OCP Global Summit 2026 | Nokia | Armaghan Eshaghi | Architecting Near-Packaged Optics for AI Datacenters

1. Why Nokia is the one talking about NPO trade-offs

There were many NPO sessions at this year's OCP, and Eshaghi opened by admitting the audience had probably heard the term a hundred times already. His topic was not "why NPO" but this: the closer optics move to the ASIC, the less settled the modulator, laser and packaging choices become, and the three are coupled, so they cannot be designed separately.

Nokia stresses that it is not tied to any single technology and has capabilities in InP, SiPh and TFLN; but the last technical slide shows three InP fabs. In other words, this is an InP manufacturer explaining why NPO gives monolithic InP lasers a second chance. That stance does not change the data, but it is worth keeping in mind when reading the comparison tables.

2. Power: NPO and CPO are in the same class; the difference is replaceability

Nokia ranks four link-power tiers by SerDes count:

Architecture	Electronics	Link power	Service model
Re-timed pluggable	1 SerDes + 1 DSP	~15 pJ/bit	Hot-swap the whole module
LRO	1 SerDes + half a DSP	~10 pJ/bit	Hot-swap the whole module
LPO / NPO / CPO	Host SerDes only	~6 pJ/bit	NPO: replace engine; CPO: laser externalized
Die-to-die integrated optics	0 SerDes	Below 4 pJ/bit	Package-level; years from volume

Going from 15 to 6 is a reduction of about 60%. Eshaghi added verbally that Nokia's own LPO pluggable runs at about 5 pJ/bit. He also cited reports from Oracle at ECOC and from Alibaba earlier that day: LPO failure rates are lower than fully re-timed modules because there are fewer SerDes and DSPs to fail.

▲ Four tiers of link power: about 15 for retimed, about 10 for LRO, about 6 for LPO/NPO/CPO, and below 4 pJ/bit for integrated optics; NPO's service model is a replaceable optical engine (Image source: Nokia | 2026 OCP Global Summit presentation (p. 2))

The key line on this slide is the bottom row. CPO's service model reads "laser externalized": the engine is soldered onto the package and hard to replace, so the laser naturally gets pushed to the front panel. An NPO engine can itself be replaced, which loosens that premise.

3. Modulators: five options, none wins everything

Criterion	InP EML/EAM	InP MZM	SiPh MZM	SiPh MRM	TFLN MZM
Drive voltage	Low	Low	High	Lowest	Low
Bandwidth headroom at 400G/lane	High	High	RC-limited	Moderate	Highest
Modulator length	Short	Moderate	Long	Smallest	Longest
Loss (incl. coupling)	Moderate	Moderate	Higher	Moderate	Lowest
Thermal sensitivity	May need TEC	May need TEC	Near-athermal	Active tuning	Excellent

STT's count of the ratings: TFLN takes three "best" cells but is weakest on length; SiPh MRM takes two "best" cells but is weakest on thermal sensitivity; SiPh MZM is weakest on drive voltage, and Eshaghi acknowledged that both silicon photonics modulators still face challenges at 400G/lane. The two InP options have no "best" and no "weakest" cell, making them the most balanced on the table.

Two verbal notes: InP's TEC may be avoidable with liquid cooling, and MRM active tuning gets harder as systems scale.

4. Laser placement: every step outward costs more laser output

The most useful slide was the laser placement table, which treats laser-to-modulator distance as a continuum:

Laser placement	Laser-to-modulator loss	Thermal	Service	Manufacturing
Monolithic (InP)	~0.25 dB	In engine heat zone	Replace engine	Single InP wafer
Integrated on PIC	0.5–1 dB	In engine heat zone	Replace engine	Two platforms, yields compound
In-package	~1.5 dB	In engine heat zone	Replace engine	Known-good, burned-in laser die
External source (ELSFP)	~2 dB	Cooler front panel	Replace laser in field	One source shared by 4–16 lanes; PM fiber handling

▲ Moving the laser from monolithic integration out to ELSFP raises loss from about 0.25 dB to about 2 dB, in exchange for thermal isolation and field replaceability (Image source: Nokia | 2026 OCP Global Summit presentation (p. 5))

Converting dB to multiples makes it more intuitive. Using monolithic as the baseline, delivering the same optical power to the modulator requires about 1.33x the laser output for an in-package laser and about 1.5x for ELSFP (STT estimate, counting only the laser-to-modulator loss on the slide).

▲ The farther the laser sits from the modulator, the greater the extra loss; ELSFP adds about 1.75 dB over monolithic integration, requiring about 1.5x the laser output (Image source: Simple Tech Trend original, data from Nokia 2026 OCP Global Summit presentation)

Nokia lists four benefits of internal lasers: simpler system design, less reliance on ultra-high-power lasers (UHPL), no PM fiber, and relief for front-panel ELSFP density. Eshaghi noted verbally that UHPL capacity is tight today. The front-panel point is especially clear in the 200T generation: assuming 200G per lane, 204.8T is 1,024 lanes, and with one source per 4–16 lanes the front panel needs 64–256 laser sources (STT estimate; halved at 400G/lane).

This is where Nokia's stance is clearest: all three internal options sit "in the engine heat zone," and Nokia's answer is that the NPO engine has already left the ASIC heat zone. We discussed the same direction in CPO lasers need not chase 400 mW: integrated lasers in NPO; what Nokia adds is the loss number for each position.

