OCP Global Summit 2026 | LightCounting | Vlad Kozlov | Market Forecast for CPO and NPO

11 hours ago

8 min read

1. Why This Talk Deserves Its Own Breakdown

LightCounting is one of the most widely cited market research firms in optical communications. Founder and CEO Vlad Kozlov spoke for just 11 minutes with 9 slides at the IPEC NPO workshop during OCP Global Summit, but the information density was high.

He opened with a joke: "We no longer just publish forecasts; we show how forecasts change every three months." Over the past two years LightCounting has raised its forecast almost every quarter, with a single brief cut after the US tariff announcement, which turned out to be benign.

In July, in CPO Is No Longer "Crying Wolf", we covered the CPO/NPO online conference LightCounting hosted. The theme then was "deployment beats innovation" and that NPO and CPO will coexist; it was about direction. This time is different: LightCounting turned direction into numbers, and laid out three possible endings at once.

When a forecaster admits its own uncertainty is widening, that is itself the signal.

2. Breaking Down the Revisions: 2027 Raised 87% in Six Months

The first data slide is blunt: capex has accelerated sharply over the last 6 to 9 months, and "we have never increased our forecast as fast in the past. Is this sustainable?"

Values read from the slide curves (approximate):

Year	March 2026 version	July 2026 version	September 2026 version
2026	~$29B	~$35B	~$44B
2027	~$34.5B	~$46.5B	~$64.5B
2028	~$40.5B	~$55B	~$77B
2031	~$59B	~$80.5B	~$94.5B

▲ LightCounting's Ethernet optical module market forecast in its March, July and September versions: 2027 revised up from about $34.5 billion to about $64.5 billion (Image source: LightCounting | 2026 OCP Global Summit presentation (3:10))

On the September version, 2025 is about $20B and 2026 about $44B, up 120% year over year. The curve implies roughly +45% to +50% for 2027 (Kozlov said verbally that growth would go "from 120% to 80%, 70%", which does not match the curve; this article follows the slide). After that, 2028 is about +19%, 2029–2030 is nearly flat, and growth only returns to double digits in 2031.

Kozlov gave two concrete reasons for the flat stretch:

Scale-up optics arrive in 2029–2030, leaving no new high-volume application to take over before then.
Fewer transceivers per GPU. Multi-plane architectures and higher-radix switches start improving optics efficiency from 2027–2028. He estimates transceivers' share of capex falls from about 5% today to 4%, or even 3.5%, by 2029–2030.

In other words, capex keeps growing, but the share flowing into transceivers is shrinking. This matches recent market sentiment; we saw the same tug-of-war in Goldman Raises, Eoptolink Delays.

3. Three Scenarios: a 4x Gap by 2031

The second slide is the heart of the talk. LightCounting no longer publishes a single scenario. The detailed forecast uses "Soft Landing" as the base case, with two other lines alongside:

Scenario	2027	2029	2031	Curve shape
Soft Landing	~$64.5B	~$79B	~$94.5B	Flat in 2029–2030, rebounds in 2031
Bumpy Ride	~$56B	~$36.5B	~$52B	Peaks in 2027, trough in 2029 about 35% lower
Fantastic Growth	~$78B	~$143B	~$213B	Roughly linear, growth rate slows each year

▲ LightCounting's three scenarios range from about $52 billion to about $213 billion in 2031, a spread of roughly 4x; the bumpy scenario peaks in 2027 and bottoms out in 2029 (Image source: Simple Tech Trend; data from LightCounting's 2026 OCP Global Summit presentation)

Three points stand out:

All three lines assume growth rates moderate. Even Fantastic Growth is linear, not exponential. When LightCounting said last month "we are probably at the peak of this cycle", it meant the peak in growth rate.
Bumpy Ride is the one that fits history. Kozlov said the firm's mathematician-trained analyst argues every year that only this red line should be published, "but we cannot just publish a forecast for a declining market".
He would rather the correction come early. "Rather than growing like crazy for another three years and then crashing, I would prefer a moderate recalibration next year."

The message is clear: LightCounting's Soft Landing base case is a compromise, not a statement of confidence.

4. CPO/NPO Shipments: 2029 Is the Inflection Point

Next comes the separate CPO and NPO forecast. It was published in May and only raised slightly in July and September, because it looks further out and 2026–2027 changes do not move it much. Units are ports, and the slide notes one optical engine includes 4 to 8 ports.

▲ Combined NPO and CPO ports grow from about 4.5 million in 2027 to about 190 million in 2031, with scale-up accounting for about 58% in 2031 (Image source: Simple Tech Trend; data from LightCounting's 2026 OCP Global Summit presentation)

Approximate values read from the bar charts:

Year	NPO ports	CPO ports	Total
2027	~2M	~2.5M	~4.5M
2028	~7.5M	~9M	~16.5M
2029	~22M	~20M	~42M
2030	~44M	~47M	~91M
2031	~97.5M	~91.5M	~189M

We cross-checked this against the speed breakdown on the sixth slide: in 2031, CPO/NPO ports across 800G, 1.6T and 3.2T total about 188M, consistent with about 189M here, so both slides come from the same model.

How to read it:

2026–2027 is a validation phase. The two years combined are under 6M ports, roughly 1M optical engines (STT estimate at 4–8 ports per engine).
Scale-out first, scale-up later. Scale-up is about one third in 2029 and rises to about 58% in 2031, with a higher scale-up share on the CPO side.
NPO and CPO split roughly 50/50, with NPO slightly ahead in 2029 and 2031.

5. Customers Have Picked Sides: NPO for Hyperscalers, CPO for System Buyers

Kozlov used a Venn diagram of adopters and stressed it is LightCounting's analysis, not endorsed by these companies:

NPO camp: Amazon, Alibaba, Google, Meta, Microsoft, Tencent
Both: Bytedance, Meta, Microsoft
CPO camp: NeoClouds, Oracle, enterprise customers

▲ LightCounting splits the NPO and CPO camps by major customers' current plans: hyperscalers lean toward NPO, while NeoClouds, Oracle and enterprise customers lean toward CPO (Image source: LightCounting | 2026 OCP Global Summit presentation (7:02))

He described NPO as "a democratic version of co-packaged optics". The logic: hyperscalers that design their own systems want to keep supplier choice and serviceability, while smaller customers without in-house teams are happy to buy fully integrated CPO systems from Broadcom or NVIDIA. This points the same way as Huawei's ECOC scorecard; see ECOC 2026 | Huawei Scores CPO vs NPO, NPO Leads 4:2.

Notably, the NPO list includes three Chinese cloud providers (Alibaba, Tencent, Bytedance), while the CPO side has none.

6. By Speed: 1.6T Is the Main Battlefield, 3.2T Is Where CPO/NPO Dominate

The sixth slide splits 2031 port shipments by speed and type:

2031 port type	Shipments	Retimed modules	LPO/LRO	CPO/NPO
800G (100G/lane)	99M	89%	3%	8%
1.6T (200G/lane)	194M	43%	9%	48%
3.2T (400G/lane)	116M	23%	2%	75%

▲ LightCounting expects 75% of 3.2T ports and 48% of 1.6T ports to be CPO/NPO in 2031, while 89% of 800G ports remain retimed pluggable modules (Image source: LightCounting | 2026 OCP Global Summit presentation (8:34))

The table supports three conclusions:

800G does not disappear. About 99M ports still ship in 2031, nearly 90% of them retimed pluggables. LPO and CPO/NPO remain trials in this generation.
1.6T is the largest generation by ports, with CPO/NPO near half; this is where old and new architectures truly compete.
3.2T only starts in 2029–2030 but ramps fast, with CPO/NPO at three quarters. In Kozlov's words, the higher the per-lane rate, the clearer the advantage of putting optics close to the ASIC.

Overall, the three speeds total about 409M ports in 2031, with CPO/NPO at about 46% and retimed pluggables at about 48% (STT weighted estimate from slide percentages).

7. The Lesson from the 2018 Forecast

In his final two minutes, Kozlov looked back. LightCounting built its first CPO forecast in 2018 for ARPA-E at the US Department of Energy. They joked that a five-year forecast was easy because the answer was zero, so they built a ten-year one.

In hindsight, the timing was roughly right, with first CPO deployments expected in 2026–2027, but units were about 10x too low. In 2018 only front-end networks were visible, and they now take only about 15% of transceivers; back-end scale-out networks use 4x more, and scale-up was an application nobody anticipated.

So he announced two things: next month's CPO/NPO forecast will add scale-in and extended-memory applications, and in December LightCounting will publish a ten-year forecast. His closing line: "Don't worry about the downturn, plan ahead, because new applications will make this forecast look ridiculously low five years from now."

8. Risks and Counterarguments

Numbers are chart readings: there is no official PDF for this session, so table values were read from video-captured curves and bars, with errors possibly in the hundreds of millions of dollars or hundreds of thousands of ports.
No CPO/NPO revenue forecast: the slides give ports only, with no dollar value or per-port price assumption, so they cannot be converted directly into laser or module revenue.
Verbal and slide figures disagree: the 2027 growth rate was stated as 70–80% but the curve implies about 45–50%, which may be a transcription error or a different basis.
No scenario probabilities: LightCounting says Bumpy Ride fits history best yet uses Soft Landing as the base case, so users must assign their own weights.
The 2018 lesson cuts both ways: new applications made that forecast 10x too low, but the same logic means today's forecast cannot be sure when new applications will arrive either.

9. What It Means for Taiwan's Supply Chain

The pluggable business stays large through 2031. 89% of 800G and 43% of 1.6T remain retimed modules, more than 170M ports combined. For Taiwan's module, component and contract manufacturers, CPO/NPO is a generational switch, not an overnight replacement.
Capacity planning should account for the Bumpy Ride timing gap. If a correction lands in 2028–2029, capacity added now on 120% growth would face utilization challenges as pluggable demand flattens, while CPO/NPO only reaches about 42M ports in 2029.
Laser suppliers' opportunity follows 3.2T. CPO/NPO take 75% of 3.2T, so light source specs and packaging for this generation are what matter. We discussed whether NPO integrates the laser and how much power it needs in CPO Lasers Don't Have to Chase 400 mW: NPO Integrated Lasers; this forecast puts the ramp at 2029–2030.
Customer camps determine the channel. The NPO camp is hyperscalers designing their own systems, so suppliers must win designs directly; the CPO camp buys systems from Broadcom and NVIDIA, so suppliers must first enter those two supply chains. Qualification and pricing work very differently on each path.

10. Conclusion

The most valuable part of this talk is not the raised blue line but that LightCounting admitted two things at once: the market grows 120% in 2026, and that pace does not fit industry history. Soft Landing is the base case, Bumpy Ride is history, and Fantastic Growth is what happens if new applications arrive faster.

CPO/NPO play the same role across all three scenarios: 2026–2027 is validation, volume starts in 2029, about 189M ports by 2031, and three quarters of the 3.2T generation. It is not next quarter's revenue story but the protagonist of the next cycle. Over the next 12 months, watch three things: how much LightCounting raises CPO/NPO after adding scale-in next month, the scenario design of December's ten-year forecast, and whether 2027 shipments track Soft Landing or Bumpy Ride.

A forecaster that raises its numbers four times a year while warning they might crash is telling you the useful part is not the number, but the timing gap between scenarios.

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