OCP Global Summit 2026 | Capital Group | Subu Subrahmanyan | NPO Through an Investor's Lens

1. Why a fund manager took the stage at the IPEC workshop

The IPEC NPO Workshop ran more than a dozen talks, most of them from hyperscalers, laser makers and optical-engine vendors. Capital Group is a mutual fund manager; Subrahmanyan covers public equities with a focus on AI infrastructure. He opened with a joke: "Investors usually don't present at forums like this. By the end, you'll know why."

The value of the talk is precisely that it does not talk technology. He put it plainly: the end customer decides which innovation goes mainstream, and the most useful thing an investor can do is show up and listen to the people deploying it.

Engineers ask whether NPO can be built. Investors ask who keeps the money once it is.

2. Unpacking the five layers: NPO sits at the bottom

Subrahmanyan's first content slide breaks AI investing into five questions, each one layer deeper than the last:

Layer	Question an investor must answer	Context from the talk
1	Is capex sustainable?	Google and Meta capex expected to grow 10x from 2023 to 2027
2	How much of capex is AI?	Hyperscaler capex expected above $1 trillion in 2027
3	Compute, memory or connectivity?	Compute led, memory was short last year, connectivity is next
4	Copper or optical?	Plus protocol debates such as Ethernet vs. NVLink
5	Which optical architecture?	NPO, CPO and pluggables

Two inferences are worth keeping. First, each layer is its own debate, and an investor has to be right at every one. Second, most investments are debated at the top layer: big swings in optical stocks usually come from a shift in the macro view on capex, not from something at a single company. Adding neoclouds, sovereign AI and enterprise builds, he expects the total to exceed $2 trillion by 2030.

That explains why optical stocks often have "the right theme but a flat share price": the variables that drive the stock sit four layers above NPO.

3. Unpacking the optical leverage: connectivity outgrows capex

Subrahmanyan skipped market forecasts (leaving those to LightCounting at the same session) and made a structural argument for why optics can outgrow AI capex: accelerator chips are at their size limit, so compute can only scale by connecting more of them. Scale-out, scale-across, scale-up, and now scale-in all need more bandwidth.

▲ Capital Group uses three numbers to show optical leverage: 4 million km of fiber in a 1 GW campus, optical modules possibly doubling to about 200 million units in 2027, and 4–8x more optical links once scale-up spans racks (Image source: Capital Group | 2026 OCP Global Summit IPEC workshop presentation (5:38))

The three numbers come from different sources and should be read separately:

• 4 million km of fiber: from Meta (September 2026), for the 1 GW Prometheus campus; the 5 GW Hyperion site will be several multiples of that. Subrahmanyan said this exceeds the entire cloud data center fiber market of a few years ago.

• About 2x, about 200 million units: transceiver units more than doubled in each of the last two years and could double again in 2027, labeled as an industry estimate.

• 4 to 8x: the increase in optical links between racks when scale-up spans multiple racks, labeled as the author's own estimate.

His conclusion: whatever rate AI spend grows at, optics grows faster. He added an important caveat: this only means optics outgrows capex, it does not protect you if capex itself declines. And none of this yet counts an optical rack backplane.

4. Unpacking the three calls: direction, timing, value capture

Midway through, he borrowed Richard Sutton's "The Bitter Lesson" and said investing has its own bitter lesson: being right about the technology is not the same as being right about the investment. Underwriting an architectural transition is really three separate calls:

Call	His verdict	Historical lesson
Direction	Usually right	Architectural trends are visible early and the industry tends to call them correctly
Timing	Usually wrong	The copper-to-optics case held for years but needed AI as the use case; the stocks went sideways for years before that
Value capture	Hardest to call	In the cloud era the early winners were vertically integrated, then disaggregation won

"Early" and "wrong" look exactly the same in a share price; that is the point of the second row. The third row matters more to optical suppliers: leading at the inflection is not enough, the advantage has to survive the next architecture. A rising tide lifts the whole sector; the companies that last are the ones that still hold a position after the next architecture shift.

5. Copper exits in three places, on three clocks

This is the most timing-specific slide of the talk. Subrahmanyan splits the copper-to-optics move into three locations, each on its own schedule:

▲ Copper-to-optics happens in three places: between racks and inside the box in 2027–28, and the in-rack backplane after 2029 and only partially; the direction is set, and timing depends on how long copper holds out (Image source: Capital Group | 2026 OCP Global Summit IPEC workshop presentation (10:30))

• Between racks: scale-out is already optical; what changes is density as scale-up spans racks, with links rising 4 to 8x. He sees this as likely the first major volume application, starting late 2027 and ramping in 2028, which is also the timeline the market currently assumes for optical stocks.

• Inside the box: today the signal leaves the switch or accelerator on copper traces and only becomes optical at the front panel. NPO moves the optics next to the chip and copper traces give way to fiber; both NPO and CPO live here, on a late-2027 to early-2028 timeline.

• Within the rack: the copper backplane moves last, from 2029, and only in part. Passive copper is the cheapest and lowest-power option, and everyone he speaks with wants to extend it as long as possible.