5. Packaging: sockets, thermal path and assembly

Packaging determines signal integrity, thermals and the service model, in four parts:

• Host attach: the engine can sit on the ASIC substrate (connectorized), on the main board (worst RF performance) or on an interposer card. Interposer card contacts include metal-contact compression LGA, elastomer membranes with metal contacts, and conductive elastomers.

• OE socket: a two-piece connector (such as the CPX Type 1/2 MSA or other proprietary approaches), or soldered down to the paddle card. Separable interfaces preserve serviceability at the cost of some electrical performance. We covered the CPX ecosystem in ECOC2026: Ciena's Open CPX turns 6.4T optical engines into off-the-shelf parts.

• Thermal path: share the ASIC cold plate while keeping lasers cool. The slide says "a cold plate works only if the heat path is designed in," and Eshaghi added verbally that it must be designed in from day one.

• Assembly: known-good-die lasers and PICs, wafer-level test, passive alignment and automation.

6. Nokia's bet: three InP fabs and a 204.8T NPO switch

The final technical slide is titled "Manufacturing scale and performance will pick the winners." Nokia's demand signal: customers are asking for roughly 10x InP volume by around 2030. Its answer is three InP fabs: the existing 3-inch fab in Sunnyvale, plus two new 6-inch fabs in San Jose and Arizona.

The platform covers monolithic, internal and external lasers, and InP, SiPh and TFLN modulators. The "Nokia DC Switch with NPO Engines" drawing on the right shows four engines on each side of the ASIC, each labeled 12.8T, for 204.8T in total (STT sum).

▲ Nokia's 6-inch InP platform and three wafer fabs; on the right, an NPO switch concept with 16 × 12.8T optical engines surrounding the ASIC (Image source: Nokia | 2026 OCP Global Summit presentation (p. 7))

▲ Modulator and laser are a coupled choice: InP can integrate the laser monolithically, while TFLN and SiPh must pair with integrated, in-package or external light sources (Image source: Simple Tech Trend original, data from Nokia 2026 OCP Global Summit presentation)

7. Risks and counterarguments

• The heat-zone problem moves, it does not disappear. All three internal lasers sit in the engine heat zone, and the NPO engine still shares the ASIC cold plate. Nokia gave no data on laser efficiency and lifetime at 200T-generation thermal density, and it keeps external sources as its own hedge for the hottest, densest scenarios.

• Replacing an engine costs more than replacing a laser. A failed internal laser means swapping the whole engine; with ELSFP only the laser is swapped. Which has the lower total service cost depends on the real FIT of lasers and engines, which the talk did not provide.

• The loss figures are typical values, and internal lasers do not win every row. The 0.25 dB to 2 dB values carry no stated conditions; the same table rates RIN as Good for monolithic and PIC-integrated lasers and Great for in-package and ELSFP. Because one ELSFP source feeds 4–16 lanes, amortizing laser count and cost may partly offset the 1.5x power gap.

• The framing leans toward InP. The table rates InP as the most balanced, and Nokia owns InP fabs, so it should be read as one vendor's assessment rather than a neutral benchmark.

• No timelines or product specs. The 204.8T switch is only a concept drawing, and no volume dates were given for the 6-inch fabs.

8. What it means for Taiwan's supply chain

• The talk named no Taiwanese companies; what follows are architecture-level observations. Laser placement changes the shape of laser shipments: if 200T-generation NPO adopts internal lasers, demand shifts from high-power CW laser modules at the front panel to laser die integrated in the engine; if it adopts ELSFP, demand stays concentrated in high-power lasers shared across many lanes. CW laser and ELSFP vendors need to be ready for both forms.

• The InP capacity race has moved from chip fabs to wafer size. Nokia's two in-house 6-inch InP fabs are vertical integration, and its stated support for multiple architectures positions it as a supplier of InP sources and optical engines. For Taiwan's compound semiconductor players, the entry ticket is 6-inch InP yield and capacity.

• Packaging is the segment Taiwanese firms know best. CPX sockets, elastomer and LGA contacts, interposer cards and cold plates shared with the ASIC fall within the scope of connector, PCB and thermal suppliers; the more NPO emphasizes separability, the more this segment is worth. Further reading: NPO is not the appetizer for CPO.

9. Summary

The value of this talk is that it turns NPO serviceability from an operations topic into a prerequisite for component choices. Because a CPO engine cannot be replaced, the laser moves to the front panel at the cost of about 2 dB of loss, PM fiber and panel density; because an NPO engine can be replaced, the laser can return inside the engine, cutting the loss to about 0.25 dB.

STT's view: the 200T generation will be the first test of the internal laser's comeback, and the biggest beneficiaries will be manufacturers that can take monolithic InP to 6-inch volume, Nokia among them. But the heat-zone and service-cost questions remain open, and external sources will not disappear. Three things to watch over the next 12–24 months: the laser configuration of 200T NPO switches, the volume timeline of Nokia's 6-inch InP fabs, and which light source TFLN and SiPh engines adopt.

CPO pushed the laser out of the optical engine, and NPO invites it back; those extra few dB of loss end up as a race for InP capacity.

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