OCP Global Summit 2026 | NVIDIA | Daniel Kuchta | Optics for Spectrum-X Multiplane Networks

1. The cleanup hitter: why NVIDIA pushed back at an NPO workshop

This was the final talk of the IPEC (NPO standardization) workshop held during the OCP Global Summit. Kuchta opened with a baseball analogy: the last batter is the cleanup hitter, and NPO has a few things that "need to be cleaned up."

He set the stage first: agentic AI is the most complex computing workload in history, the data center is the computer, the network defines the data center, and optics determine performance and scale. NVIDIA's flagship Vera Rubin platform combines the Vera CPU, NVL72 compute and NVLink switch trays, BlueField-4 storage, Spectrum-6 SPX switches, and the LPX tray from the Groq team. One Vera Rubin POD is 40 racks, and Kuchta said 10 of them are Spectrum-X switch racks. The slide also splits Spectrum-6 SPX into east-west (E-W) and north-south (N-S) networks, which is the basic outline of the "multiplane" architecture in the title, although the talk did not break down optics counts per plane.

The weightiest slide compared two 1 GW AI factories:

1 GW AI factory	x86 + Hopper	Vera Rubin
Number of GPUs	600K	300K
Tokens per second	2M	700M

Half the GPUs, 350x the token output; per GPU, that is roughly 700x (STT calculation from the slide numbers). Under the same power cap, every extra watt spent on networking and optics erodes that multiple, and that is NVIDIA's starting point for choosing optics.

▲ For the same 1 GW, Vera Rubin produces 350x the tokens with half the GPUs (Image source: Simple Tech Trend; data from NVIDIA's 2026 OCP Global Summit presentation)

2. CPO breakdown: Spectrum-X Photonics is in production, and soldered down

NVIDIA's first card is Spectrum-X Ethernet Photonics, a CPO switch chip built for AI factories. The slide lists four technical points:

• Micro ring modulator CPO chip, in production

• 3D-stacked silicon photonics engine on the TSMC COUPE process

• High-power, high-efficiency external lasers

• Detachable fiber connectors

Versus pluggables, NVIDIA gives three numbers: 1/4 the discrete components, 5x lower power, and 10x higher MTBI (mean time between interruptions).

▲ Spectrum-X Photonics CPO: micro-ring modulators, TSMC COUPE, 1/4 the discrete components, 5x lower power and 10x higher MTBI (Image source: NVIDIA | 2026 OCP Global Summit IPEC workshop presentation (3:06))

These three numbers should be read separately. 5x power and 1/4 components are structural results of CPO removing the DSP and module housing, so they are more credible; 10x MTBI depends on keeping the laser external and separately replaceable, which is why NVIDIA insists on external lasers. We discussed the laser's role under CPO and NPO in Is Nvidia's shift from CPO to NPO bad news for the laser supply chain?, and this talk confirms at least one thing: NVIDIA's switch CPO roadmap has not changed course, and high-power external lasers remain a core component.

3. Pluggable breakdown: OSFP is still the workhorse, and its successor is USFP

CPO is not NVIDIA's only option. Kuchta stated plainly that OSFP is still the industry's "work horse": he said the Vera Rubin NVL72 compute tray carries 8 OSFP modules, and Spectrum-6 is also offered in a fully pluggable version, the SN6600-LD. The slide title adds that OSFP is gaining TRO (transmit retimed optics) and DR4+.

For OSFP's successor, the slide lists four requirements: 400G-capable signal integrity, x8/x16 lane options, simple top/bottom cold plates, and reliable, cost-effective design. The answer is USFP (Ultra-dense SFP), with the form factor co-defined and the MSA co-driven by Cisco and NVIDIA:

Item	USFP spec
SerDes per lane	200G and 400G
First generation	16 lanes, 6.4T
Second generation	32 lanes, 12.8T
Bandwidth density	4x
Power efficiency	1.3x (400G vs 200G SerDes)
Ecosystem partners	Amphenol, TE

▲ USFP, co-defined by Cisco and NVIDIA: 200G/400G per lane, 6.4T and 12.8T, and 4x the bandwidth density (Image source: NVIDIA | 2026 OCP Global Summit IPEC workshop presentation (4:58))

One gap deserves attention: the slide says 1.3x power efficiency (roughly 23% less power per bit), while Kuchta said "at least 30%" lower power. The baselines differ: the former compares 400G with 200G SerDes, the latter gave no reference point. Cite the slide figure.

The more important signal is timing. USFP is "a pluggable designed for the 400G generation," meaning NVIDIA judges that pluggables still have a place in the 400G-per-lane era and will not fully give way to CPO or NPO.

4. Socketed vs soldered: a socket does not mean field replaceable

Kuchta spent the most time on a common fear: CPO has no socket, so one failure scraps the whole board. His answer was "Been There, Done That." NVIDIA's Quantum InfiniBand CPO switch used sockets with pigtails; today's Spectrum-X switch has 32 CPO engines, all soldered down.