He was equally cool on the low-power argument: it was made for LPO too, but without an ecosystem and a driver behind it, hyperscalers don't move away from solutions that work well and cost less. We covered the market split on CPO/NPO timing in Goldman Raises Targets, Eoptolink Pushes Out; this talk's answer is clear: the direction is settled, and the timing is set by how long copper lasts.

6. Why NPO comes first, and the signals he is watching

The last slide is the investment conclusion. Subrahmanyan sees NPO as the most probable first step toward integrated optics, for three reasons:

▲ Three reasons NPO comes first and four signals to watch; what would change his mind is CPO yield arriving early, or copper lasting longer than expected (Image source: Capital Group | 2026 OCP Global Summit IPEC workshop presentation (12:46))

• It keeps the optics ecosystem open and multi-vendor. CPO in its truest form probably needs a switch or accelerator vendor to deliver it, which narrows vendor options; NPO, designed as a standardized pluggable socket, preserves today's many-vendor optical industry.

• No switch die redesign, and the optics stay replaceable. Yield, serviceability and repair time are what operators care about most.

• History repeats. The last time networking chose between a vertically integrated stack and a disaggregated one, hyperscalers chose disaggregation, and that shaped who captured value for the next decade.

This matches our view in NPO Is Not the Appetizer for CPO: NPO is not just a transition step, it decides where value is distributed. The figure below puts the two models side by side:

▲ CPO in its purest form pulls the optical engine and yield risk into the switch or accelerator vendor; NPO keeps the optical engine multi-sourced through a standard socket, so a failure means swapping optics, not the switch (Image source: Simple Tech Trend; data from Capital Group's 2026 OCP Global Summit presentation)

His watch list has four items, plus two conditions that would change his view. Sorted by which outcome they push toward, they form an investor's tracking sheet:

▲ Capital Group's four watch signals and two mind-changers: fiber-coupling yield and standards traction decide whether NPO reaches volume, early CPO yield favors vertical integration, and longer-lived copper delays everything (Image source: Simple Tech Trend; data from Capital Group's 2026 OCP Global Summit presentation)

"Standards traction" refers to socket standards such as Open CPX; for progress, see ECOC2026 | Ciena Turns 6.4T Optical Engines Back into Shelf Parts with Open CPX.

7. Risks and counterarguments

• The speaker is an investor, not a deployer. The talk had no volume yield, cost or power data; "NPO first" is an inference from industry structure and historical analogy, not a measured result.

• 4 to 8x is the speaker's own estimate, and about 200 million units is an industry estimate; neither is an official shipment figure.

• "NPO first" was close to consensus at this workshop. When speakers at one venue largely agree, market expectations tend to price it in early; the real point of disagreement is whether the late-2027 ramp slips again.

• CPO yields could mature faster than expected. That is one of his own view-changers, and Broadcom is already reporting field reliability for its third-generation CPO (see 2026 OCP APAC Summit | Broadcom | CPO's Third-Generation Report Card).

• Copper's lifespan is the biggest timing variable. The slide asks how long copper holds at 400G, while he said 200G and 400G aloud; if 400G copper still covers in-rack reach, in-rack optics will keep sliding past 2029.

8. What it means for Taiwan's supply chain

The talk named no Taiwanese companies; the points below are inferred only from his framework and from information already public at the same summit:

• Disaggregation favors Taiwan, because Taiwanese suppliers mostly sit in replaceable layers. NPO keeps a multi-vendor structure in which switch ODMs, sockets and cable assembly remain separate procurement items. Ciena's Open CPX slides at the same summit showed member logos including Accton, Alpha Networks, BizLink, FIT and LOTES, all Taiwanese or Taiwan-backed; standards traction is exactly one of the signals Subrahmanyan is watching.

• If CPO yields arrive early, value concentrates in chips and advanced packaging. In that scenario Taiwan's most direct position is TSMC: NVIDIA's Spectrum-X Photonics CPO optical engine, presented at the same summit, uses TSMC's COUPE process. Vertically integrated CPO shrinks the role of assembly and module makers but enlarges the packaging side.

• Use late 2027 to 2028 as the baseline. That is the timeline he says the market currently assumes for optical stocks; NPO-related revenue in Taiwan's supply chain would appear in that window at the earliest, while in-rack backplane optics would not start until after 2029.

9. Conclusion

This 12-minute talk had no new technology, but it translated the whole IPEC workshop's technical debate into the language of capital: direction is not in dispute, timing is where money is made or lost, and value capture depends on whether the industry chooses disaggregation or vertical integration. Subrahmanyan bets on NPO first, not because it is the most advanced technology, but because it changes the existing ecosystem the least.

Watch the clock: between-rack and in-box optics start in late 2027 and ramp in 2028, while the in-rack backplane waits until after 2029 and converts only in part. Three things to watch over the next 12 months: NPO fiber-attach yield at volume, whether socket standards such as Open CPX are adopted by switch platforms, and how far copper reaches at 400G.

Technology decides whether optics arrives, copper decides when, and disaggregation decides who keeps the money.

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