MRVL (Trans): No change to ASIC guidance; data center to grow 60%+ next year.

Below is Dolphin Research's compiled Trans of Marvell FY27Q2 earnings call.

For the full take, see: Marvell: Growth Trails NVIDIA; Are Rich Valuations Due for a Rinse?.

I. Core Financial Highlights Recap

1) Guidance lifted across all three horizons. Q3, FY27 and FY28 were all raised.

a) Q3: revenue of $3.15bn (±5%), +15% QoQ and over +50% YoY; non-GAAP EPS of $1.05–1.15.

b) FY27: revenue around $12.0bn, approx. +45% YoY, up from $11.5bn last quarter; data center growth raised from ~50% to 60% (no ASIC growth figure provided this time; last quarter was 20%+).

c) FY28: revenue around $18.0bn, up from $16.5bn, with growth revised to ~50% from ~45%; data center growth raised to 60%+; custom revenue to more than double (unchanged from last quarter's outlook).

2) GPM structurally pressured by custom scale-up, while opex still operating with leverage. Mix is the driver on margins.

a) Q3 non-GAAP GPM guided to 57.5%–58.5% (vs. 58.9% this quarter), mainly on rapid custom ramp mix; Q4 and FY28 both held in this band.

b) FY27 non-GAAP opex about $2.55bn, above prior $2.45bn, with Q3 around $655mn; FY28 opex growth to be about half of revenue growth.

c) This quarter non-GAAP OPM was 36.6%, +180bps YoY and +160bps QoQ; Q4 expected to enter the 38%–40% target range, moving to the upper end within FY28.

3) Tax rate and share count. Q3 non-GAAP tax rate 11%, moving to ~13% in FY28 on higher revenue and profits.

Q3 basic shares about 900mn and diluted shares about 921mn, essentially flat QoQ.

4) Cash deployment shifting from shareholder returns to capacity locking. The focus pivots to prepayments.

a) Operating cash flow was $606mn this quarter, slightly down QoQ, mainly due to higher capacity prepayments to suppliers.

b) FY27 capacity prepayments remain on track at approx. $1.0bn, consistent with last quarter's framework; prepaids will offset future materials purchases, affecting both the balance sheet and operating cash flow.

c) This quarter buybacks totaled $200mn and dividends were $54mn, with ongoing buybacks to offset dilution; ending total debt was $4.96bn and inventory was $1.36bn. Total debt/EBITDA was 1.32x and net debt/EBITDA 0.27x.

II. Earnings Call Details

2.1 Management Commentary: Key Messages

1) Trajectory: accelerating, not peaking. Momentum is building rather than topping out.

a) Revenue hit a record $2.739bn, +13% QoQ and +37% YoY; non-GAAP EPS was $0.94, +40% YoY, both beating the mid-point of guidance.

b) Data center revenue was $2.17bn, 79% of total, +18% QoQ and +46% YoY, a clear acceleration vs. Q1 (+11%/+27%); Q3 is expected to grow over +20% QoQ and approx. +75% YoY, with further acceleration in Q4.

c) Carrier and Other were $568mn, -3% QoQ and +10% YoY; Q3 is guided to decline double-digit QoQ and YoY to mid-teens, then rebound meaningfully in Q4. Full-year growth is tracking near the +10% target.

2) Scale-out: three lines each approaching a $1bn run-rate. Execution is broad-based.

a) In optical DSP, 800G demand remains strong, with higher-speed products ramping fast and accelerating again in FY28; scale-out switching is set to more than double this year, driven by broader 51.2T deployments across customers.

b) Broadband analog TIAs and drivers are exceeding expectations; broadband analog, scale-across DCI modules and scale-out switching have each reached or are ahead of last quarter's path toward a $1bn annualized run-rate.

3) Scale-across: interconnecting multiple data centers unlocks new demand. Very large clusters are spanning sites.

These super-clusters will need total bandwidth over 10x that of today's front-end AI networks; the company is addressing this with coherent DSP-based 1.6T ZR/ZR+ DCI modules.

4) Scale-up: migrating from copper to optics is the largest new TAM. This is the next leg of growth.

a) As clusters get larger, copper distance and bandwidth limits are pushing customers to optical interconnects, and to three classes of proprietary switching: UALink, ESUN and NVLink.

b) Technology choices have not converged: customers are evaluating NPO vs. CPO packaging and MZM/EAM/MRM modulation, each with cost/power/maturity trade-offs; copper and optics will coexist for years, but some customers will deploy scale-up optical interconnects as early as next year.

c) On switching, the firm is developing UAL and ESUN in-house and collaborating with NVIDIA on NVLink Fusion; it is in deep discussions with multiple tier-1 customers, with each project's lifetime revenue opportunity measured in the multi-billions.

5) Custom: the warrant agreement expands the long-term scope. The commercial envelope is broader.

Last week's 8-K disclosed an expanded commercial agreement plus warrants with a large hyperscaler, spanning in-flight programs, new design wins and future potential projects; related revenue before FY28 is already embedded in the existing custom target.

6) People and calendar. Leadership changes and Investor Day timing were outlined.

a) The CFO transitioned in mid-Jun from Willem Meintjes to Dan Durn.

b) Head of IR Ashish Saran will retire in Apr 2027, to be succeeded by former Deutsche Bank semi analyst Ross Seymore; Analyst Day is set for Oct 6 in New York.

2.2 Q&A

Q: For the warrant agreement, which products drive the associated revenue? Can you split inference accelerators vs. XPU attach?

A: No product-level mix disclosed; coverage is broad, with the peak opportunity spanning the next 6–6.5 years. The agreement spans multiple lines, including inference accelerators, storage controllers, NICs, memory interface controllers and near-memory computing.

The company defined the XPU attach category two years ago and detailed it at the Custom Chips event in Jun 2025; management said every forecast so far has been 'too low', and the opportunity keeps expanding. They declined to name any single product, emphasizing that the total envelope is transformational in scale for Marvell.

Q: Analysts estimate the agreement at ~$120bn over six years, or ~$18.5bn per year, while custom in CY2027 is still only $5–6bn; how do you view the prior FY29 custom target of $10–11bn now?

A: No new FY29 number, but bias is upward vs. the $10bn handle. FY28 revenue tied to the warrant is already in guidance, as some programs are underway; programs in execution or just ramping to volume will mainly contribute in FY29.

For FY28, management reiterated only that custom is set to 'more than double', without a ceiling or precise quantification. The $10bn for FY29 is not new — first stated in Jun 2025, with a range of $8–10bn given in Apr 2024.

Management said the agreement substantially raises the revenue opportunity for custom, with some portion likely starting in FY29, but declined to quantify on the call; they will lay out the full revenue step-up — beyond FY29 and into the end of the decade — at Analyst Day.

Q: How is the other large XPU program progressing? The market assumes only $600–700mn next year — what is the long-term scope?

A: No quantification for that program; the largest piece of the $1.5bn raise comes from scale-up optical interconnect. The program is indeed part of the 'more than double' for custom, with quarterly progress in design execution, supply and commercial terms, and opportunity size — and it is 'only getting better'.

Of the $1.5bn FY28 raise, management flagged the heaviest contribution from scale-up optical interconnect and switching, with custom only part of it. They maintained this is one of Marvell's largest revenue opportunities, confident in year-on-year layering next year and beyond, but gave no dollar figure.

Q: How is CXL evolving in architecture and ecosystem?

A: CXL is in large-scale deployment at multiple hyperscalers, with tight memory supply driving further adoption. This multi-year internal investment, initially aimed at server/compute-centric memory architectures, has proven well-suited to memory expansion and inference, with heterogeneous customer architectures.

Demand comes from two sources: inference requires memory expansion, and current memory supply is tight, prompting customers to modify and adjust designs and use more of this technology. The company won new design-ins over the past two quarters and sees the opportunity as 'huge', with memory expansion to be a dedicated topic at Analyst Day.

Q: How does scale-up optical interconnect track vs. the initial targets? Are Celestial AI numbers materializing?

A: Outlook moved up: CPO at $150mn next year, and scale-up optical interconnect overall at $300mn. Last quarter, the framework was Celestial AI's CPO photonic fabric at ~$150mn next year and scale-up optical interconnect, including NPO, at ~$300mn overall.

Both baselines are moving higher and are key drivers of the $1.5bn FY28 raise. The uplift is not a single-point story and is broadening.

Management stressed that CPO is critical in the next two years, while NPO is scaling in parallel at a faster clip; on NPO, the company participates with SiGe broadband TIAs and drivers (high share and content), and a light engine developed in-house over years and shown at OFC.

