Earnings Digest: Marvell (MRVL) | Q2 FY2027 — FY2028 Raised to $18B, and the Biggest Upgrade Comes from a Business That Didn't Exist a Year Ago

2 days ago

12 min read

1. This quarter isn't about the beat; it's about the size of the raise. Q2 revenue came in at $2.739B, up 36.6% YoY, but the real story is the outlook: FY2027 was raised from ~$11.5B to ~$12B, and FY2028 from $16.5B to ~$18B. Adding $1.5B in one quarter is like growing another mid-sized optical communications company. And despite a larger base, growth actually accelerates from +45% to +50%.

2. The largest piece of the $1.5B raise is not custom silicon; it's scale-up optics. Matt Murphy said this himself in Q&A. Last quarter, Marvell's FY2028 figures for scale-up optics were "~$150M for Celestial AI's CPO, ~$300M for scale-up optics overall." One quarter later, the category is "much bigger than we thought a quarter ago," and the growth driver has broadened from CPO to NPO. A brand-new category priced in hundreds of millions last quarter has become the single largest driver of this year's raise.

3. The real impact of the Google warrant lands in FY2029. Management was explicit: revenue from programs covered by the warrant through FY2028 is "already embedded in our existing custom targets"; the real step-up comes in FY2029 and beyond. The market annualized $120B over 6.5 years to roughly $18.5B per year, and Murphy's response was "your math is not wrong" — but he also said a range will only come at the October 6 Investor Day.

Last quarter, our Marvell headline was "From riding CapEx to defining the shape of scale-up." This quarter, that shape started showing up in the guidance. For the full context, see Earnings Digest: Marvell (MRVL) | FY27 Q1.

1. Revenue and Financials

Q2 FY2027 (quarter ended 2026-08-01) key figures, all cross-checked against the SEC 8-K Ex-99.1 press release and Marvell IR release:

・Revenue $2.739B: +36.6% YoY, +13.3% QoQ, a quarterly record and above the guidance midpoint

・Gross margin (GAAP) 53.1%: +2.7pp YoY; non-GAAP gross margin 58.9%, slightly above the guidance midpoint

・Operating income (GAAP) $459.7M, operating margin 16.8%: +58.5% YoY; non-GAAP operating margin 36.6%, +180bps YoY and +160bps QoQ

・Net income (GAAP) $308M: +58.1% YoY; non-GAAP net income $865.9M

・Diluted EPS (GAAP) $0.33: +50.0% YoY; non-GAAP EPS $0.94, +40% YoY, one cent above the guidance midpoint

Segments (Marvell discloses only two end markets, which sum to total revenue):

・Data center $2.1715B: +45.7% YoY, +18% QoQ, 79% of total revenue

・Communications and other $567.8M: +10.1% YoY, -3% QoQ

The data center line deserves a separate look at its acceleration: Q1 was +27% YoY / +11% QoQ, Q2 jumped to +46% YoY / +18% QoQ, and Q3 guidance goes further to roughly +75% YoY / more than +20% QoQ. Three consecutive quarters of stacked acceleration is a rare shape in semiconductors — usually a company accelerates for one quarter before hitting a base effect or a capacity ceiling.

Cash and balance sheet: operating cash flow was $605.5M (down slightly QoQ, mainly due to higher capacity prepayments to suppliers), inventory $1.361B (roughly flat QoQ), total debt $4.963B, gross debt/EBITDA 1.32x and net debt/EBITDA 0.27x. The company repurchased $200M of stock and paid $53.9M in dividends this quarter.

2. The Most Important Thing This Quarter: The $1.5B Raise Isn't Mainly Custom

For the past two years, the market has looked at Marvell almost exclusively through custom silicon (custom ASICs). The most counterintuitive line of this call came after questions from Aaron Rakers and James Schneider:

"The most meaningful part of this $1.5B raise actually comes from areas like scale-up optics and switching — custom is of course part of it too."

To grasp the weight of that statement, look back at last quarter's framing. Marvell's figures then: Celestial AI's CPO Photonic Fabric solution at about $150M in FY2028; scale-up optics overall (including NPO) as a category at about $300M. A $300M category becoming the single largest driver of a $1.5B annual raise one quarter later doesn't signal linear growth; it means the adoption timeline for the whole category has been pulled forward.

Murphy specifically corrected a market misconception: this is not a route war where CPO replaces NPO or NPO replaces CPO.

