OCP Global Summit 2026 | Meta | Nan Ge | Beyond Pluggables: Building the NPO Ecosystem

1. Why listen to a sourcing manager talk about NPO

What makes this talk unusual is the speaker. Most NPO sessions before it were given by engineers covering optical engines, packaging and measurement; Nan Ge opened by saying he had "decided not to talk about technology at all." He sits on Meta's technical sourcing team, and he came to cover three things: industry trends, how to build the ecosystem, and why Meta believes NPO is a good answer for next-generation AI interconnect.

That angle turns out to be the most valuable one. Which technology path wins is often decided not by pJ/bit, but by how much risk the buyer is willing to carry. When a hyperscaler sourcing manager says publicly "where we deploy first, what backup we keep, and what we need from suppliers," the buying logic is on the table.

Engineers tell you whether NPO can be built; sourcing managers tell you when NPO will be bought.

2. Demand breakdown: from 128 GPUs to 1 GW, and 100 million transceivers

Meta started with its scale: 3.6 billion daily active users and 200 billion-plus messages flowing across its platforms every day. AI is reshaping every product: Meta launched the personal AI agent Muse in September and followed with Muse Wonder and Muse Wonder Full in October. Nan Ge stressed that model iteration cycles have shrunk from one or two years to months.

The infrastructure build-out has moved even faster:

• 2022–2023: from 128 GPUs with no back-end network, to 2K, 4K, 8K, then a 24K building-sized cluster

• 2024: a 129K-GPU multi-building cluster

• 2025 onward: 1 GW-class multi-region clusters, still in progress

▲ Meta's cluster scale grew from 128 GPUs in 2022 to 129K GPUs in 2024, and to 1 GW cross-region clusters from 2025 onward (Image source: Meta | 2026 OCP Global Summit presentation (05:26))

In transceiver terms: Nan Ge's rule of thumb is about 6 transceivers per GPU (higher or lower depending on topology), and on Meta's forecast for next year, total data center demand will exceed 100 million units. Another slide shows AI network scale growing 10x from 2024 to 2030 and names two scaling challenges: the slowest node gates overall performance, and latency overhead turns into stranded compute.

Put those two numbers together and the message is clear: the transceiver bottleneck has shifted from "can it be built" to "every watt and every nanosecond gets multiplied 100 million times".

3. Power breakdown: count the SerDes and you know who saves power

This is the slide to take home. Citing a paper by Drew Alduino et al. at ECOC 2026, Meta offers a minimalist method: count the SerDes along the entire link (including the ASIC side), and you can rank the relative power of each option.

Option	Relative SerDes count	Link power on chart (with range)
Retimed pluggable (5nm SerDes)	2x	16
Retimed pluggable (3nm SerDes)	2x	12.5 (+1)
LRO, half retimed	1.5x	9 (+2)
LPO / NPO / CPO	1x	6 (+2)
Integrated optics / OCI	0x	3.5 (+1.5)

▲ Link power ladder ranked by SerDes count: retimed pluggable 16, LPO/NPO/CPO 6, OCI 3.5; each set of SerDes removed drops one step (Image source: Simple Tech Trend; data from Meta's 2026 OCP Global Summit presentation)

The slide does not label units, so the chart values are taken as given. Speaking about an 800G module, Nan Ge said typical power is about 16 W and NPO or CPO can bring it down to "almost 10 W." That does not fully match the chart's 6 (+2); the spoken figure may be a conservative number near the top of the range.

The real point is not the decimals but Meta's verdict: "Would that be enough? We don't think so." He said plainly that SerDes is the key power driver behind the scenes, not the supporting architecture; remove the reliance on SerDes and power drops to a level Meta can accept. In other words, for Meta NPO is not the destination, but the fastest way to remove the DSP without changing the ASIC interface.

That also explains why LPO, NPO and CPO sit on the same row: from a SerDes perspective, all three are in the same power class. The differences are serviceability, interoperability and who owns the whole link, which is exactly what sourcing cares about.

▲ Meta's slide "Beyond Pluggables: The Case for NPO", tracking the relative power of each option by SerDes count (Image source: Meta | 2026 OCP Global Summit presentation (12:26))

4. Why scale-out first: you only take risk when you have a backup

Nan Ge noted that Meta's use of NPO is "slightly different from our partner hyperscalers": Meta's home ground for NPO is scale-out, not scale-up.

The reasoning is pure risk management:

• Scale-out has a fallback. This tier already has mature alternatives, including pluggable optics; if NPO's time to market slips, Meta has a backup position to mitigate the risk.

• Scale-up is not ready to switch. NPO, CPO and OCI are all under exploration but are "not mature enough to completely replace the copper backplane," and Meta has not made the call to move the copper backplane to NPO.

So Meta's conclusion is: roll out NPO in the scale-out network as a pilot platform. That order is the opposite of the common narrative that "NPO is a transition step for scale-up." We argued in NPO Is Not a Stepping Stone to CPO that NPO has its own position; Meta's sourcing decision now shows NPO's first volume use case may appear where it is least needed but easiest to back out of.