Just like 800G to 1.6T is not a one-time switch, optical interconnect will see coexistence rather than 'either-or', with customers adopting multiple approaches simultaneously. Management believes next year's scale is much larger than a quarter ago, and the following year 'far larger than imagined when we looked at Celestial last year'.

Q: What is the progress on copper-based scale-up? Is copper leadership the foundation for optical wins, or are they independent decisions?

A: Copper switching is shipping; long-term wins will depend on end-to-end optical capability. On Ethernet, the Teralinx architecture from the 2021 Innovium acquisition has clear momentum, with a broader customer base and meaningful revenue, proving volume delivery of complex switch silicon.

This capability portably extends to scale-out, and at 100T the scale-out approach also applies to scale-up; progress on optical interconnect is independent. UAL switching also sees heavy investment, with development on track and able to support any electrical or copper approach.

Management believes customers now trust the company to deliver complex switches with large reticles and high SerDes speeds — where many others have failed. The next bar is a credible, fully integrated optical roadmap from NPO to CPO, since scale-up customers need a single vendor to anchor both ends of the link.

Existing copper/electrical solutions will do well, but management is more bullish on optical attach scaling over the next few years — which can nearly double the attach value available to switching.

Q: Based on the ~$120bn magnitude, will FY29–FY30 see big step-ups? Are margins comparable to existing biz.?

A: The math is fair; the impact lands mainly in FY29 and beyond, with margins following the custom model. Management said 'your math is not wrong' and confirmed that since much of next year's volume is already in guidance, the real impact is in FY29 and later — a 'monster' number at full potential.

They expect FY29+ custom numbers to be well above prior Street models — investors had questioned whether $8–10bn was doable. They declined to quantify on the call and will provide scenario ranges at Analyst Day.

On margins, the company applies distinct financial models to custom vs. standard/commercial products, and this agreement follows the custom model. For scale context, revenue was $8.0bn last year, $12.0bn this year, and $18.0bn next year.

Q: Under cost trade-offs among compute, networking and memory (echoing NVIDIA), are new design wins skewing to memory controllers and custom HBM, or is XPU also strengthening?

A: No product split; the substitutability under triple constraints is amplifying custom demand. Management agreed, noting that in a market with supply, power and architecture constraints, rapid custom/semi-custom and fast product tweaks to match architecture shifts are core strengths.

At the 2021 Analyst Day, the firm projected some customization at every network hop — not only in accelerators — and that is materializing five years on. CXL and memory expansion are one example, and AI inference accelerators optimized for inference are another (also within the warrant's scope).

The industry has shifted rapidly from training to inference and now to monetization, with customers optimizing cost and performance per token; architecture iteration is unprecedented, lifting not only custom, but also CPO/NPO and the full optical and copper chain.

Q: How have sub-segment growth guides changed vs. last quarter?

A: The $500mn raise for this year and $1.5bn for next year are broad-based, with connectivity the largest contributor. Within connectivity, beyond scale-up optical interconnect (NPO/CPO), growth for optical modules and optical DSPs serving scale-out has also been revised higher.

Switching expectations for this year and next are also up. Together with custom, these areas are all tracking above prior expectations, but net-net connectivity is the largest piece of the $1.5bn uplift.

Q: Data center is guided to grow over 20% QoQ in Q3, but GPM is down 90bps QoQ; is this due to faster custom growth dragging mix, or is there another mix shift within data center?

A: Confirmed: Q3 GPM compression is mainly custom ramp; Q4/FY28 remain in the same range. The CFO said Q2, Q3 and Q4 are sequential accelerations, with Q3 driven mainly by mix — the strong custom ramp directly shows up in the margin curve, a ramp signposted multiple times.

Q4 has an even larger step, implied by the $12.0bn full-year target, and it is broad-based: custom, connectivity, and a rebound in Carrier and Other all contribute. Q4 GPM sits in the same range as Q3; the initial view for FY28 is the same as 2H of this fiscal year, ultimately depending on product mix.

OPM still has meaningful leverage: it rises QoQ in Q3, enters the 38%–40% long-term target range by fiscal year-end, and moves to the upper end within FY28; the long-term target model will be reset at Analyst Day in a few weeks.

<End of text>

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure