shipments on 1.6 t at five nanometers. We have very strong demand for three. And we've executed very well on that product development so far. … from a 1.6 t kind of lifespan perspective, I think that's where the volume is gonna be. And I think that's gonna be a much stronger ramp next year. … There's just a lot of demand to, you know, initial ramps and customers will be on five nanometers. But there's a big push in the ecosystem to drive it to three nanometers given the compelling power savings and the performance of our product."
— Matt Murphy, CEO

Custom gross margin sustainability

Q: "I think you guys kind of put a line in the sand at a certain gross margin for ASIC. I'm curious if that's still the right level. And when you look at these next generations, I'm glad you're involved with all those. You know, how should we think about, like, sustainable AI gross margins?"
— Blayne Curtis, Jefferies

A: "When we look at our overall, you know, custom business as a whole, we've been able to manage the gross margins, you know, even with these ramps in the range we've talked about. But within custom, there is a range. … the higher the volume, you're gonna see lower gross margins, but you're going to see a lot more operating income. … From a percent standpoint, we're managing the business fine, and the AI portion is always gonna be lower. … one thing that's consistent across all these programs is that we're driving accretive operating margin to our model."
— Matt Murphy, CEO & Willem Meintjes, CFO

Dual-sourcing risk and NVLink Fusion model

Q: "I wanted to follow-up on the deck question or maybe just a clarification on whether these ASIC programs could potentially be dual-sourced. … To the extent that they're dual-sourced, you spent years designing these programs. Do you guys have guarantees that you don't get blanked at the end of the year if they, you know, take the other path?"
— Quinn Bolton, Needham & Company

A: "I was quite lengthy in my prepared remarks and even my, I think, Vivek answer in terms of what I've the perspective I can offer you on that. So I think I've said it, and if I wasn't clear, you know, I did say that given the volumes, it's certainly possible that there are multiple paths pursued. For us in our business and what I said in the prepared remarks as well as we feel very confident and very good about our revenue continuity on the initial program and the revenue ramp on our additional program."
— Matt Murphy, CEO

What They're Not Saying

• No explicit confirmation of exclusivity on lead-XPU 3nm follow-on. The closest Murphy came was "multiple paths" — the implicit concession that multi-vendor sourcing is possible.
• No update on Trainium 3 revenue magnitude. Marvell is appropriately tight-lipped, but the magnitude of any potential multi-vendor share split is the central debate.
• No quantification of optics %, custom %, and other within data center. Last update was Q4 FY25 (optics ~50%, custom ~25%, other ~25%). The mix has shifted materially in Q1.
• No FY27 framework yet. Murphy explicitly defers to June 17.
• No commentary on potential additional M&A. The $2.5B Auto Ethernet proceeds plus the buyback step-up implies capital is flowing to capital return — but tuck-ins remain on the table.
• Limited Microsoft Azure / Google specifics. The second hyperscaler XPU customer (announced AI Day '24) is presumed to be Microsoft; Marvell has not named.

Market Reaction

• Pre-print (May 28 close): ~$65. Stock down from early-2025 highs of ~$120 due to (a) Trainium 3 controversy, (b) general AI-semis pullback in March-May 2025, (c) tariff/trade concerns affecting Asia exposure.
• Day-of (May 29): Print landed after-hours. Initial reaction +6% on the beat + $2B Q2 guide. Most of the call-time spent addressing Trainium controversy.
• Day-after (May 30): Stock closed ~$67 (+3%). Volume ~25M shares (~1.5x 30-day average).
• Peers (day-of): AVGO +1%, AMD +0.5%, NVDA flat. AI-semis mixed but not strongly differentiated.
• Sell-side flow: Mixed reactions; bulls highlighted the 3nm capacity confirmation; bears noted the "multiple paths" framing as concession. Most analysts hold ratings pending the June 17 event.

Interpretive read: The market processed Q1 as confirmation that the fundamentals are intact, but Murphy's "multiple paths" framing keeps the Trainium 3 ASIC concentration debate alive. Stock at ~$67 reflects: strong P&L trajectory + reasonable valuation (~22x forward P/E on FY27 estimates), but with the ASIC competition tail keeping the multiple compressed vs the 2024 high. June 17 is the catalyst.

Street Perspective

Debate 1: Is Marvell losing share on Trainium 3 / next-gen?

Bull view: Murphy's 3nm-secured statement is a concrete operational fact that has been validated by industry checks. The "multiple paths" framing reflects the reality that AWS Trainium volumes are growing far faster than any single vendor's capacity — multi-vendor sourcing is not a Marvell loss, it's a Marvell capacity-constrained world. Marvell will continue to grow Trainium-attributable revenue significantly through follow-on generations.