▲ Meta's adoption logic: scale-out has pluggables as a fallback, so NPO starts as a pilot; scale-up stays on copper backplanes; the long-term goal is OCI removing the SerDes from the optical link (Image source: Simple Tech Trend; data from Meta's 2026 OCP Global Summit presentation)

Suppliers often ask him: "What do you actually want, pluggable, half-retimed, NPO, or CPO/OCI?" His standard answer: it is not an either-or choice. Each technology has its own use case, and they are not competing with each other.

5. The ecosystem challenge: NPO without a standard needs both sides to co-design

This is the kind of candor only a sourcing manager offers. Nan Ge said most NPO talks never show "the co-design work between the switch vendor and the module partners," and that this is exactly the hardest part of moving to the next architecture:

• Optics vendors care only about module performance, but once a module is designed into a switch system, the switch partner comes back with integration, signal integrity and link reliability issues.

• The pluggable world is manageable, because mature standards define the boundary conditions and ensure interoperability.

• NPO does not have that standard yet, so the market and customers demand close collaboration between both sides to optimize the whole link as a system.

The concrete technical requirement fits in one sentence, but it matters: linear drive must deliver enough reach from the ASIC to the front port and meet IEEE standards. Power optimization is not only the optics vendor's job; it is system-level work.

That is why the industry has started trying to standardize NPO interfaces. We broke down the Open CPX approach in ECOC 2026: Ciena's Open CPX Turns the 6.4T Optical Engine Back into a Shelf Part; Meta's remarks amount to a buyer-side endorsement of such specs: no boundary conditions means no multi-sourcing, and no multi-sourcing means no negotiating leverage.

6. The north star: take SerDes out of the optical link

Nan Ge laid out Meta's long-term direction clearly:

• Today: LPO, NPO and CPO can all remove the SerDes on the optics-side DSP, but the SerDes on the ASIC side still has to stay.

• Next: through OCI (Optical Compute Interconnect), built with industry partners, the north star is to remove SerDes from the optical link entirely and use optics directly for high-speed interconnect.

That matches the direction we saw in Meta, Broadcom and AMD Define the OCI 200G Line-Side Spec. Read together, Meta's technology path is a straight line: retimed pluggable → LPO/NPO (module DSP removed) → OCI (all SerDes removed). NPO's role on that line is the stop where scale-out runs first and builds supply chain and reliability experience.

7. Risks and counterarguments

• The power chart has no units, and the spoken numbers do not match. The chart shows 6 (+2) for LPO/NPO/CPO while the talk said 800G drops to almost 10 W; until the ECOC paper is in hand, the ranking is credible but absolute numbers deserve caution.

• "Meta itself says it is not enough". NPO only cuts SerDes from 2x to 1x, still short of Meta's goal; if OCI moves faster than expected, NPO's window in scale-out may be shorter than the market assumes.

• A pilot is not volume production. Meta said explicitly that NPO is a pilot platform in scale-out, and the talk gave no timeline or volume for broader deployment.

• The missing standard slows multi-sourcing. Until the interface spec takes shape, NPO is likely a tied project between a switch vendor and a few optical engine makers, with limited leverage on price and lead time.

• 100 million units is a spoken forecast. It comes from the 6-per-GPU rule of thumb, does not appear on the slides, and covers all transceivers, not NPO volume.

8. What it means for Taiwan's supply chain

• The talk did not name any Taiwanese company; the points below only infer positions from Meta's stated needs.

• Switch system makers gain weight. Meta wants switch vendors and optics partners to jointly own signal integrity, link reliability and linear-drive reach; until NPO is standardized, system makers that can handle the ASIC, board design and optical engine integration together will have more leverage than pure assemblers.

• Connectors and sockets are the key to standardization. Whether NPO can be multi-sourced depends on standardizing the mechanical and electrical interface between optical engine and main board. The Open CPX ecosystem list shown in other sessions at the same OCP summit includes Taiwanese suppliers such as Accton, BizLink, LOTES and FIT (Foxconn Interconnect Technology); what to track is when those names move from logos to pilot orders from hyperscalers like Meta.

• Lasers and optical engines: Meta did not discuss light source architecture. Removing the DSP only raises the bar on laser and optical engine linearity, power and reliability; see Is Nvidia's Shift from CPO to NPO Bad News for the Laser Supply Chain?.

9. Conclusion

This talk did not introduce a single new technical number, yet it is the one to read in the IPEC NPO series, because it answers what technical talks cannot: who buys NPO first, where, and why.

Meta's answer is pragmatic: power is a SerDes count, deployment depends on having a backup, and scale depends on having a standard. So NPO goes first into scale-out, where pluggables are the fallback; scale-up keeps the copper backplane; and the long-term goal is OCI removing all SerDes. For suppliers, Meta's ask is equally clear: do not just deliver an optical engine with great specs; deliver, together with the switch vendor, a link that is reliable, interoperable and IEEE-compliant.

Over the next 12 months, watch three things: when Meta announces a broader NPO scale-out deployment timeline, which organization converges the NPO interface spec, and which year OCI reaches its first volume node.

NPO will be won not by lab pJ/bit, but by whether the sourcing manager dares to remove the backup.

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