OCP Global Summit 2026 | Genuine Optics | David Huff | The Case for NPO Today

Oct 22

7 min read

1. The tenth NPO talk, and Genuine changed the question

Huff opened with a joke: "Number ten on the list of NPO presentations. What can I possibly say that would be new?" His answer was to change the question. Genuine's executive session had already covered its full-stack silicon photonics capability; this talk answered only one thing: why NPO, and why now.

He first cooled down the power argument. Optics is about 5–10% of data-center power, "important, but not a huge number," and nowhere near the GPU's share; latency is "pretty much already solved." The slide cites an IEA projection that AI data centers will consume about 12% of US power generation by 2030, but Huff put the weight on the last bullet: industry scaling.

The slide credits the figure to "AB, 260316," which Huff described as Andy Bechtolsheim's estimate: the next generation needs about one billion 1.6T-equivalent transceivers, in the 2027–2028 time frame.

Not just whether we can do it technically, but whether we can do it at that kind of scale and volume.

That line is the thesis of the talk. CPO is the most elegant technology, but whether a billion units can be carried by an ecosystem that is not yet standardized is a different question.

2. Power breakdown: NPO concedes only 1–4 pJ/bit to CPO

Genuine put five options on one table:

Option	Power (pJ/bit)	Latency (ns)	Front-panel optics density
Fully retimed (reference)	16	>100	1x
LPO (linear optics)	6–8	<1 + length	1x
XPO (linear)	6–8	<1 + length	4x
NPO	4–6	<1 + length	>6x
CPO	<1–4	<1 + length	>6x

The density column carries a footnote: it excludes front-panel space for the external laser source (ELS) and assumes MMC connectors for on-board optics. Huff added verbally that MMC-type connectors can, in some cases, push density to 8–9x that of OSFP.

Converting the table to watts (STT calculation): one 6.4T optical engine draws about 26–38 W as NPO versus about 102 W fully retimed. Scaled to a 102.4T switch (16 x 6.4T), NPO is about 0.41–0.61 kW, retimed about 1.64 kW, and CPO below 0.41 kW.

▲ Genuine's comparison table converted to optical power for a 102.4T switch: about 1.64 kW retimed and about 0.41–0.61 kW for NPO, with an NPO-to-CPO gap of about 0.1–0.4 kW (Image source: Simple Tech Trend; data from Genuine Optics' 2026 OCP Global Summit presentation)

That is what Huff meant in his summary when he said NPO is missing "maybe 1 to 4 picojoules per bit": moving from retimed to NPO already captures most of the power savings; the remaining gap is traded for maintainability and supply chain.

3. Deployability breakdown: CPO scores 0 on RAM

Genuine's "Ready for Prime Time?" scorecard rates four technologies on scalability, RAM, technical novelty and ecosystem readiness:

Technology	Scalability	RAM	Tech novelty	Ecosystem readiness
LPO	+++ (note)	+++	+	+++
XPO	+++ (note)	+++	++	++
NPO	+++	++	++	++
CPO	++++	0	+++	+

Note: scalable within current module capacity limits, but further expansion presents significant challenges.

▲ Genuine's "Ready for Prime Time?" scores: CPO has the highest scalability but a RAM score of 0 and only one plus for ecosystem maturity, while NPO lands at two to three pluses across the board (Image source: Genuine Optics | 2026 OCP Global Summit presentation (05:26))

The qualitative notes are blunter: LPO uses existing pluggable technology with minor changes and is easily repaired; XPO needs more changes, with easy but more expensive repairs; NPO uses the current pluggable supply chain in a new form factor and is repairable with difficulty; CPO uses proprietary ASIC/optics packaging and is not usually repairable.

Huff has worked on CPO himself since 2021. His view: proprietary systems can be bought today, but "a CPO where I can mix and match vendors and suppliers, not today and not for a long time." This matches what we saw in ECOC 2026: Huawei Scored CPO vs NPO, and NPO Led 4:2: once maintenance and ecosystem enter the scorecard, NPO comes out ahead.

4. Connector breakdown: one optical engine, three sockets

6.4T NPO currently has three connector routes:

Specification	UPO (ODCC)	American/European	Custom protocol
Application	Scale-out switch	Scale-up server/switch	Not specified
Dimensions (mm)	17x28, 20.4x34.2, 26x39.15	17x36, 22x50	26.6x41.6
Connector	LGA	CPX T2	MEZZ (e.g. MME, Gemini)

Huff pointed to a geographic divide: the LGA approach seems to be winning in Asia, especially China, while CPX and Mezz connectors are doing better in Europe and America. He did not pick a side. Genuine builds all three, because "once you have the platform, the individual connectors can be changed pretty easily."

He also gave a shout-out to Open CPX, which puts pluggability on a common substrate, somewhere between NPO and CPO. For the Open CPX design logic, see ECOC 2026: Ciena Uses Open CPX to Turn the 6.4T Optical Engine Back into a Shelf Part.

5. The 6.4T NPO hardware: specs you can see today

The 6.4T NPO Genuine showed at its booth:

Connector: Molex Mirror Mezz Enhanced (MME), active differential-pair column pitch of 3 mm or less
Fibers: 32 RX + 32 TX + 8 laser fibers (ELS), 4 x MPO16
Chipset: 2 PICs, 8 TIAs and 4 drivers, in an advanced package
Errors: the slide says loopback error better than 1E-8; Huff said verbally 32 channels x 200G, error rate below 1E-10 and 18 dB bump-to-bump loss
Timing: verbally targeted for GA in early 2027

▲ Genuine 6.4T NPO: MME connector, 32 RX + 32 TX + 8 ELS laser fibers, 2 PICs + 8 TIAs + 4 drivers (Image source: Genuine Optics | 2026 OCP Global Summit presentation (07:42))

Eight laser fibers mean this version uses an external laser source. The slide states the trade-off plainly: ELS occupies front-panel space but offers improved RAM. We broke down the integrated-laser versus ELS trade-off in CPO Lasers Don't Have to Chase 400 mW: NPO with Integrated Lasers.

Packaging matters even more. This 6.4T NPO uses 2.5D packaging, which Genuine had to develop in-house. Huff's point: NPO is a place to validate 2D, 2.5D and 3D packaging "one step at a time," instead of putting optics into the ASIC package all at once. That echoes NPO Is Not CPO's Appetizer: NPO's value is not only as a bridge, but as a way for the supply chain to build packaging yield at a controlled pace.

6. Standards timeline breakdown: mid-2027 is the convergence point

Genuine compiled the NPO timelines of four standards bodies:

Open CPX MSA: started 2026/3/23, published 2026/9/16, currently at V1.0
ODCC (UPO): UPO 1.0 released in 2026Q3, 6.4T UPO samples in 2026Q4, 6.4T UPO volume in 2027Q2, 12.8T (448G) UPO definition in 2027Q3, currently at V0.95
IPEC 7.2T NPO: started 2026/4, first draft 2026/12, standard published 2027/6, demo and test 2027/9
OIF 12.8T NPO (covering 6.4T and 7.2T): started 2026/5, baseline 2026/11, ballot 2027/5, publication 2027/8

▲ NPO timelines at four standards bodies cluster between 2026Q3 and 2027Q3, overlapping the early-2027 GA announced for Genuine's 6.4T NPO (Image source: Simple Tech Trend; data from Genuine Optics' 2026 OCP Global Summit presentation)

Huff's logic: once the standards are in place, connectors, FAUs and even packaging systems can be standardized, and everyone competes for the same sockets. With four timelines landing around mid-2027, the window for NPO to move from projects to multi-vendor supply is roughly the next 12 months.

7. Genuine's roadmap: engineering that carries across generations

Genuine's product line runs on one reuse axis: 1.6T LPO/LRO → 6.4T NPO (MEZZ, CPX and LGA versions) → 12.8T XPO, sharing a common silicon photonics PIC (800G, 1.6T to 3.2T, 400G/lane roadmap), a common optical engine (PIC + FAU + fiber + laser), common test (BER, thermal, SI, reliability) and a common supply chain (driver, TIA, laser, connector, dual-sourced critical parts).

▲ Genuine's NPO roadmap: 1.6T LPO/LRO, 6.4T NPO in three connector versions and 12.8T XPO share the PIC, optical engine, test and supply chain (Image source: Genuine Optics | 2026 OCP Global Summit presentation (09:50))

Verbally: 1.6T LPO and LRO are both sampling today; Genuine is a founding member of the XPO coalition, with a 12.8T XPO running as an LRO implementation at the booth and an LPO version to follow. The slide's conclusion: engineering investment compounds across generations instead of being stranded in a single form factor.

8. Risks and counterarguments

Spoken and slide numbers differ: the slide says loopback error better than 1E-8, while Huff said below 1E-10; the 18 dB loss and early-2027 GA were stated only verbally.
The power table gives ranges, not measurements: 4–6 pJ/bit is Genuine's own compilation with no test conditions; STT's wattage conversion is an order-of-magnitude reference only.
Scoring CPO's RAM at 0 reflects a position: Genuine is heavily invested in NPO, the scorecard comes from the NPO camp, and field reliability data from CPO vendors will challenge it.
Standards could slip: IPEC and OIF are still at the draft and baseline stage; a delay at either pushes the multi-vendor window back.
Connector fragmentation splits capacity: with LGA, CPX and Mezz in parallel, optical-engine makers gain flexibility, but connector and test-fixture makers face duplicated investment.

9. What it means for Taiwan's supply chain

The talk did not name any Taiwanese company; the points below are inferred from the slides only.
"Reuse the existing supply chain" is the argument most favorable to Taiwan. Huff stressed that NPO uses the same people who make PC boards, do the packaging and attach FAUs, with "not a lot of invention required in the supply chain." Taiwan has mature capacity in PCBs, substrates, OSAT and optical-module manufacturing, and scaling NPO sits closer to those existing capabilities than CPO does.
The connector route decides who benefits. Europe and America lean toward CPX and Mezz, Asia toward LGA, so Taiwanese connector makers need to serve both. Open CPX ecosystem logos shown in other sessions at the same OCP summit include Taiwanese companies such as LOTES, FIT (Foxconn Interconnect), BizLink and Accton; the next thing to watch is when these names appear in volume 6.4T NPO part numbers.
The ELS route creates demand for laser modules. Genuine's 6.4T NPO uses 8 laser fibers per engine, and external laser modules and high-power CW lasers are a segment Taiwanese compound-semiconductor makers can address, though the talk did not disclose suppliers.

10. Conclusion

The talk had no headline new number. Its value is in shifting the NPO argument from "saves power" to "can be mass-produced." Genuine's reasoning is clear: optics is only 5–10% of data-center power, and NPO trails CPO by just 1–4 pJ/bit, while CPO trails NPO by two levels on RAM and one level on ecosystem readiness. Against demand of a billion units, the option that can scale on the existing supply chain matters more than the option with the lowest power.

Timing should be read separately. Open CPX V1.0 is already published, IPEC and OIF target publication between mid-2027 and Q3, ODCC targets 6.4T UPO volume in 2027Q2, and Genuine targets 6.4T NPO GA in early 2027. Three things to watch over the next 12 months: whether IPEC 7.2T and OIF 12.8T publish on schedule, who takes the first volume 6.4T NPO shipments, and whether the LGA versus CPX connector split converges.

NPO gives up 1–4 pJ/bit and gets a supply chain that can scale today.

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