Approach	Pros	Cons
Socketed CPO	Good for initial deployments, OSAT rework, required for pigtail solutions	Adds cost and area, impacts signal integrity and thermals
Soldered CPO	Best signal integrity, smallest footprint, lowest cost	Hardest to implement, limited rework

His conclusion was blunt: sockets are useful for prototypes and factory rework and not much more, "socketed does not mean field replaceable," and anyone who thinks otherwise "doesn't know how these boxes are put together." The slide adds one line: the same goes for NPO. That line is the turning point of the talk, because field serviceability is exactly the advantage the NPO camp most often claims over CPO.

5. NPO's two selling points, and NVIDIA buys neither

Kuchta joked that NPO might stand for "No Pluggable Option," but his position is not a blanket rejection:

• More customers are exploring NPO, and NVIDIA will work with ODMs to enable it and will also support it through NVLink Fusion

• CPO, pluggables and NPO will likely co-exist for scale-out

He accepts two reasons for NPO: no faceplate space and a need for more density that forces optics inside the box; or flyover cables and connectors that are too lossy, pushing a linear or TRO solution, about which he said "I wish you luck, it's harder than it looks."

He rejects the other two:

• Multiple suppliers: every NPO prototype today looks different, and there are four MSAs, all brand new. He even said the only way LightCounting's projections come true is if many people ship single-sourced.

• Serviceability: NPO is assembled in the factory, likely in a clean room, so it will not be taken apart in the field but packed up and shipped back; woven pigtails are very hard to repair, and with fiber shuffles it is almost impossible.

▲ Three packaging distances: in-package at about 30 µm pitch, co-packaged at about 150 µm and near-packaged at about 350 µm; I/O energy follows A < B << C and scales nonlinearly (Image source: NVIDIA | 2026 OCP Global Summit IPEC workshop presentation (9:26))

The diagram on this slide is more damaging than the text: in-package (~30 µm pitch), co-package (~150 µm) and near-package (~350 µm) layers, with I/O energy of A < B << C, and not linear. In other words, each step that pushes the optical engine further out makes the electrical power penalty jump, rather than grow proportionally.

▲ NVIDIA's optics menu: soldered CPO on switches, OSFP moving to USFP on servers, and NPO supported through ODMs and NVLink Fusion (Image source: Simple Tech Trend; data from NVIDIA's 2026 OCP Global Summit presentation)

This points in exactly the opposite direction from what we saw in ECOC 2026 | Huawei scored CPO vs NPO, and NPO leads 4:2. Huawei's scorecard counted serviceability and ecosystem openness as NPO advantages; NVIDIA says neither holds up in practice.

6. Risks and counterarguments

• NVIDIA has a stake: NVIDIA already ships its own CPO, so questioning NPO multi-sourcing also protects its vertically integrated roadmap. The talk gave no direct NPO-vs-CPO power comparison.

• MSAs are converging: Kuchta says there are four different MSAs, but specs like Ciena's Open CPX, which turns the 6.4T optical engine back into an off-the-shelf part are tackling exactly the multi-source problem, and the picture may differ in a year.

• CPO numbers are self-reported: the talk did not state the baseline or measurement conditions for 5x power and 10x MTBI.

• USFP is still at the MSA stage: there is no timeline for first-generation 6.4T volume or module power.

• LightCounting's forecast is not necessarily wrong: we covered its NPO forecast in CPO is finally no longer "crying wolf"; single-sourced shipments may be an acceptable cost for hyperscalers.

7. What it means for Taiwan's supply chain

• TSMC COUPE is the only Taiwanese company named. The 3D-stacked silicon photonics engine in Spectrum-X Photonics uses COUPE and is in production, and with NVIDIA's switch CPO roadmap unchanged, COUPE's shipment base continues.

• USFP's connector partners are Amphenol and TE, with no Taiwanese connector maker on the list. A new form factor with 4x density will drive redesigns of module housings, thermals and test fixtures; whether Taiwanese vendors get in depends on second sources after the MSA is published.

• ODMs, not NVIDIA, drive NPO. NVIDIA said it will work with ODMs to support NPO but named none and gave no timeline. For Taiwanese assemblers and module makers, NPO orders will come more directly from hyperscaler and ODM specs than from NVIDIA reference designs.

• Demand for high-power external lasers continues. NVIDIA's production CPO explicitly uses high-power, high-efficiency external lasers; the larger switch CPO scales, the greater the demand for external laser sources.

8. Conclusion

The value of this talk is that NVIDIA laid out its optics choices very clearly: soldered CPO for switches, OSFP on the server side, USFP as the next pluggable, and NPO left to customers and ODMs. It does not oppose NPO's existence; it opposes the pitch that NPO is "multi-sourced and easy to service."

On timing, watch three things: the first USFP MSA spec and 6.4T module volume timing, whether the four NPO MSAs converge before 2027, and field MTBI data from Spectrum-X CPO switches. If NPO specs are still fragmented in 2027, Kuchta's "single-source" prediction will come true.

NVIDIA is not suppressing NPO; it is reminding everyone that sockets and multi-sourcing are not free, and the real bottleneck is how much energy you are willing to pay for.

For technology and industry trend analysis only; not investment advice.