OCP Global Summit 2026 | Lumilens | Samuel Liu | NPO to Enable Scale-up & Scale-out Interconnects

3 days ago

7 min read

1. Why a Two-Year-Old Startup Talked Manufacturing at an NPO Workshop

Samuel Liu opened with a joke: every speaker before him came from a large company, and he was the only one representing a startup. Lumilens is about two years old and went public with its launch only a few months ago. Two years ago, industry friends asked him why anyone would start yet another optical transceiver company, and he admits he had no good answer; now the answer is simple: the NPO market has arrived.

The market figures on his slide come from two firms:

AI data center infrastructure grows to $2.8 trillion in 2029 (Citigroup, 2025), with the slide noting that "network has become the bottleneck."
Photonic interconnects for AI grow from $28B in 2026 to $65B+ in 2031 (LightCounting, 2026), roughly an 18% CAGR (STT calculation).

These forecasts line up with the LightCounting framework we broke down in CPO Is No Longer "Crying Wolf". For a startup, the point is not how big the pie is, but that while it grows, the architecture is changing generations, and that is what opens the door to new entrants.

Startups get in not because their technology is stronger, but because the rules happen to be changing.

2. Breaking Down Scale-up: NPO Is What Can Be Deployed Today Once Copper Runs Out

Lumilens' read on scale-up is blunt: copper dominates today, the design is still evolving, and it will move to NPO and CPO. The slide lists four reasons:

Front-panel density: faceplate area, not silicon, caps how many GPUs a scale-up switch can reach. NPO moves the optical engine onto the board, so the switch gets more ports.
Power per bit: removing the long electrical host channel and the module DSP takes power out of every GPU-to-GPU hop. Liu added verbally that the target is below 10 pJ/bit.
Cost per lane: short reach and a low link budget allow relaxed optics, fewer components and higher assembly yield. His example: a 6.4T NPO carries 4x the bandwidth of a 1.6T pluggable, and its price cannot also be 4x.
Short reach wins: inside a controlled chassis, under roughly 20 m, linear optics close the link without a retimer in the module.

▲ Lumilens' four reasons scale-up moves to NPO: front-panel density, power per bit, cost per lane, and winning at short reach within 20 meters (Image source: Lumilens | 2026 OCP Global Summit presentation (05:50))

The "Why NPO first" strip at the bottom is the clearest positioning of the talk: pluggables are serviceable but faceplate- and power-bound; CPO has the lowest pJ/bit but is the hardest to service and to produce in volume; NPO puts the engine beside the ASIC and is the compromise that can be deployed now. That is close to our view in NPO Is Not CPO's Appetizer, except that Lumilens still treats CPO as the long-term destination.

One easily missed detail: Liu noted that scale-up NPO mostly sits not next to the GPU but close to the scale-up switch, supporting high-density, cross-rack scale-up switch connectivity.

3. Breaking Down Scale-out: NPO Takes Only the Shortest Hop

On scale-out the stance is more conservative. Lumilens says scale-out is dominated by pluggables that are still growing; NPO is not there to displace them but to open a new market. He split scale-out into four options:

Option	Role	Reach	Position
OSFP / QSFP-DD pluggables	Full retimed, LRO, TRO, field replaceable	100 m to 2 km	Carries the fabric today
Next-gen XPO pluggables	Lower power and higher density in the same cage	Same as pluggables	Density step, same operations
NPO	Linear optics, limited budget, board-mounted or pluggable	Short in-row hops	NIC to ToR / Leaf
CPO	Optical engine inside the switch package	Highest-density tiers	Switch tiers by 2030

▲ Lumilens' scale-out portfolio: NPO covers only the short NIC-to-ToR/Leaf hop, retimed pluggables and XPO still handle 100 m to 2 km, and CPO enters at the switch level in 2030 (Image source: Simple Tech Trend; data from Lumilens' 2026 OCP Global Summit presentation)

He also noted that 51.2T switches with 64 ports of 1.6T pluggables are already deploying, and that the density advantage of XPO and NPO grows as networks move to 200T switches. The judgment is pragmatic: no single option will take all of scale-out, and NPO plays a supporting role there, not the lead.

Scale-out needs a portfolio, not one answer.

4. Breaking Down the Laser: Use ILS Where You Can, ELS Only Where You Must

This is where Lumilens' technical stance is sharpest. It splits NPO optical engines into two types:

ILS (integrated laser source, preferred): laser, silicon photonics modulator, driver and TIA in one package. Benefits: no PM fiber, lower required laser output power, smaller footprint, fewer optical connectors and splices; the trade-off is that thermal headroom and laser reliability must be managed in-house.
ELS (external laser source): the laser sits in a remote module, away from the heat, and can be replaced in the field; the costs are added loss, more parts and a larger footprint.

