MRVL (Trans): Raises revenue guide again; second XPU to enter mass production in FY28---

Below is Dolphin Research's Trans of Marvell's FY2026 Q4 earnings call. For the earnings read-through, see: Marvell: ASIC card miss, winning back AI via Interconnect.

$Marvell Tech(MRVL.US) Key takeaways:

1) Outlook: FY2027 total revenue raised to $11bn (prior $10bn), with Data Center up 40% YoY and Comms & other up 10% YoY. FQ4 FY2027 revenue to top $3bn, with YoY growth accelerating by quarter. FY2028 total revenue guided to around $15bn (prior $13bn), up nearly 40% YoY, with Data Center up nearly 50% YoY and Comms & other growing low single digits.

2) ASIC progress: FY26 revenue reached $1.5bn. FY27 to grow 20%+ (bias to the upside), and FY28 to at least double (not an aggressive case). The flagship XPU program (Amazon) is migrating to next gen, with firm orders for the full year in hand. A second top-tier cloud XPU program is on track, with mass production in FY28.

XPU adjacencies, booked under Custom, were at several hundred million dollars last year. This year is expected to double, and next year could double again to about $1bn. For CY2028, management targets a total XPU-adjacent TAM of ~$15bn, aiming for a 20% share (roughly $3bn).

3) Customer concentration: Across the four hyperscalers, Marvell ships highly diversified, tailored product mixes. While Custom draws attention, it is only one part of the story by numbers. Looking to FY2028–FY2029, with 20+ designs in production or entering production, customer diversity within Custom should improve further.

4) Electro-optics: 800G remains the shipment workhorse with momentum through this year and next. Management once expected growth to track CapEx, but performance now looks more like an acceleration, with growth already 50%+ this year. The strength should extend into FY2028, driven by three factors: higher optical attach with new XPU/GPU launches; 1.6T penetration lifting ASPs and revenue; and a pipeline of new wins ramping.

Overall, Marvell raised FY2027 and FY2028 outlook again, which directly boosts confidence. Prior market concerns focused on growth uncertainty (customer concentration, Alchip competition), but the higher guide is a strong rebuttal. Management addressed the debate with clearer visibility and momentum.

While ASIC is guided to grow just 20%+ in FY2027, Data Center still grows 40%+ on strong electro-optics demand. In other words, the focus had been on ASIC, but optics provide a more stable growth engine. In addition, a second XPU major customer ramps in FY28, likely re-accelerating ASIC-related revenue.

In AI data centers, hyperscalers value scalable clusters over simply selling cards. The Celestial AI and XConn deals, though not near-term revenue drivers, expand interconnect capabilities and position Marvell as a scale-out/scale-up cluster enabler. That should enhance customer reach and competitiveness.

I. Marvell earnings highlights recap

1. FY2027 Q1 guidance

a. Revenue: $2.4bn (+8% QoQ), with Data Center up ~10% QoQ. Comms & other to grow low single digits QoQ and ~30% YoY, continuing a steady recovery.

b. Non-GAAP GPM: 58.25%–59.25%.

c. OpEx: Q1 non-GAAP expected at $575mn, driven by seasonality and integration of Celestial AI and XConn (approx. $75mn annual OpEx).

d. Non-GAAP EPS: $0.74–$0.84.

2. Capital return: Full-year buybacks and dividends of $2.245bn.

II. Earnings call details

2.1 Management highlights

1. Interconnect

a. Scale-out (optics/PAM4): 800G demand remains strong. 1.6T is in production, expected to ramp fast in FY27 and become mainstream in FY28. The company underscores clear leadership at 200G/lane.

b. Scale-across (DCI/coherent): 400G/800G modules now serve all five US hyperscalers. Launched the industry's first 2nm 1.6T ZR/ZR+ module, with sampling expected later this year.

c. Scale-up (new vectors/AEC): With Celestial AI integration, its photonic architecture will contribute revenue in FY28, targeting a $500mn annualized run-rate (by FY28 Q4) and doubling to $1bn in FY29, with market size potentially >$10bn by 2030. In AEC, wins include three leading US hyperscalers plus model developers and OEMs, and AEC plus retimers are expected to more than double revenue YoY in FY2027.

