OCP Global Summit 2026 | Source Photonics | Frank Chang | Revolutionizing AI Scale Up with Pluggable NPO

1. Why Source Photonics Opened the Session

This was the optical interconnect workshop hosted by IPEC (International Photonics & Electronics Committee) at the OCP Global Summit, titled "Revolutionizing AI Scale Up with Pluggable NPO: Why Does It Matter?". It was co-organized by Frank Chang of Source Photonics, Eric Bernier of Huawei and Jianwei Mu of Ligent, with 14 speakers in the afternoon, from LightCounting, Alibaba, Meta, Oracle and Microsoft to Ciena, Lumentum, NVIDIA and Huawei.

IPEC positions itself as an open, unified, equal and innovation-driven industry forum. Frank Chang said it now has more than 55 members, spanning system vendors, telecom operators, and component and module makers. The member list on the slide is dominated by Chinese companies, including Huawei, ZTE, China Telecom, Accelink, Innolight, YOFC and Hengtong, alongside Intel, Marvell, Keysight, Semtech, MACOM and Senko.

Source Photonics comes to this table as a "pluggables company". The deck cites Cignal AI's 1Q26 leaderboard: it ranks 4th in 400GbE shipments, 5th in 1.6TbE, and 4th in datacom optics revenue. A module maker that lives on pluggables opening an NPO workshop is itself a statement: in its view, NPO is an extension of the pluggable roadmap, not a replacement.

2. Following the Money: About $5.5 of Every $100 Goes to Optical Transceivers

The starting point is hyperscaler capex: Amazon, Microsoft, Google and Meta are spending roughly $765 billion on AI in 2026, and about $7.6 trillion cumulatively from 2026 to 2031.

Where it goes, per an illustrative BNP Paribas breakdown (via a16z, August 2026) shown on the slide:

• Semiconductors $50: accelerators 25, memory 15, CPUs and other chips 10

• Networking $15: network processors 3, cabling 2, switches 4.5, optical transceivers about 5.5 (the figure is obscured by the highlight circle on the slide; inferred from the total)

• Power $20, cooling $7.5, facilities $7.5, with land at only $1

▲ Of every $100 of AI capex, $15 goes to networking and about $5.5 to optical modules (Image source: Source Photonics | 2026 OCP Global Summit presentation (03:22))

The market-size comparison is even more direct. Global chip market value grew from $2.44 trillion in November 2022 to $18.8 trillion in July 2026; the optics market grew from about $10 billion at ChatGPT's launch to about $50 billion in 2026, and Frank Chang estimated it could double again to around $100 billion by 2030-2031. The line at the bottom of the slide sets the tone for the whole session: "Optics has moved to the top of the semiconductor value chain".

3. The Scale-Up Market Math: 10% Penetration Recreates a Whole Market

Cignal AI data shows optical transceivers grew 4x in two years, but the slide is explicit: most revenue today comes from Open MSA pluggable modules used in scale-out. The next leg of growth has to come from scale-up.

EA splits AI networks into three tiers by reach and gives relative market sizes:

• Scale-across (10-100 km, between data centers): coherent or IMDD pluggables, about 0.1x

• Scale-out (100 m-2 km): IMDD, DSP, LPO and LRO pluggables, 1x

• Scale-up (1-10 m, inside the rack): DAC, AEC, ACC and AOC cables, 5-10x

▲ EA estimates the scale-up market at 5 to 10 times scale-out, currently dominated by copper (Image source: Source Photonics | 2026 OCP Global Summit presentation (06:58; data from EA))

Scale-up is still copper territory today. Frank Chang's argument is that as lane rates rise, copper runs out of reach and space, which is where optics starts to come in. The math is simple: scale-up is 10x, 10% penetration is 1x, which doubles today's entire optics market.

▲ With scale-up at 5 to 10 times scale-out, 10% to 20% optical penetration equals 0.5 to 2 times today's market (Image source: Simple Tech Trend; data from Source Photonics' 2026 OCP Global Summit presentation)

The force of this math lies at the low end of the range: even at 5x and 10%, it adds half of today's market. That is also why the same workshop features a dozen approaches, from slow-and-wide to fast-and-narrow: everyone is chasing the same 0.5x-2x increment.