Bear view: The competitive supplier (Alchip, with MediaTek packaging) has been increasingly aggressive on cost. AWS may shift increasing share to the competitor for Trainium 3 and beyond, particularly given the production-cost-sensitive nature of inference workloads. Marvell could see Trainium-attributable revenue grow more slowly than the underlying program.

Our take: The reality is between the two. Marvell's installed-base advantage (proven NPI, established trust with AWS engineering) protects against wholesale displacement, but multi-vendor sourcing at the margin is highly probable. We model Trainium-attributable revenue growing roughly 30-40% per year through FY28 — solid but materially below what the equity market was pricing in early 2025 at $120. The June 17 event will be the next anchor.

Debate 2: How big can the custom business become?

Bull view: Marvell's AI Day '24 framework — 15% of $75B 2028 TAM = ~$11B revenue from custom alone by 2028 — looks understated. The June 17 event will likely raise the TAM and the share target. Combined with optical (~$3-4B by 2028) + NIC / PCIe / switching attach (~$2-3B by 2028) + comms recovery — Marvell could be a $18-22B revenue company by FY29-FY30.

Bear view: Custom margin compression as scale grows. The 15% target depends on Marvell winning incremental sockets at multiple hyperscalers, which is a multi-year sales execution risk. If competitive pressure intensifies, the share target slips to 10%.

Our take: The bull case is well-grounded in the multi-quarter execution to date. We model FY28 revenue ~$15B with custom at ~$8-10B. June 17 will refresh.

Debate 3: Is the multiple too low or appropriately discounted?

Bull view: At $67 (~22x forward P/E on FY27E ~$3 EPS), Marvell trades at a discount to AVGO (~35x) and to its own historical multiple (28-32x). With AI mix expanding, gross margin pressure absorbed, and EPS compounding 30-40% YoY, the multiple should re-rate to ~25-28x in 12 months — implying $80-95 stock.

Bear view: The discount reflects the Trainium concentration risk and the lack of multi-customer custom visibility (vs AVGO's diversified Google + Apple + Meta + ByteDance custom franchises). Until Marvell can demonstrate similar diversification, the discount is warranted.

Our take: The multiple discount vs AVGO is rational. The compression should narrow as Marvell delivers multi-customer custom visibility — which the June 17 event will frame. Hold rating reflects this opportunity.

Model Implications & Thesis Scorecard

Model Assumptions

• FY26 estimates: Revenue ~$8.0-8.2B (+35-40% YoY); non-GAAP gross margin ~59-60%; non-GAAP op margin ~34-35%; non-GAAP EPS ~$2.75-2.90; buyback ~$1.2-1.5B.
• FY27 estimates: Revenue ~$10-11B (+25-35%); non-GAAP op margin ~37-38%; non-GAAP EPS ~$3.75-4.25.
• FY28 estimates: Revenue ~$13-15B; non-GAAP EPS ~$5-6.
• Long-term framework: $20B revenue + ~40% op margin + $7-8 EPS by FY29-FY30 — the "AI Day '24 framework refreshed at June 17."

Thesis Scorecard

Rating & Action

Initiating Coverage at Hold. Marvell's Q1 FY26 is a high-quality print with all the right operational signals: record revenue, +158% EPS, multi-generational custom roadmap confirmed at 3nm, optical leadership defended, comms recovery durable. The reasons for Hold rather than Outperform: (1) the Trainium 3 multi-vendor question is acknowledged but not resolved; (2) the June 17 custom-silicon investor event has not yet happened and will materially recalibrate the multi-year framework — we want to upgrade post-catalyst rather than ahead of it; (3) gross margin compression from custom mix creates near-term modeling noise that needs to be balanced against EPS growth trajectory. The reasons against Underperform: the operational execution is excellent, the trajectory is multi-year compounding, capital allocation is aggressive, and the valuation at ~$67 already reflects substantial Trainium 3 discount.

Fair value range: $60–$85. Stock at ~$67 sits in the lower-middle of our range. We re-evaluate up to Outperform on (a) June 17 event materially expanding the framework, (b) Q2 print delivering above-guide on custom + optics, (c) explicit second/third hyperscaler XPU customer disclosure. We re-evaluate down to Underperform on (a) June 17 disappointing the market with vague framing, (b) Q2 custom soft, (c) further negative Asia-supply-chain noise that proves substantiated.

Key watch items into Q2:

• June 17 custom silicon investor event — multi-year framework refresh; TAM update; market share target; customer count disclosure.
• Q2 print: $2B guide validation; custom contribution magnitude; optical 1.6T ramp.
• Auto Ethernet close timing + redeployment of proceeds.
• Comms recovery trajectory — does the 5th consec quarter of growth continue?
• Any further Asia supply-chain commentary on Trainium 3 / next-gen volume splits.
• NVLink Fusion customer engagement signals.