Marvell's Interconnect Business Set to Double the Company in Three Years, With Vertical Scaling Optics and Custom ASICs as the Biggest Growth Drivers
Alina Collins
Marvell disclosed at the Evercore conference that its interconnect business is growing over 70% this year and will materially outpace cloud capex next year, with scale-up optics and custom ASIC entering steep ramp trajectories from low bases — effectively rebuilding another Marvell from incremental businesses alone.
Why can interconnect growth decouple from cloud capex?
Management stated explicitly: interconnect growth next year will materially outpace cloud capital expenditure. This means → Marvell's growth is no longer just riding the industry wave — it is driven by share gains and product-mix upgrades, i.e. company-specific alpha.
Two parallel curves underpin this: scale-out (mature PAM DSP, TIA, and driver products delivering steady revenue) + scale-up (ramping from a near-zero base on a steep trajectory).
In plain terms = scale-out is "the established business holding steady," scale-up is "the new business surging." Both legs running together is what supports outperformance.
How steep is the scale-up optics revenue curve?
Marvell doubled its scale-up optics revenue guide for next year from $150 million to $300 million. The figure covers the full product suite — optical engines, TIAs, and drivers — not just Celestial AI (an optical-interconnect company Marvell acquired) alone.
The ramp timeline was quantified for the first time: quarterly run-rate of ~$500 million annualized by end of next year → doubling to ~$1 billion annualized by end of 2028. That is roughly five quarters from production start to a $1 billion scale.
A leading hyperscale customer selected the solution after multiple rigorous qualification rounds including live-product testing; capacity is production-ready. This means → "technology-validation risk" is largely cleared — what remains is execution and ramp.
Why is the custom ASIC doubling considered high-certainty?
The XPU/ASIC business is ~$2 billion this year, growing over 20%, with a target to double to $4 billion+ next year.
Growth sources break into roughly equal thirds: ① natural growth of existing programs ② 10+ XPU attach programs entering production next year ③ entirely new XPU programs. In plain terms = the doubling does not hinge on a single large customer — a substantial portion is already "locked in" via programs approaching production.
Marvell's moat rests on proprietary commercial IP: high-speed SerDes, die-to-die interconnect, custom HBM interfaces, high-density SRAM, and advanced packaging — capabilities that third-party design houses cannot replicate.
Which assets is the market undervaluing?
TIA and driver business: acquired via Inphi and predating PAM DSP, this line is on track to reach $1 billion+ scale. Management openly called it "under-recognized by the market" and signaled plans to disclose its scale more visibly. This reflects a classic precursor to a management-led revaluation campaign.
XPU attach upside optionality: the 2028 target is ~$3 billion (CXL and NICs each contributing $1 billion+, plus storage and security accelerators from the Cavium heritage). Crucially, this guidance excludes incremental CPU demand from Agent AI. This means → with Agent AI driving higher CPU utilization and elevated DRAM prices reinforcing the value of CXL — a protocol that lets CPUs share a common memory pool — this is an uncaptured call option not yet in consensus models.
What does the Nvidia partnership really mean?
The core is not component supply. Marvell is using its networking and custom IP to bridge Nvidia's commercial infrastructure with customers' proprietary infrastructure.
In plain terms = hyperscalers no longer face a binary choice between Nvidia's stack and their own custom stack. Marvell acts as the "translation layer" between the two, fundamentally expanding the joint addressable market for both parties.
On the optics side (NPO/CPO collaboration) and AI RAN (Octane baseband integrated with Nvidia's platform), Marvell locks in medium-term and emerging-market entry points respectively.
How large is the long-term opportunity — and where is the risk?
Marvell reiterated a 5× expansion of the DCI (data-center interconnect) TAM by 2030 and, unusually, said the forecast "still has room for upward revision" as AI and mixture-of-experts models are driving network complexity beyond expectations set just one or two years ago.
Scale-across back-end data volumes run 10× front-end, requiring bandwidth upgrades from today's mainstream 400G/800G to 1.6T. Of last year's ~$3 billion interconnect revenue, scale-across contributed only ~$500 million; meaningful revenue is expected from next year onward — yet another growth curve off a low base.
Risk boundary: lasers face no near-term supply constraint, but Marvell has no plans to vertically integrate laser production. Long-term laser-source demand depends on industry-wide investment. This means → the scale-up optics ramp carries an external dependency outside Marvell's direct control.
Content is for reference only, not financial advice.