top of page

📢 STT Premium is now open

Every free article stays free — nothing moves behind a paywall. Premium adds the weekly deep-dive, full earnings breakdowns and all long-form reports. Note: premium articles are published in Traditional Chinese.

Free keeps you current. Premium lets you judge.

NT$199/month · NT$2,000/year (about NT$167/month)

👉 Subscribe: vocus.cc/salon/simpletechtrend

Barclays 23rd Annual Global Technology Conference | Marvell Technology | From Scale-Out to Scale-Up: Marvell's Full-Stack AI Interconnect Strategy and ASIC Growth Path

Oct 5
3 min read

This Barclays session featured Marvell (MRVL) President & COO Chris Koopmans and SVP of IR Ashish Saran.

In the conversation, Marvell presented a very clear strategic blueprint: not only defending its dominance in scale-out (rack-to-rack) optical connectivity, but also using the newly acquired Celestial AI and UALink switches to attack the scale-up (in-rack) market long monopolized by Nvidia or dominated by copper. Management also pushed back forcefully, with data, on recent market rumors of lost ASIC orders.

Here are my key takeaways and analysis:


1. Defining the New Battlefield: Scale-Out vs. Scale-Up

Through its acquisitions of Inphi and Innovium, Marvell built a commanding lead in scale-out (rack-to-rack, DC-to-DC) connectivity. But in an AI data center, 80–90% of traffic actually happens in scale-up (in-rack GPU-to-GPU interconnect).

  • Status quo: Marvell already rules the scale-out market (optical DSPs, switches). In scale-up, however, there is currently only one mainstream architecture (an implicit reference to Nvidia NVLink), most of it still on passive copper, and Marvell's share is minimal.

  • Opportunity: Marvell expects that over the next five years, scale-up architectures will move from passive copper to active switching and optics — an entirely new, zero-to-one TAM (total addressable market).

  • Weapons: Marvell will launch a 115T UALink switch (sampling 2H 2025) and the Photonic Fabric technology gained through the Celestial AI acquisition, going all-in on a territory that was never theirs.


2. The Strategic Meaning of the Celestial AI Acquisition: Scale-Up Optics

The market often conflates where CPO (co-packaged optics) applies. Chris Koopmans drew a sharp distinction:

  • CPO for scale-out (delayed): Because pluggable optical modules (800G → 1.6T → 3.2T) are advancing so quickly, customers prefer to keep using pluggables, keeping CPO adoption low in scale-out.

  • CPO for scale-up (a must-have): Inside the rack, when reach shrinks to millimeters or inches and bandwidth density requirements are extreme, electrical signaling hits its physical limits and pluggables simply don't fit. That's where Celestial AI's Photonic Fabric becomes the only solution.

  • Specs: Celestial's first-generation chiplet delivers up to 16 terabits of bandwidth (10x the density of today's 1.6T scale-out modules).

  • Timeline: Expected to reach a $500M run-rate by the end of 2027 and double to $1B by the end of 2028 — the first large-scale commercial deployment of scale-up optical interconnect.


3. Optics and the Golden Cable Initiative

  • 1.6T leadership: Marvell claims to be the main driver of 1.6T DSPs today. Because 1.6T requires an advanced 3nm process and low-power design, with long qualification cycles (which began a year ago), this forms a strong moat.

  • Golden Cable (AEC): In the copper cable market, hyperscalers want multiple sources. Marvell's strategy is to provide a reference design (Golden Cable) to cable makers (such as Foxconn), built around Marvell's DSP. This lets Marvell dominate the AEC market through DSP sales without making cables itself.


4. The Growth Path for Custom ASICs (XPUs)

Responding to market rumors (such as losing Customer C's orders), Marvell was very firm, flatly rejecting them and stating it already has purchase orders for the full year.

  • 2026 outlook: Revenue will come mainly from existing Customer A (broader adoption) and continued ramp of existing programs.

  • 2027 inflection: Custom AI revenue is expected to double, driven by:

    1. Customer C's large XPU program entering volume production (new ramp).

    2. XPU attach (companion chips): An often-overlooked but high-potential segment, including SmartNICs and CXL memory controllers.

  • XPU attach potential: As LLMs need large KV caches, CXL memory expansion becomes a must-have. Marvell estimates that SmartNIC + CXL alone could contribute $2B in revenue by 2028.


5. The Protocol War: NVLink, UALink, and Ethernet

In the scale-up protocol war, Marvell is pursuing an "optionality" strategy:

  • NVLink Fusion: Marvell participates by supplying chiplets or interconnect IP, letting customers plug into the Nvidia ecosystem.

  • UALink: The open standard Marvell is championing, designed specifically for scale-up (based on PCIe/memory semantics, but faster).

  • Ethernet: Marvell's home turf, where it continues to supply hyperscale Ethernet switches.

This strategy positions Marvell as an "arms dealer": whichever architecture customers choose (closed vs. open), they need Marvell's SerDes, DSP, or switch IP.


Conclusion

Marvell is at a pivotal point in transforming from an "optical component supplier" into a "full-stack AI infrastructure platform."

Investors used to view Marvell mainly as an optical DSP play riding cloud CapEx, but now they need to watch the new scale-up opportunity (Celestial AI/UALink) and the step-change growth of custom ASICs in 2027. Management expressed strong confidence in $5 EPS potential for calendar year 2027, backed mainly by orders that are already designed in with long lead times.

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page