Earnings Digest: Marvell (MRVL) | FY27 Q1 — From Riding CapEx to Defining the Shape of Scale-Up

2 days ago

10 min read

Executive Takeaway

Interconnect growth raised from 50% to 70%+ YoY is the single most important number from this call. Marvell is no longer just a PAM4 DSP supplier; it's a solutions provider for the entire scale-out + scale-across + scale-up connectivity architecture.
FY28 revenue raised to $16.5B (up $1.5B), with Data Center accelerating to +55% YoY; custom XPU raised from +20% YoY to “more than double”; scale-up optics now expected to exceed $300M, double the prior outlook ($150M, all from Celestial AI).
A three-pillar partnership with NVIDIA went public — silicon photonics, NVLink Fusion, and AI-RAN — not a press-release-grade tie-up, but a structural positioning move that upgrades Marvell from “merchant chip supplier” to “the customization bridge into the NVIDIA ecosystem.”

In one sentence: the signal from this call isn't “AI demand is strong” (that's a given) — it's that “Marvell has swallowed the entire scale-up interconnect BOM.”

1. The Numbers Are Wild Enough, but the Pace of Raises Is Wilder

First, this quarter's results and outlook:

Metric	FY27 Q1 Actual	FY27 Q2 Guidance	FY27 Full Year (Raised)	FY28 (Raised)
Total revenue	$2.418B (+28% YoY, +9% QoQ)	$2.7B (+35% YoY, +12% QoQ)	~$11.5B (+40% YoY)	~$16.5B (+45% YoY)
Data Center revenue	$1.8B (+27% YoY, +11% QoQ)	mid-to-high teens QoQ, mid-40% YoY	+50% YoY	+55% YoY
Interconnect	—	—	+70% YoY (previously +50%)	continues to outpace CapEx
Custom business	—	—	+20% YoY	more than double YoY
Non-GAAP EPS	$0.80 ($0.01 above guidance midpoint)	$0.88-$0.98	—	OpMargin reaching the upper end of 38-40%

The key isn't the absolute numbers — it's the curve of the raises:

- A year ago, they said FY27 quarterly revenue would grow sequentially in the high single digits, exiting Q4 at $3B - Last quarter, they said high-single-digit sequential growth - This quarter, they said Q2/Q3/Q4 will each grow 10%+ QoQ, reaching $3B in Q3 and accelerating to ~$3.4B in Q4

In other words, this company is raising its own full-year expectations every single quarter. Given the industry's current supply-chain structure, there are only two possible explanations: either order visibility has extended dramatically, or it's taking share from others. Judging from the call details, both are true.

2. Interconnect: From PAM4 Leader to a Full-Chain “Scale-Up + Scale-Across + Scale-Out” Solution

If you only read one section of this call, read the interconnect section. It's the real engine of Marvell's FY27/FY28 story.

2.1 Scale-Out PAM4 — 1.6T Already Shipping in Volume, 400G/Lane on the Way

The call reconfirmed Marvell's lead at 200G/lane (i.e., 1.6T optical modules):

- 800G demand keeps strengthening (not replacement, but dual-track growth) - 1.6T began volume shipments in FY26 H2, ramping fast in FY27 and set to “step up again” in FY28 - 400G/lane (next-gen PAM4, i.e., 3.2T) was first demonstrated at OFC in April 2025, and Marvell stresses it has been first-to-market in every PAM4 generation

Even more noteworthy is the TIA + Driver business sold alongside: management guided outright to “a $1B annualized run-rate in quarterly revenue within the next few quarters” — equivalent to 25% of the roughly $4B annual market for small-form-factor module electrical chips, a hidden ace overshadowed by the PAM4 DSP story.

2.2 Scale-Across DCI — Read This Section Carefully, It's a New Story

The most easily overlooked but structurally most significant shift in this call is Marvell repositioning DCI from “traditional long-haul transport between data centers” to “the scale-across back-end AI network.”