▲ Lumilens' NPO laser architecture: ILS is the preferred option, with ELS used only when customers require serviceability (Image source: Lumilens | 2026 OCP Global Summit presentation (11:38))

Liu's verbal logic was cost: 400 mW-class lasers are already on the market, but much of that power is lost on the way from the laser to the modulator, and an integrated laser eliminates that loss. He was candid that some customers specify ELS for serviceability, and Lumilens supports both. This matches the direction we discussed in CPO Lasers Don't Have to Chase 400 mW: NPO Integrated Lasers: in this NPO generation, integrating the laser lets the laser power spec come down.

Another approach worth noting is the laser optical sub-assembly: although the laser is integrated in the module, it is tested, assembled and burned in as a separate building block before entering the module, to cut module build loss and raise yield.

5. Breaking Down Manufacturing: The Winner Is Whoever Ramps Fastest

Liu said the topic most speakers skipped is what actually decides NPO and CPO: who controls manufacturing and who can improve yield. Building samples, running live demos and showing good data is not hard; the hard part is who can ramp to volume fastest.

The slide breaks volume NPO into four stages and stresses co-designing the assembly and test flow with the system builder, not just the module:

Silicon photonics + electronic IC: wafer-level process control and co-designed EIC/PIC.
Advanced packaging: optical engine assembly, fiber attach and thermal path.
Known-good engine: engine-level test and burn-in before the system build.
System integration: attach, rework and repair flows the factory can actually run.

▲ Lumilens: volume NPO is won in manufacturing, and the optical engine must be co-designed with the GPU rack, compute tray and scale-up/scale-out switches (Image source: Lumilens | 2026 OCP Global Summit presentation (15:50))

He criticized the optical module industry's habit of drawing a demarcation line at the module boundary: "You define the spec, I deliver, and the rest is the system's problem." NPO cannot work that way; signal integrity, power integrity and thermal design have to be co-designed with the system.

6. Lumilens' Differentiation: Team Mix and an Asset-Light Model

The slide lists four differentiators:

Single platform: the LumiCore platform spans pluggables, NPO and CPO, with in-house silicon photonics, high-power lasers and optical interposers.
Systems background: the team comes from Juniper, Cisco, Meta, Marvell, Lumentum and Coherent. Verbally, he added that roughly one-third are system domain experts and roughly one-third are advanced packaging people.
Manufacturing ownership: Lumilens owns the process recipes, automation, test equipment design and MES, while partners provide operators and facilities; the slide claims much lower capex than IDMs.
Hyperscaler design partner: a tight design relationship with a top hyperscaler, aiming for first-pass qualification.

▲ Lumilens' four differentiators: the single LumiCore platform, systems background, manufacturing autonomy and hyperscaler design partners (Image source: Lumilens | 2026 OCP Global Summit presentation (19:46))

His closing judgment is worth noting: NPO and CPO will increasingly become custom products, designed for one large customer's requirements, rather than standard pluggables competing on capacity and price.

7. Risks and Counterarguments

No hard specs at all: the talk gave no per-lane rate, total engine bandwidth, measured pJ/bit or yield figures. "Below 10 pJ/bit" and "under 20 m" are directional statements.
ILS thermal and reliability is an admitted challenge: putting the laser in the hot environment beside the ASIC, with no lifetime data published yet.
Customization cuts both ways: tying to one hyperscaler speeds qualification but concentrates revenue, and the vendor can be replaced at the next generation.
Asset-light manufacturing depends on partners: recipes and MES stay in-house while operators and facilities belong to partners, so process transfer and quality consistency are real challenges.
Distance from the incumbents: Broadcom is already pushing CPO and pluggable vendors are doing XPO and NPO, leaving a startup at a disadvantage in capacity and supply-chain bargaining power.

8. What It Means for Taiwan's Supply Chain

"Partners provide operators and facilities" is a contract-manufacturing signal. Lumilens has not named its partners, but its model keeps recipes, automation and MES in-house and hands the production lines out. Optical contract manufacturing and EMS are Taiwan's strengths; it is worth tracking when such startups' manufacturing partners become public.
Switch-side design decisions land directly on system makers. The slide's scale-up switch design points, including high-radix NPO faceplate escape, thermal budget and repair strategy, plus engine placement, host channel length and board-level assembly yield on the compute tray, all fall within the scope of switch and server ODMs.
Laser optical sub-assemblies and engine-level burn-in drive test demand. Integrated lasers must be tested and burned in separately before entering the module, and known-good engines must be tested and burned in before system build, adding volume for test equipment and test services.

9. Summary

Lumilens gave no spec sheet, but it made one point very clearly: NPO competition has shifted from "whose pJ/bit is lowest" to "who can ramp first." In scale-up it positions NPO as the first deployable option after copper, in scale-out it takes only the NIC-to-ToR hop, and on lasers it bets on integration.

Timing needs to be read separately: scale-up NPO is happening now, scale-out NPO fills a gap, and CPO reaches switch tiers only by 2030. Over the next 12 months, three things to watch: the first public bandwidth and pJ/bit figures for LumiCore, whether the hyperscaler first-pass qualification succeeds, and who the manufacturing partners are.

A startup without its own factory betting that NPO will be won in manufacturing tells you where the main battlefield now is.

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