"It's an 'and,' not an 'or.' It's a bit like when everyone assumed 800G DSPs would switch straight to 1.6T — that's not how it works anymore. Customers have different architectures, and they'll adopt several solutions at the same time."

Marvell plays in NPO in two ways: first, SiGe-process TIAs and drivers — an existing strength where it has high share and content; second, its in-house light engine, developed over many years and shown at several OFCs. Add the CPO Photonic Fabric brought in by Celestial AI, and Marvell now pitches itself as "one of the largest NPO enablers in AI infrastructure."

We've broken down the technical trade-offs of this path: After Copper Gives Out for AI: Seven Paths to Scale-Up Optical Interconnect; and what the $3.25B Celestial AI acquisition actually bought is covered in full here: Celestial AI Explained: The $3.25B Bet on Bringing Light Inside the Chip.

Notably, Marvell also laid out its modulator technologies this time: customers are evaluating three approaches — MZM (Mach-Zehnder modulator), EAM (electro-absorption modulator), and MRM (micro-ring modulator) — across both NPO and CPO packaging, each with different trade-offs in cost, power, and maturity. Management's pitch is "we have all three, so whichever route customers choose, we're there" — a classic optionality argument: the upside is never betting on the wrong side; the cost is R&D spread across three lines.

3. Three Connectivity Lines, Each Running Toward $1B Annualized

Last quarter, Marvell set a goal: beyond optical DSPs, three more businesses would each reach $1B in annualized revenue. This quarter's update: "every one is on track or ahead":

・Scale-out switching: revenue this year will "more than double," driven by strong ramps of the 51.2T product line and a broader customer base. This line traces back to the Teralynx architecture from the 2021 Innovium acquisition; Marvell says it has "proven to the market that we can deliver switch silicon of this complexity in volume"

・Broadband analog (TIA/driver): demand is "above expectations," and it's also a key content source for NPO

・Scale-across DCI: the most overlooked of the three. When AI clusters grow large enough to span multiple data centers, total bandwidth demand for scale-across networks is projected at more than 10x that of today's front-end DCI networks. Marvell's entry point is 1.6T ZR/ZR+ coherent DSP modules

As for optical DSPs, 800G demand remains strong, 1.6T is ramping quickly, and growth will accelerate further in FY2028.

4. Scale-Up Switching: All-In on Three Protocols, Each Tier-1 a "Multi-Billion-Dollar" Opportunity

Marvell's stance on scale-up switching this quarter was emphatic: it supports all three dedicated protocols — UALink, ESUN, and NVLink Fusion — the first two with in-house switch silicon, NVLink Fusion through an expanded partnership with NVIDIA. Management said it is in deep discussions with multiple tier-1 customers, and each engagement represents a "multi-billion-dollar lifetime revenue opportunity".

Joe Moore asked a key question: customers use copper for scale-up today, so can success in copper extend to optics? Murphy answered on two levels —

First, trust has already been earned: Teralynx has a volume track record in Ethernet switching, and "many companies have tried to build this kind of complex, high-radix, high-SerDes-rate switch silicon and failed." The second level is what decides the game:

"But the next thing customers want is whether you can bring a credible optical roadmap, from NPO all the way to fully integrated CPO. In scale-up, they need both ends of the link covered — they need to believe there's a partner who can handle both ends of that link and everything in between. Point solutions won't pass at this stage."

Murphy also gave a very concrete business implication: once optical attach takes off, it nearly doubles the switch SAM (serviceable addressable market). That explains why Marvell is willing to pour resources into both switching and optics — it's not selling two chips; it's selling an entire link.

Who's on the other side of this fight? We wrote about it: 2026 OCP APAC Summit | Ethernet Moves into Scale-Up: Broadcom's Keynote Is Aimed Squarely at NVLink.

5. The Google Warrant: FY2028 Already in the Numbers, the Step-Up Comes in FY2029

An 8-K Marvell filed last week disclosed an expanded commercial agreement with a "large hyperscaler" that includes a warrant. In the Q&A, analysts referred to it directly as Google, and management mentioned the "TPU ecosystem" and being "very honored to be part of it."

The warrant's scope is broad: existing custom programs already in execution, new design wins, and potential future programs, spanning AI inference accelerators, storage controllers, network interface controllers (NICs), memory interface controllers, and near-memory compute.

On the timeline, management was more restrained than the market imagined:

Period	Warrant impact
FY2027 (current year)	Some programs already shipping; already in current numbers
FY2028	Revenue from covered programs is "already reflected in existing custom targets" (custom more than doubling YoY)
FY2029 and beyond	"Whatever number you had in your custom model, it needs to go up" — but no range given this quarter

Harlan Sur and Ben Reitzes both pressed with the same math: $120B cumulative revenue ÷ 6.5 years ≈ $18.5B per year, equivalent to "adding another FY2028 Marvell every year." Murphy replied "your math is not wrong," but added three caveats: it assumes all milestones are fully achieved; the Investor Day needs to provide a range of outcomes; and gross margins still follow custom's existing financial model (i.e., below the corporate average — the warrant doesn't turn this into a high-margin business).

The more telling part is this: in April 2024 Marvell put its custom FY2029 target at $8–10B, and in June 2025 at $10B+, and the market spent two years doubting it could get there. The warrant's real effect is less about how much revenue it adds and more that it takes the question of "can they do it" off the table.

Incidentally, this also explains why XPU attach — a category Marvell defined itself two years ago — matters so much: what's truly broad in the warrant isn't the XPU itself, but the whole ring of peripheral silicon growing around it. We've unpacked this "not selling memory, selling the data-movement layer" positioning: Marvell Isn't Selling Memory: It Wants to Be the "General Contractor" of AI's Data-Movement Layer.

6. CXL Becomes the Quarter's Hidden Thread: Memory Shortages Push Structera onto the Stage

C.J. Muse asked a good question: with memory prices this high, how does the "fungibility" between compute, networking, and memory affect Marvell's design wins? Murphy's answer revealed a line that wasn't on the market's radar.

Marvell's CXL product line (Structera) was originally designed for traditional server memory expansion. That investment unexpectedly ran into two things:

1. Inference itself needs memory expansion — an architectural need, not a cyclical one

2. Memory shortages — to work around HBM/DRAM supply bottlenecks, customers are changing architectures and shifting more workloads onto these technologies

The result: "multiple hyperscalers, multiple architectures, very high volumes," with new design wins in recent quarters. Murphy called it a "home run" and previewed a full memory-expansion segment at the Investor Day.

This matters more for the supply chain than for Marvell itself: if memory shortages systematically push workloads toward CXL and near-memory compute, the beneficiaries aren't just Marvell but the entire "connectivity and data-movement" layer — exactly where Astera Labs, Credo, and their peers sit.

7. Management Outlook: Everything Raised, but Gross Margin Is the Price to Pay

Q3 FY2027 guidance (official figures)

Item	GAAP	Non-GAAP
Revenue	$3.15B ± 5% (+15% QoQ, more than +50% YoY)	Same as GAAP
Gross margin	52.9%–53.9%	57.5%–58.5%
Operating expenses	~$1.015B	~$655M
Diluted EPS	$0.53 ± $0.05	$1.10 ± $0.05

Data center is expected to grow more than 20% QoQ and roughly 75% YoY; communications and other is expected to decline "low-to-mid teens" percent both QoQ and YoY, before rebounding clearly in Q4.

Full-year and multi-year outlook

・FY2027 revenue of about $12B (roughly +45% YoY), raised from ~$11.5B a quarter ago; FY2027 data center growth of about +60% YoY, raised from ~+50%

・FY2028 revenue of about $18B (roughly +50% YoY), raised from $16.5B; FY2028 data center growth above +60% YoY, with custom more than doubling YoY

・Non-GAAP operating expenses: FY2027 about $2.55B (raised from $2.45B); FY2028 opex growth roughly half the rate of revenue growth

・Non-GAAP operating margin: entering the 38%–40% long-term target range in Q4 FY2027, reaching the upper end during FY2028; the Investor Day will reset the long-term model

・Non-GAAP tax rate: 11% in Q3, about 13% in FY2028

・FY2027 capacity prepayments of about $1B, unchanged from last quarter

Confidence tiers in the language (on earnings calls, "confirm," "expect," and "target" signal three different levels of certainty):

・Locked in: Q3 guidance, FY2027 full year of ~$12B, custom more than doubling in FY2028

・Strongly expected but not quantified: scale-up optics in FY2028 "much bigger than a quarter ago" and CXL a "huge opportunity" — both deliberately left without numbers

・Explicitly deferred to the Investor Day: FY2029 custom size, long-term operating margin model, actual number of scale-up switch design wins

Gross margin is the price of this growth. The Q3 non-GAAP gross margin midpoint is about 90bps below Q2, and Dan Durn said plainly the main cause is mix dilution from accelerating custom shipments; Q4 stays in the same range, and the "preliminary view" for FY2028 is the same range too. In other words, Marvell's profit growth over the next two years will come from operating margin (opex leverage), not gross margin — which is why management keeps stressing that opex grows at only half the pace of revenue.

8. Supply Chain and Customer Clues

A few threads that weren't named directly but send clear signals:

・Capacity was "grabbed," not just "ordered". Murphy said the operations team "secured additional capacity in an industry-wide constrained environment." The $1B in capacity prepayments is the financial evidence — the money offsets future material purchases and is essentially trading balance sheet for a place in line for wafer and packaging capacity

・The second large XPU customer is on track. Vivek Arya noted the market assumes $600–700M for this program next year; Murphy didn't dispute the figure, saying only that design execution, supply, and commercial terms are progressing every quarter, and that "this has always been one of the largest revenue opportunities Marvell has"

・The NVLink Fusion partnership is expanding. Marvell works on both UALink/ESUN (the open camp) and NVLink Fusion (the NVIDIA camp), a rare dual position in the supply chain — but it also means it isn't the dominant party in any protocol war; it's the pick-and-shovel seller

・Management turnover is complete. The CFO role passed from Willem Meintjes to Dan Durn (effective mid-June), who has more than 30 years of finance experience in semiconductors and enterprise tech and was already a Marvell board member; head of IR Ashish Saran will retire in April 2027, succeeded by Ross Seymore, who covered semiconductors at Deutsche Bank for 25 years. At a company scaling revenue from $8B to $18B, replacing both the CFO and IR lead with people "fluent in large-scale capital allocation and sell-side language" is a signal in itself

9. Risks and Counterarguments

To be balanced, this report carries several real risks:

1. Structural gross margin dilution. Every step up in custom mix pushes blended gross margin down a step. Marvell chooses to offset it with opex leverage, but that path has less margin for error than gross margin expansion — if revenue growth falls short, the opex has already been spent

2. Customer concentration and bargaining power. The flip side of the warrant is tying one customer's interests to equity. That's a plus for growth visibility, but not necessarily for long-term pricing power

3. Communications and other is still a drag. Q3 is expected to fall "low-to-mid teens" both QoQ and YoY, and the full year is only "close to the 10% growth target." This segment is 21% of revenue, and its swings are amplified in the totals

4. The scale-up optics timeline is still a "customer plan," not a "locked order". Murphy himself said copper and optics will coexist for years, and customers "could start aggressively deploying scale-up optics as early as next year" — "as early as" and "aggressively planning" are both conditional

5. The FY2029 upside hasn't been endorsed by the company. The $18.5B/year is an analyst back-calculation from the warrant ceiling, not Marvell guidance. Murphy's exact words were "if all programs are realized at maximum performance"

Conclusion

What really changed at Marvell this quarter isn't how much it earned, but that its growth gained a line that wasn't in anyone's model.

A year ago, the bull-bear debate on Marvell had essentially one question: will it win custom ASICs, and how much? This call's answer: custom will be won, and bigger than anyone thought (the warrant is the endorsement) — but the biggest raise this quarter didn't come from custom; it came from scale-up optics. A category at only ~$300M last quarter became the main driver of a $1.5B annual raise.

This matters because it validates a bigger industry thesis: as scale-up domains grow, copper's reach and bandwidth density can't keep up, so optics must move next to the XPU and the switch. And Marvell is one of the few companies that holds switch silicon, SerDes, optical DSPs, TIAs/drivers, silicon photonics, and all three modulator technologies at once — it doesn't need to bet on the right route; it sells both ends of the link and everything in between.

Three metrics to track next quarter (Q3 FY2027, ending late October):

1. Whether non-GAAP gross margin holds at 57.5%–58.5%. This is the only hard metric to verify that custom dilution is "controllable, not a bottomless pit"

2. The FY2029 custom range given at the October 6 Investor Day. This will be the number the market uses to re-anchor Marvell's valuation, and the payoff point for every "wait for Investor Day" promise this quarter

3. Whether scale-up optics moves from a "category" to "specific design wins". This quarter management gave only direction and relative weight — no customer count, no dollar amount. If next quarter brings the same vagueness, it should be downgraded from "validated" back to "opportunity"

In one sentence: Marvell is halfway across the threshold from "AI CapEx beneficiary" to "one of the definers of AI network architecture" — the other half, see you on October 6.

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