2. Custom Silicon/ASIC

a. Status: FY26 revenue at $1.5bn. FY27 growth 20%+; FY28 at least doubling.

Three core drivers for FY28: ongoing contributions from existing Custom programs; multiple XPU-adjacent programs scaling to volume, notably custom NICs and CXL; and a brand-new flagship XPU program entering volume ramp.

b. Flagship programs: The lead XPU is moving to its next gen with full-year firm orders; a second top cloud XPU program is progressing well and slated for FY28 mass production.

c. New growth vectors: CXL expanders and custom NIC demand is accelerating, and by FY29 these two alone are expected to contribute $2bn+ revenue.

3. Switching

a. Scale-out switching: 51.2T platform is ramping strongly, with 100T sampling expected 1H this year. FY27 switching revenue outlook raised from $500mn to $600mn.

b. Scale-up switching (XConn): With XConn PCIe/CXL, Marvell is accelerating UA Link 115T, planning sampling in 2H this year and volume in FY28.

4. Comms & other remain stable, expected to grow 10% in FY27 and low single digits in FY28.

5. Strategy

a. Moat: Marvell is uniquely positioned to deliver a full-stack AI networking solution across optical interconnect + switching + Custom compute (ASIC).

b. M&A: Divested Auto Ethernet (double-digit revenue multiple) and quickly redeployed into Celestial AI and XConn.

c. Competition: Interconnect growth is expected to outpace hyperscaler CapEx growth.

6. Outlook:

a. FY2027: total revenue approaching $11bn, up 30%+ YoY, raised from $9.5bn (Sep-2025) and $10bn (Dec-2025). Data Center up 40% YoY; Comms & other up 10% YoY.

Quarterly revenue to rise sequentially, with FQ4 FY2027 above $3bn and YoY growth accelerating each quarter.

b. FY2028: total revenue around $15bn, up nearly 40% YoY, about $2bn above the Dec-2025 target. Data Center up nearly 50% YoY; Comms & other low single digits.

Non-GAAP EPS to exceed $5. Celestial AI and XConn to contribute about $250mn revenue in FY2028 as acquisitions begin to monetize.

c. Long-term: If achieved, Data Center would grow 40%+ for three consecutive years, placing Marvell in a multi-year AI infra upcycle.

2.2 Q&A

Q: Beyond the magnitude of growth, how does demand look by customer? The market worries about Custom concentration. Has the customer base broadened?

A: We are deeply embedded across the ecosystem and hold strong positions at all four US hyperscalers and second-tier players. While revenue mix and product sets differ by customer, even at the $11bn target, Custom is not the dominant piece. Concentration reflects where industry CapEx is, given the four absorb most spending.

Within these core customers, we are highly diversified, as shown by the long list of products we ship. We provide distinct, rich portfolios to each of the four. Custom draws attention but is one variable in the equation. Into FY2028–FY2029, as 20+ designs in or entering production ramp, Custom will also diversify by customer.

Our edge is breadth and depth across the stack to serve hyperscaler needs end-to-end. Recent M&A perfectly filled capability gaps in PCIe, UA Link and key silicon photonics, further extending what we can deliver to top customers.

Q: OpenAI's deal with your lead XPU customer may consume ~2GW of next-gen XPU, signaling accelerating AI compute. With 15–20 XPU-adjacent Custom projects landing over the next two years and the lead customer's next-gen XPU ramping, will growth step up in 2H or be linear? Any change to the prior $2bn annualized target?

A: Strong AI compute spend validation, especially hyperscalers building in-house XPUs, is very positive. Even where we do not design the XPU, our XPU adjacencies have broad penetration. Cross-platform engagement supports our view of upside to Custom growth this year and sets the stage for FY2028–FY2029.

On cadence, we still expect stronger 2H growth for Custom due to program transitions. On the annualized exit, the prior target stands with upside potential. Our overall exit revenue now points to $3bn+ in Q4, with Custom driving a meaningful portion of the uplift.

