OCP Global Summit 2026 | Huawei | Eric Bernier | NPO Standardization: IPEC and the Global Landscape

1. Who IPEC Is, and Why It Ran a Full NPO Workshop at OCP

IPEC (International Photonics & Electronics Committee) hosted an entire NPO workshop at the OCP Global Summit, with about 14 talks almost all centered on NPO. Bernier's talk set the tone. It ran only ten minutes, but it made IPEC's identity clear:

• Open: any interested party can join, and the speaker described the membership fee as "very low"

• Incorporated in Geneva, pitching globally unified standards, with members participating in spec development on a non-discriminatory basis

• About 60 members, plus a separate track for cutting-edge research

Huawei's own position is equally clear. Bernier said Huawei's superclusters have used all-optical interconnect for about 10 years, in both VCSEL and silicon photonics versions; the slide splits mainstream super-node designs into "intra-rack copper + inter-rack optical" and "all-optical interconnect." We covered Huawei's density argument in ECOC2026 | A 1,024-lane switch is capped at 0.07 T/mm. This time the point is not the technology but who writes the spec.

2. The Premise: How Much Is Left for Copper After 448G

Bernier starts from a reach-versus-rate chart. From 25G NRZ to 200G PAM4, usable DAC reach keeps shrinking and the point where optics takes over keeps moving inward; at 448G, the chart simply shows a question mark.

▲ As rates move from 25G NRZ to 200G PAM4, usable reach for DAC and ACC/AEC keeps shrinking, with a question mark over 448G (Image source: Huawei/IPEC | 2026 OCP Global Summit presentation (3:38))

His view on "slow and wide" is worth noting: he grants that more parallel approaches such as OCI have "a window of opportunity," but says history has been "very brutal" to systems that go slower and wider, and the industry will still have to answer how to do 448G. In other words, IPEC is betting that fast and narrow keeps going, and NPO is the engineering answer for bringing 448G optics close to the ASIC.

3. The Scorecard: From 4:2 to 4:1

The slide compares NPO and CPO for a 224G super-node as a boxing match:

Round	Criterion	Winner
1	Technical maturity and supply chain resilience	NPO
2	Serviceability and cluster availability	NPO
3	Manufacturing yield and commercial cost efficiency	NPO
4	Standardization and ecosystem compatibility	NPO
5	Ultimate energy efficiency and extreme integration	CPO

▲ Huawei compares NPO and CPO for a 224G SuperNode over five rounds; NPO wins 4:1, with CPO taking only energy efficiency and integration (Image source: Huawei/IPEC | 2026 OCP Global Summit presentation (5:50))

We wrote about Huawei's other scorecard in ECOC2026 | Huawei scored CPO vs NPO, and NPO leads 4:2. This version is 4:1 and the conclusion is the same, but Bernier's spoken aside is more interesting: he concedes that CPO is the "ultimate" in efficiency and integration, it just is not "trained enough" yet. That is not a rejection of CPO; it is a timing call: the window right now belongs to NPO.

4. The Spec Map: Nine Items, Only Half of Them New

This is the most important slide of the talk. Bernier draws the items NPO needs to standardize around an optical transceiver socketed on the PCB:

▲ The nine NPO standardization items listed by IPEC: mechanical dimensions, optical connector, system requirements, optical interface, ILS/ELSFP, electrical connector, CMIS, and low-speed and high-speed electrical interfaces (Image source: Huawei/IPEC | 2026 OCP Global Summit presentation (8:10))

Mapping the slide and the spoken comments against other standards bodies gives roughly this picture:

Standardization item	What the speaker said	Counterpart elsewhere (STT mapping)
Optical interface	Reuses existing IEEE specs	IEEE 802.3
Management (CMIS, I2C/I3C)	Reused and extended	OIF CMIS
Laser source (ILS/ELSFP)	ELSFP-related spec already exists	OIF ELSFP
High-speed electrical (448G, PAM4/PAM6)	In progress	OIF CEI, IEEE 802.3
Electrical connector, optical connector, mechanical dimension	Focus of the 7.2T project	Open CPX MSA (6.4T/7.2T sockets)
System requirement	Splits the error budget across segments	Defined by system vendors

▲ NPO standardization layers: optical and management interfaces reuse IEEE/OIF, while sockets, connectors and mechanical dimensions are the contested layer that IPEC and Open CPX are both entering (Image source: Simple Tech Trend; data from Huawei/IPEC's 2026 OCP Global Summit presentation)

Read the table and the point is plain: IPEC is not trying to replace IEEE or OIF. The optical and management interfaces are already written, and IPEC uses them as is; the real contest is in the lower half, the sockets, connectors and form factors. That layer decides whether vendor A's optical engine can plug into vendor B's switch, and it is an area bodies like OIF touch less. For OIF's role, see OIF drew an official map for AI interconnect.