The logic chain: - Traditional DCI moved from 400G to 800G, carrying front-end network traffic between data centers - Because of power and site constraints, AI clusters must be built across data centers — once back-end GPU/XPU traffic must also cross data center boundaries - Marvell estimates total scale-across bandwidth at more than 10x existing front-end DCI - And scale-across skips 400G and 800G, jumping straight to 1.6T

Marvell's position: the industry's first 1.6T ZR/ZR+ DCI module with MACsec, using its in-house 2nm coherent DSP, sampling this year.

The DCI module business is currently at a $500M annualized rate (FY26), with FY28 seen reaching a $1B annualized line — transforming DCI from a “mature business” into “the next high-growth engine.” In ZR Is the Meat, Coherent Lite Is the Bone: The Next Decade's Battle in Optical Transceivers we broke down how the ZR / Coherent Lite / CPO routes map against each other. This call effectively expands ZR's TAM from “public network hops” to “AI back-end,” a structural narrative upgrade.

2.3 Scale-Up Optics — The Real X Factor, Quietly Doubled

At the end of last year, Marvell's FY28 revenue outlook for scale-up optics was about $150M, all from Celestial AI (the acquired Photonic Fabric IP).

This call said scale-up optics is expected to more than double — i.e., exceed $300M — and it's not just Celestial contributing; Marvell's own silicon photonics products are shipping too.

The root cause: the radix and bandwidth of scale-up domains are exploding. A traditional NVLink domain had only 8-16 GPUs; the new generation is moving to 72, 144, even 576 GPUs per scale-up domain. Beyond 1-2m, copper is no longer viable — it has to go optical.

Marvell's positioning in this battle is an extremely rare “cover the entire tech stack” approach:

- All three major modulator routes — MZM, EAM, and MRM — MZM/EAM on in-house SiPh, MRM in partnership with the TSMC COUPE platform - Acquired Celestial AI (EAM + low-power SerDes, already selected by a tier-1 hyperscaler for next-gen XPU scale-up) - Acquired Polariton (plasmonic modulators, 1 THz bandwidth, 10x that of silicon photonics, ammunition for DCI at 3.2T and beyond) - Still investing in newer technologies such as micro-LED and micro-VCSEL

Add 224G SerDes in volume production, four generations of reticle-size switch silicon in volume, and 1.5 billion hours of SiPh field data — this is one of the few companies in the industry able to integrate XPU + switch + optical modules end to end.

On why scale-up (rather than traditional scale-out) is CPO's real window, we went deep in The 3D Photonic Integration Watershed: OpenLight's CEO on Five Convergence Thresholds — Scale-Up Is CPO's Real Window. This call effectively turns that thesis from industry speculation into a verifiable revenue line.

3. Custom Business: “Diversified Customers” Is Finally Not Just a Slide

Over the past year, the market's biggest worry was whether Marvell's custom business was overly concentrated in a single hyperscaler (Amazon Trainium).

This call gave a formal answer: custom more than doubles in FY28, with three engines each contributing roughly one-third:

1. Existing XPU programs (including the flagship) continue to ramp — in other words, Trainium isn't weakening

2. 10+ XPU-attach programs ramping — mainly NICs, CXL memory expanders, and retimers, with KV-caching and inference demand exceeding expectations

3. A new tier-1 XPU customer entering volume production — all next-year demand already firm, all R&D milestones met

Point 3 addresses the “Microsoft Maia or Meta in-house chip” rumors the market has fixated on for the past six months. Management said explicitly that “we have custom engagements with all five major U.S. cloud providers.” Without naming names, there are now three parallel tracks — the “flagship XPU,” the “new tier-1 XPU,” and “another group of XPU-attach customers” — far healthier diversification than the single AWS bet of a year ago.

What long-term investors should watch more closely: the FY29 target of $10B+ in custom revenue was reaffirmed, and Matthew Murphy stated clearly that “this counts only sockets already won and locked in — no new design wins are needed to fill it.” Put that next to “we'll win more designs for FY29,” and $10B is effectively the floor.

On questions about SRAM-heavy inference ASICs (like the Groq LPU), management's answer revealed Marvell holds the industry's first 2nm SRAM IP for SRAM density — a “not loudly advertised but critical” differentiator in inference-heavy custom ASIC design.