As 2H starts feed full-year next year, FY2028 revenue should inflect. Drivers are higher product value, tiered XPU adjacency ramps, and a new Tier-1 hyperscaler XPU program. Our current internal wafer starts and production plans exceed the financial forecast, and if trends hold, there is strong upside to next year's results.

Q: Can Electro-optics match 60%+ hyperscaler CapEx growth? Is high growth sustainable into FY2027–FY2028?

A: That is accurate. One key reason for the higher guide is the strength in optics. Previously in Sep and Dec, we expected growth to track CapEx, but performance now looks like an acceleration with 50%+ growth this year.

The momentum should persist into FY2028, driven by three factors: higher optical attach with new XPU/GPU launches; 1.6T penetration lifting ASPs and revenue; and a series of exciting new ramps. Since the Inphi deal, DC interconnect has compounded strongly, and tech-driven growth should endure even if absolute rates fluctuate.

Q: For CY2026 (FY2027), what is the Custom base? Does 20% growth move closer to ~30%? And confidence in timing and the 'doubling' target for the second XPU customer?

A: For FY2027, our view is above the prior 20% growth marker. I cannot provide a precise revision, but the bias is clearly upward, so adjust above 20%, which also raises the base for next year.

We are confident given Marvell's track record in large Custom ramps. We have clear gating milestones and deep alignment with customers on production plans, and our reserved capacity for FY2028 is well above what we have shared publicly.

Our budgets factor in potential delays and lingering skepticism in parts of the market, so the forecast is pragmatic. If we fully utilize the reserved capacity, upside is large, and doubling next year is not an aggressive assumption.

Q: Since Dec, the FY2028 outlook is up by $2bn. What is driving the lift, and can it extend into CY2028?

A: Visibility has improved step by step. From ~$9.4bn last Sep to $10bn in Dec and now $11bn+, each raise reflects clearer line of sight and concrete program ramps.

The biggest delta is Interconnect. Frankly, our prior model anchored it to total hyperscaler CapEx, but as some have noted, it should correlate more with GPU/XPU demand. Orders and backlog confirm this, so Interconnect estimates moved up materially and pulled forward next year's view.

All of this is backed by robust orders and deep supply planning with customers. Back in Apr-2024, we set a CY2028 target assuming 20% Data Center share, and many thought $15bn Data Center revenue was unrealistic. By last Jun's AI Day, as TAM expanded, that implied number had risen to $18bn+.

Execution to date is tracking those ambitions. It is not merely a numeric raise but validation of strategy set four years ago. With orders, forecasts and supply alignment, we are confident for this year, next, and CY2028.

Q: You said AEC and retimer will more than double this year. What is the base, and how do these fit into end-to-end Interconnect and contribute over the next few years?

A: It remains nascent for us. Even with more than doubling this year, revenue is around $200mn, and there is room to go higher. In semis, once such products start doubling, momentum tends to persist, and we feel very good about our tech position.

We are leveraging DSP and PAM4 extensively. We made a strategic call to enter retimers at the NRZ-to-PAM transition. We may be later than some in certain sockets and still early in the cycle, but we plan to invest and engage at scale.

Longer term, electrical and optical interconnect are complementary. We bet on pluggable optics with Inphi and added CPO via Celestial, but electrical remains essential at specific reaches and use cases. The goal is to be a one-stop, end-to-end interconnect partner across electrical, optical, silicon photonics and distances.

Customers need reusable IP, firmware, software and system-level reliability from a single partner. While AEC and retimers are still small, they are in a high-growth loop and are the final pieces completing our full-stack interconnect.

Q: Can you size last year's XPU adjacency revenue and expectations for FY2027–FY2028? Also, given competitive noise on the first-gen major customer, how defensible is the second major XPU project launching next year?

A: We are very confident in the competitive position of the second customer's program. It is a deep, co-steering engagement, and with rapid tech cycles, customers benefit from multi-gen planning with us, not just the next gen. Our roadmaps are tightly aligned, and the customer's CapEx can absorb substantial volume.

On size, we will not give exact numbers but provide order-of-magnitude. For CY2028, XPU adjacency TAM is about $15bn and our target share is 20% (~$3bn).