The slide gives three reasons to standardize: technical optimization (for example, splitting the error budget optimally across system segments), ecosystem expansion, and cost reduction.

5. Progress: Hardware Was Already on Display at CIOE

IPEC demonstrated at CIOE 2026 (the China International Optoelectronic Exposition in Shenzhen), which Bernier grouped into five categories: electrical connectors, optical connectors, optical modules, an NPO switch, and NPO-DCO interoperability.

• Electrical connectors: the slide shows a Yamaichi YTM-series mezzanine connector, an Amphenol NPO socket solution, and NanoLGA/MicroLGA sockets

• Systems: NPO switches from Huawei and HGTECH

• Interoperability: logos from HiSilicon Optoelectronics, Kuaishou, Meituan and Accelink appear on the interoperability demo, and the speaker said a hyperscaler took part

▲ IPEC's NPO demo at CIOE 2026: connectors, optical modules, an NPO switch, and NPO-to-DCO interoperability (Image source: Huawei/IPEC | 2026 OCP Global Summit presentation (9:02))

The close has just one hard data point: IPEC has started a 7.2T NPO project and invites others to join. In the same workshop, Ciena's talk presented the Open CPX MSA, founded in March 2026, which published Rev 1.0 in September covering 6.4T and 7.2T (32 or 36 lanes of 200G); details in ECOC2026 | Ciena uses Open CPX to turn the 6.4T optical engine back into shelf stock. Whether IPEC's 7.2T and Open CPX's 7.2T share the same mechanics and pinout was not addressed, and that is the single most important thing to watch next.

6. Risks and Counterarguments

• The cost numbers are illustrations, not data. The "Why Standardize?" slide shows hardware cost −25%, power cost −15%, total cost −20%, latency from 10 µs to 2 µs, and power from 10 W to 7 W, but the graphic is an illustration, power cost appears twice, and there is no baseline or method. These should not be quoted as measured values.

• "Globally unified" does not match the membership. Of the 56 logos, the US names are component and test vendors such as Intel, Marvell, MACOM, Semtech, Keysight and VIAVI; there is no Nvidia, Broadcom or US hyperscaler. Without platform owners, a spec struggles to become the global default.

• The relationship with OIF ELSFP is unclear. The speaker said IPEC has standardized ELSFP but did not say whether it is fully compatible with OIF's ELSFP spec; if the two evolve separately, laser module makers will need two versions.

• Geopolitics is a variable. With Huawei leading the editing and a membership weighted toward mainland China, whether US platforms adopt IPEC specs is not purely a technical question.

• The CPO window may open sooner than expected. Bernier himself calls CPO the ultimate solution; once CPO's serviceability and yield problems are solved, the payback period for NPO socket specs gets compressed.

7. What It Means for Taiwan's Supply Chain

• NPO may end up with dual-track specs, so connector makers need to sample to both sides. Among IPEC's member logos, the only Taiwanese name visible is Foxconn (Hon Hai group); the Open CPX members on Ciena's slides in the same workshop include Accton, Alpha Networks, BizLink, FIT (Foxconn Interconnect Technology) and LOTES. The Hon Hai group is the one Taiwanese name visible on both lists.

• Sockets and connectors are the main battlefield of the spec war. IPEC's 7.2T project focuses on exactly the electrical connector, optical connector and mechanical dimension, and the CIOE demo named Yamaichi and Amphenol. For Taiwanese connector makers aiming at mainland China NPO switch supply chains, the IPEC 7.2T draft is a document they must track.

• The laser source is the overlap worth watching. Both IPEC and OIF have ELSFP; if the two are compatible, an external laser module can sell into both markets as one part, and if not, Taiwanese makers of CW lasers and ELSFP will have to weigh the cost of developing two versions.

8. Conclusion

This ten-minute talk had no new technology, but it made one more important point: in NPO standardization, the contest is at the socket layer, not the optics layer. The optical interface has IEEE and management has CMIS, and everyone uses them; what decides whose optical engine plugs into whose switch is the electrical connector, optical connector, mechanical dimension and laser form factor, and both IPEC and Open CPX are moving at the same 7.2T capacity point.

On timing, IPEC's 7.2T project has just begun, while Open CPX Rev 1.0 is already published. Three things to watch over the next 12 months: when the IPEC 7.2T draft goes public, whether its mechanics are compatible with Open CPX 7.2T, and whether any US platform owner joins IPEC.

NPO won its first match against CPO; the next one is whether NPO's own spec splits into two.

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