4. The Three Pillars of the NVIDIA Partnership — Bigger Than It Looks

The call devoted a long stretch to the NVIDIA partnership, and the market's first reaction might be “just another strategic-partnership press release.” But look closely at the content, and each of the three pillars is a structural business:

Pillar 1: Silicon Photonics

Marvell has long been NVIDIA's DSP, TIA, and Driver supplier; now it's expanding to collaboration at the silicon photonics level — meaning NVIDIA's next-gen Quantum-X / Spectrum-X CPO/NPO platforms will incorporate Marvell's SiPh process. Compared with the thesis we laid out in The USD 725B Question: As the Big Four CSPs Push 2026 Capex to Astronomical Levels, Optical's Real Bottleneck Quietly Moves Upstream (W21) — that NVIDIA's own CPO platform timeline is confirmed — this partnership is official endorsement that “when NVIDIA's own CPO launches, Marvell will be one of its optical front-end suppliers.”

Pillar 2: NVLink Fusion Integration

This is the most imaginative one. NVLink Fusion lets third-party custom chips plug directly into NVIDIA's scale-up domain — meaning Marvell customers can build their own XPUs but connect to NVIDIA GPUs via NVLink rather than PCIe or UALink.

Put plainly: previously, a hyperscaler using its own XPU plus NVIDIA GPUs ran two separate systems; now they can mix both in the same NVLink domain, with Marvell as the bridge. Strategically, Marvell turns “customization” and “the NVIDIA ecosystem” — once opposing options — into choices that can coexist. Huge flexibility for hyperscalers, pure upside for Marvell's custom business.

Pillar 3: AI-RAN

Integration of Marvell OCTEON base-station SoCs with NVIDIA GPUs to “run 5G/6G RAN and AI inference on the same hardware.” This TAM is longer-dated, but it's a new story for Marvell's carrier business — recovering from last year's inventory-correction trough, AI-RAN gives it a structural narrative beyond “just a cyclical recovery.”

NVIDIA also made an equity investment in Marvell (mentioned in this call's increase in diluted share count). When NVIDIA puts its own money in, the market should take this partnership as more than words.

5. Switching: Scale-Out Crosses $1B, Scale-Up Is Pure Whitespace

Scale-out switching

- FY27 expected to exceed $600M (2x FY26) - FY28 seen at an annualized $1B - 12.8T remains strong, 51.2T ramping, 100T announced (with emphasis on industry-best power efficiency)

From a $100-150M revenue line at the time of the Innovium acquisition to a $1B annualized run-rate in five years — this trajectory matters for validating Marvell's integration capability.

Scale-up switching (a pure greenfield opportunity)

After the XConn acquisition, Marvell strengthened its PCIe / CXL switch capability. But management stated clearly: PCIe is no longer sufficient due to radix and bandwidth limits, and will be replaced by UALink, ESUN, and NVLink.

Marvell has products across all three protocols: - In-house UALink switch - In-house ESUN switch - Entry into the NVLink ecosystem via NVLink Fusion

Management stated: “every tier-1 customer's lifetime engagement in scale-up switching is multi-billion-dollar scale.” If true, scale-up switching in FY29-30 is an order of magnitude larger market for Marvell than scale-out switching.

This narrative lines up exactly with the signal we flagged in CPO's Commercial Year One Begins — TSMC COUPE Volume Production and the 200G EML Bottleneck Draw 2026's Optical Winners and Losers (W21), where Marvell's Analyst Day laid out the ESUN / UALink 2026 timeline. Of the three stages — “timeline → customer commitment → order value” — we're now at the start of the third.

6. The Polariton Acquisition: Plasmonics Finally Moves from Papers to the Roadmap

The other acquisition announced on this call is Polariton (an ETH Zurich spin-off in Switzerland), which makes plasmonic modulators.

Plasmonics is a technology “hyped in academia for 15 years but long avoided by industry,” because its strength (modulator bandwidth up to 1 THz, 10x silicon photonics) and weakness (high optical loss) have been locked in a tug-of-war. The signal from Marvell acquiring it:

- At 3.2T and beyond, silicon photonics modulator bandwidth is approaching its ceiling - Plasmonics has insertion-loss issues, but in applications like CoherentLight and DCI that already have strong amplification chains, the weakness can be masked - Marvell lists Polariton as “key IP for the 3.2T and 4T+ roadmap”

This strategy shares the same roots as the electro-optic polymer story in Earnings Digest: Lightwave Logic (LWLG) | 2026 Q1 — AI Pushes Silicon Photonics to Its Limit, and EO Polymers Go from Decade-Long Backup to Lead — both are different answers to the same basic judgment that “at 3.2T and beyond, silicon photonics can't keep up and new materials are needed.” Marvell bets on plasmonics, LWLG on EO polymers; the two technologies will face off in 2028-2029.

7. Supply Chain and Capital Allocation: Running the 2020-21 Playbook

CFO Willem Meintjes disclosed $1B in prepayments in the Q&A (paid in installments over the year) to lock in long-term capacity at key suppliers. COO Christopher Koopmans added:

Since 2020-21 we have never been in an unconstrained state; AI-related products have been short of supply the whole way.

The implication: Marvell expects the AI supply chain to stay tight for the next 2-3 years — if you don't lock it in now, you won't have it later. It's the same playbook as locking up wafer capacity in 2020, except this time the target is high-end capacity at the level of advanced packaging + HBM + CoWoS.

The OpEx structure is also worth noting:

- FY27 non-GAAP OpEx of about $2.45B (including Celestial AI + XConn integration) - FY28 OpEx growth only mid-to-high teens — while revenue grows 45% - Result: late-FY28 operating margin reaches the upper end of the 38-40% target model

In other words, Marvell is simultaneously “increasing R&D investment” + “delivering clear operating leverage” + “continuing buybacks ($200M/quarter)” — something only a company with pricing power can pull off.

8. What to Watch Over the Next Two Quarters

1. New tier-1 XPU ramp progress — management says “all FY28 demand is firm”; next quarter, watch for guidance on ramp timing (volume starting in Q3 or Q4)

2. 1.6T DCI ZR/ZR+ module sampling timeline — this determines whether scale-across DCI can actually hit the $1B line in FY28

3. Leaks of the scale-up optics customer list — so far only “a tier-1 hyperscaler has selected Celestial,” but who's the next customer for Marvell's in-house SiPh? That drives whether the $300M figure can be raised again

4. The first NVLink Fusion deployment — for the market to truly size this TAM, it needs to see a first hyperscaler explicitly choosing “Marvell custom XPU + NVIDIA NVLink”

5. Polariton integration progress — the 3.2T DCI module sampling timeline will directly affect the credibility of the FY29 plasmonics story

STT Take

The most important thing about this call isn't any single number — it's the upgrade in narrative structure.

For the past three years, the market's story on Marvell was “PAM4 DSP leader, riding along with AI CapEx.” After this call, the story becomes:

Marvell is one of the few companies able to field full-stack solutions across all three tiers — scale-out, scale-across, and scale-up — and NVIDIA has already chosen it as the bridge into its customization ecosystem.

The value of this narrative upgrade isn't an extra $0.10 of EPS this year — it's the redefinition of the FY28-FY30 TAM: as the scale-up interconnect BOM shifts from “copper cables + connectors” to “DSP + driver + TIA + SiPh + modulator + switch,” all on silicon, Marvell is one of the most complete vendors on this new BOM.

Where to stay level-headed: scale-up is true greenfield, and the standards aren't even settled yet (a three-way tug-of-war among UALink, ESUN, and NVLink). Marvell is betting on all three — which today looks like “maximizing TAM coverage,” but in two years could become “diluted resources, third-rate at everything.” The thing to watch is which protocol actually reaches meaningful shipments in FY28 H2 — by then Marvell should become more focused.

The other point of caution: while the FY29 $10B+ custom target is backed by orders, “the flagship XPU customer concentration is still high” hasn't gone away — it's only been partly diluted by new customers and the XPU-attach business. Any change in Trainium's cadence will still flow straight through to Marvell's results.

Overall, though, the strength of this call's signal is the highest Marvell has shown in five years. The turning point from “peripheral participant” to “core rule-setter” may well be this quarter.

*This article is for technology and industry trend analysis only and does not constitute investment advice.*