Over the past year, the business was at several hundred million dollars; this year should double, and next year double again. By next year, it should scale to roughly $1bn, booked under Custom, on a very steep growth trajectory with a doubling-plus cadence.

Q: Update the optics product mix, please. 1.6T has started shipping, but consensus is 800G remains mainstream this year. What is the FY2027 split among 1.6T, 800G and 400G?

A: Broadly right. 800G will have a long life, stay strong this year and next, and represent the majority of units. That said, 1.6T shipments picked up meaningfully late last year and will ramp steeply this year.

One reason for raising Interconnect is that all customers are lifting demand, with 1.6T especially strong. Early customers are expanding rapidly, and more will join over the next two years. Given the fluid order environment, we cannot provide precise mix yet, but clarity will improve as the year progresses.

Q: With such strong growth, any supply-chain challenges? How is your preparation vs. a few years ago?

A: Since ChatGPT, leading-edge wafers, advanced packaging and large substrates have been tight. Even so, we grew total revenue 40%+ last year, showing very strong supplier relationships.

We anticipated the upcycle early and provided multi-year visibility to suppliers. Based on locked-in resources, we are confident we can secure the supply needed to support growth over this year, next, and beyond.

Q: FY2028 revenue is guided to $15bn with EPS at about $5. That revenue is ~15% above Street, but profit uplift seems only half as large. What are the below-the-line drivers — mix vs. higher R&D to deliver growth?

A: The '>$5' EPS is a floor, not a precise forecast. You can model it with the current framework.

On revenue, use a macro framework off the $3bn+ exit quarterly run-rate this year. On the P&L, with the OpEx guide and GPM trajectory, we are nearing our target operating model by year-end.

A conservative assumption is to carry that margin profile into next year. At $15bn revenue under the current model, EPS clearly comes in above $5, with no signal of margin pressure or lost leverage. OPM is around 35% and trending up through year-end, and next year should at least match the year-end run-rate. Put simply, EPS should land comfortably above $5.

Q: On CPO at scale, how do you see deployments across architectures? With Celestial AI integrated, will you offer UA Link-based switch-tray products, and when would volume start?

A: We have long believed CPO will be limited at scale in scale-out networks vs. pluggables, and that view remains. We can integrate Celestial tech with Innovium or Teralynx and have proven concepts, but it is not our current focus and we will flex to market demand.

The true inflection is in scale-up architectures where UA Link-like protocols shine and CPO delivers large value. The Celestial deal brought a key design win and deep partnership here, and we are driving to a volume ramp by late next year.

The solution will serve scale-up use cases, integrating PF chiplets on the XPU and switch sides. While most scale-up switching will remain copper-based for some time, industry interest in CPO exploding in scale-up in a few years is very strong. That underpinned our recent M&A and team build-out.

In short, we will ship CPO for scale-up to a major customer next year and are developing follow-ons. Most other deployments will stay copper for now.

Q: Given Interconnect is the current AI bottleneck with higher margins and clearer leadership, why not tilt R&D fully to Interconnect instead of continuing to invest in Custom ASIC/XPU? Does it weigh on the multiple?

A: I am not sure which is the 'bigger' bottleneck — both Interconnect and processors have major challenges. We are going all-in on Interconnect, as seen in last year's M&A — there is no compromise there.

We persist in Custom for strategic reasons. When we bought Avera in 2019, we did not foresee a $50bn TAM, but we wanted a vanguard for Marvell. Custom forces leadership in advanced nodes, packaging, and core IP, turning Marvell from a fast follower into a true tech leader.

I acknowledge noise around Custom over the past year hurt the multiple. But the facts are: it scaled from zero to $1.5bn, grows again this year, and doubles next year. Customers fund large NRE and commit deeply, so this is not a resource drain.

Rumors about share loss or program changes proved wrong. Look at this year's guide and next year's outlook — do I look shaken? Not at all. We will ignore the noise and keep building Marvell as a company with impact across the AI stack, from electrical to optical, Interconnect to compute.

<End here>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure