OCP Global Summit 2026 | Microsoft | Yawei Yin | How NPO Can Continue the Ecosystem Success of Pluggable Optics

1. Why Microsoft Switched the Question from "Why" to "How"

Yawei Yin (Microsoft, Sr. Principal Network Developer) spoke for only 12 minutes at the IPEC NPO workshop, yet he was one of the few speakers talking from the buyer's and operator's side.

His starting point: pluggable optics built today's data center ecosystem on four pillars:

• PMD standards: SR, DR, FR, LR, ER and ZR unified reach, wavelength, lane count and link budget.

• Form factors: SFP28, QSFP56, QSFP112, OSFP, OSFP-XD, XPO and USFP made mechanical, power and thermal integration repeatable.

• Electrical and management interfaces: AUI, CAUI, CEI plus CMIS, module firmware and telemetry decoupled host, module and component innovation.

• Multi-vendor specialization: ASIC, DSP, optics, connectors, modules, systems and test equipment each became a specialty.

Yin put it bluntly: a hyperscaler can pick parts from different suppliers, and once they are in the switch they look and behave the same. That is what Microsoft does not want to lose in the NPO era.

▲ The four pillars of pluggable optics: PMD standards, form factors, electrical and management interfaces, and a multi-vendor supply chain; the conclusion is to keep the ecosystem model, not the physical module (Image source: Microsoft | 2026 OCP Global Summit presentation (2:38))

This matches our view in NPO Is Not an Appetizer for CPO: NPO's value is not only power, but whether it can carry the pluggable business model forward.

2. Breaking Down the Three-Dimensional Bottleneck of Pluggables

Yin split the pluggable bottleneck into three dimensions that constrain one another:

• Electrical reach: the higher the lane rate, the shorter the usable PCB channel and the thinner the margin.

• Thermal envelope: the more module power, the more heat concentrates at the switch faceplate.

• Faceplate bandwidth density: as radix and bandwidth rise, more optical capacity must fit into limited panel area.

He quoted supplier feedback: if you insist on OSFP, the switch has to be 2RU or 3RU. Form factors brought standards, but they also capped faceplate density.

Pluggables use a long PCB path with connectors; NPO uses a short PCB or flyover cable to a socketed optical engine next to the ASIC; CPO puts the optical engine in the package or interposer. The three lines at the bottom of the slide amount to Microsoft's stopping rule for integration:

• Electrical loss is the forcing function.

• Move the optical engine only as close as physics requires.

• Repairability and serviceability determine where integration stops.

▲ Optical engine placement in pluggable, NPO and CPO; electrical loss pushes the engine toward the ASIC, while serviceability decides where integration stops (Image source: Microsoft | 2026 OCP Global Summit presentation (5:46))

Microsoft does not believe "closer is always better". Huawei scored the two on five dimensions at ECOC (see ECOC 2026: Huawei Scores CPO vs NPO, NPO Leads 4:2); Microsoft boils the win condition down to one line: if you can repair it, you can move it closer.

3. Breaking Down DSP: Not Yes or No, but Port by Port

DSP solved yesterday's long channel but pushes the challenge onto tomorrow's power envelope. The slide lays it out in two columns:

What DSP provides	What DSP costs
Equalization and signal recovery	Module power
FEC and coding flexibility	Faceplate heat
Host and channel tolerance	Additional latency
Vendor and system interoperability	Complexity and qualification
Diagnostics and telemetry	Larger thermal solution

NPO shortens the electrical channel and reduces the compensation burden. Yin said most NPO designs assume linear channels work, but at 400G per lane some lanes may still fall short, so the industry is also discussing adding retiming or half retiming.

More notable: NPO means not every port needs the same signal-integrity scheme, or even the same optical engine, because some ports inside Microsoft face different PMD challenges. The slide's conclusion: NPO should support FRO, TRO/LRO and linear optics rather than prescribe one architecture.

For DSP vendors this is good news and bad news: good because DSP does not disappear in NPO; bad because it shifts from "standard in every module" to "optional per port".

4. Breaking Down FIT: 10x Is the Entry Ticket

Yin noted that CPO and NPO have been proposed for 10 to 20 years without becoming reality, and the root cause is this: once the optical engine moves inside the box, there is no module to pull from the faceplate, so what happens when it fails.

It is not just lasers that fail, but also modulators, the optical engine itself, thermal mechanics and liquid cooling. So the slide recasts reliability as a system property:

▲ Microsoft defines service impact as links × component FIT × time to repair × failure blast radius; NPO diagnostics must move to CMIS and the operations model is still undefined, with a verbal target of 10x lower FIT and 10x higher MTBF (Image source: Simple Tech Trend; data from Microsoft's 2026 OCP Global Summit presentation)

How the three architectures differ:

Item	Pluggable	NPO	CPO
Replacement unit	Module	Module or board	Package or system
Blast radius	Port or module	Module or board	Possibly the whole switch or accelerator
Diagnostics	CMIS (mature)	CMIS-like (must be standardized)	Architecture-dependent
Operational model	Mature	To be defined	New model

There is only one number, and it is spoken: FIT at least 10x smaller, MTBF at least 10x larger, with room to aim higher. The reason is that a hyperscaler cannot keep swapping entire boxes as often as it swaps pluggables; when something fails, traffic must keep flowing and GPU and CPU servers must keep running.

The line module makers should read most carefully: power savings must be evaluated together with reliability and operational cost. The "to be defined" in the NPO column is Microsoft publicly admitting the operating model has no answer yet.

5. NPO Is a Bridge, Not a Smaller Pluggable

Yin listed five groups of enabling technologies that make NPO practical and interoperable:

• Electrical interface: high-density 224G-class electrical connectors (the slide labels them 224G / 12.8T class), short-reach channels, PCB or flyover cable.

• Optical interface: detachable, manufacturable high-density fiber connectors with blind-mate support.

• Optical engine: silicon photonics, EML or VCSEL by application, 200G per lane and up, with external or integrated lasers.

• Power and cooling: heat spreaders, cold plates or airflow integration, plus power monitoring.

• Management and reliability: CMIS-derived management, telemetry, firmware and qualification.

▲ The spectrum from pluggable, on-board optics and NPO to socketed CPO and integrated CPO; NPO is a replaceable optical engine close to the ASIC but outside the package (Image source: Microsoft | 2026 OCP Global Summit presentation (11:02))

Microsoft's definition of NPO is precise: a replaceable optical engine placed next to the ASIC, connected through a standardized high-density electrical interface and a detachable optical interface, but not integrated into the ASIC package.

There is a tension here. Yin said whatever can be reused from electrical, optical and management standards should be reused, but there is no need to copy every existing standard; form factors will be redesigned, and whether a new standard is needed is open for discussion. He even expects NPO to look more like "a co-designed system customized to each data center's requirements". Standardize the interfaces, customize the shell: that is the biggest difference between Microsoft and the camp pushing a single socket spec.

6. Call to Action: Set the Boundaries Before Optimizing

The closing slide breaks NPO interoperability into six blocks and lists six kinds of participants: ASIC and system vendors, optical engine vendors, connector and fiber vendors, test equipment, standards bodies and cloud operators.

Block	What must be defined
Module envelope	Footprint, height, keep-out zone, extraction
Electrical interoperability	Lane rate, channel budget, linearity, link training
Optical interoperability	IEEE PMDs, optical connectors
Power and thermal classes	Power limits, cooling boundary, thermal reporting
Management interoperability	CMIS-derived firmware, telemetry, alarms
Reliability and operations	FIT allocation, qualification, field replacement, failure isolation

▲ Microsoft's NPO call to action: six interoperability blocks and six types of ecosystem participants — the same ecosystem principles, with the optical engine closer to the ASIC (Image source: Microsoft | 2026 OCP Global Summit presentation (11:46))

The order matters: define interfaces and ownership first, then optimize individual implementations. Ciena's Open CPX is one attempt to standardize the socketed optical engine (see ECOC 2026: Ciena's Open CPX Turns the 6.4T Optical Engine Back into Shelf Stock), but Microsoft did not endorse any specific spec in this talk.

7. Risks and Counterarguments

• No hard numbers: no pJ/bit, cost or timeline; the only "10x lower FIT" is spoken, and no baseline was given.

• Customization erodes what it wants to keep: if every data center co-designs its own, multi-vendor interoperability and scale economics get fragmented.

• The operating model is still blank: if replacing an engine means opening the box, field work is nothing like faceplate hot-swap, and who does the repair is unanswered.

• More options, heavier qualification: mixing FRO, TRO/LRO and linear per port enlarges the qualification matrix, exactly one of the DSP costs on the slide.

• NPO is positioned as a transition: Yin calls NPO a bridge. At ECOC Microsoft already set an optical I/O target of >10 Tbps/mm (see ECOC 2026: Microsoft Writes Down the Optical I/O Spec Sheet), so the payback window for NPO investment may be shorter than module makers expect.

8. What It Means for Taiwan's Supply Chain

The talk named no Taiwanese companies; the points below are inferred only from the roles and technologies listed on the slides:

• High-density electrical connectors and sockets are new part numbers in NPO. 224G / 12.8T-class electrical connectors plus flyover cables are new content beyond the faceplate cage. Taiwanese connector and cable makers are positioned to enter, but this part must be co-designed with the ASIC, optical engine and mechanics.

• "Customized per data center" favors system integrators. With the shell not being unified soon, switch ODM design capability becomes worth more than plain module manufacturing.

• 10x lower FIT is a testing business. Reliability qualification, burn-in, telemetry and failure isolation are all on the interoperability list, and test equipment is one of the named participants.

• Light-source options stay open. External or integrated lasers are both allowed and SiPh / EML / VCSEL are chosen by application, so Taiwanese laser and VCSEL makers will not be locked out by a single architecture, but they need FIT data ready.

Verdict: NPO orders will not go first to the cheapest module maker, but to the one that can deliver FIT data, CMIS-style management and a field-replacement process.

9. Conclusion

The value of this talk is that Microsoft spelled out NPO's acceptance criteria: preserve the ecosystem model, improve reliability 10x, keep the architecture flexible and standardize interfaces first. It gave no power numbers, because for a buyer power has never been the only consideration.

Three things to watch over the next 12 months: whether IPEC or OCP publishes a CMIS-like NPO management spec, whether any supplier discloses FIT and field-replacement data for NPO optical engines, and whether Microsoft ends up backing a socket spec or going custom.

The success of pluggables was never the module itself but the rules that let anyone plug in; NPO only shortens the electrical path, and not one of those rules can be dropped.

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