4. A Fourth Tier, Scale-In: Deeper Than Scale-Up

Frank Chang also introduced a fourth tier, scale-in: GPU/TPU to CPU, chip to chip, with the shortest reach and the least standardization. The examples on the slide are TSMC COUPE CPO (with Ayar Labs), Ranovus NPO optical engines and the Lightmatter Passage optical interposer, over interfaces such as PCIe, CXL and UALink, all copper today.

The description of the scale-up tier deserves more attention: "the highest-bandwidth, most cost-sensitive, lowest-latency tier", with many MSAs, but highly customized for every hyperscaler. Examples include NVLink, Photonic Fabric, UCIe optical chiplets and OCI/CPX/XPO.

In other words, the biggest challenge for scale-up optics is not that the technology cannot be built, but that demand is split into several custom specifications, which directly erodes the economies of scale that pluggables do best.

5. History Repeats: The Post-1.6T Form-Factor War

This is the most opinionated slide of the session. Frank Chang compared today with the chaos after 10G:

Era	Options in play at the same time	Converged on
Post-10G	CFP, CFP2, CFP4, CFP8, CDFP, CFP-DCO, CFP2-ACO, QSFP	QSFP28 (LR4/CWDM4)
Post-1.6T	OSFP-XD, CPO, NPO, IPO, OIO, XPO, micro-LED, micro-VCSEL, 400G/lane, SiPh/EML, InP, TFLN	Not yet converged

▲ The post-10G era converged on QSFP28, while the post-1.6T era has more than ten package and material options in play at once (Image source: Source Photonics | 2026 OCP Global Summit presentation (10:34; data from EA))

In the last cycle, from 100G QSFP28 and 400G QSFP-DD to 800G OSFP, the industry spent roughly a decade converging on pluggables. Now it is back at a crossroads: NVIDIA has started shipping CPO, some hyperscalers have committed to upgrading pluggable networks to NPO, XPO drew a lot of attention after its launch at OFC 2026, and NVIDIA and Cisco announced the 12.8T, 400G/lane USFP on the day of the talk. The note beside the slide is blunt: slow-and-wide and fast-and-narrow are both running, and the market looks hardest on "fast and wide".

LightCounting's shipment forecast adds timing: 400G declines from 2026, 800G peaks in 2028-2029, 1.6T jumps from about 3 million units in 2025 to more than 20 million in 2026, and 3.2T (400G/lane) ramps from 2028 to about 46 million units in 2030.

6. Source Photonics' Bet: XPO and Open CPX

After the industry view, Frank Chang laid out his own company's roadmap: Source Photonics is a founding member of XPO and a contributor member of the Open CPX MSA, with products lined up by reach (power figures are STT readings from the slide's axis):

Product	Application	Reach	Power (approx.)
6.4T-OpenCPX-4DR8 (linear)	Scale-up	Under 100 m	Below 5 pJ/bit
12.8T-XPO-8DR8-RTLR	Scale-up / Scale-out	About 100 m	About 10 pJ/bit
12.8T-XPO-16FR4-FRO	Scale-out in DC	1 km and beyond	About 15 pJ/bit
12.8T-XPO-Coherent-Lite	DC campus	About 10 km	20 pJ/bit and above
12.8T-XPO-ZR-Coherent	Scale-across	About 100 km	About 30 pJ/bit

▲ Source Photonics' Open CPX and XPO products arranged by reach and power, from scale-up to scale-across (Image source: Source Photonics | 2026 OCP Global Summit presentation (11:30; data from MSA))

The message of this chart: one 12.8T XPO form factor spans 100 m to 100 km, while the shortest scale-up segment goes to the Open CPX linear optical engine. The Open CPX MSA aims to turn 6.4T optical engines into interchangeable standard parts (see ECOC 2026: Ciena Turns 6.4T Optical Engines Back into Off-the-Shelf Parts with Open CPX). The module makers' logic is consistent: as long as the optical engine stays pluggable and under an MSA, replaceability and multi-sourcing are preserved, which is NPO's core selling point over CPO, as we argued in NPO Is Not an Appetizer for CPO.

7. Risks and Counterarguments

• 5-10x is a relative magnitude, not a revenue forecast: EA's multiples carry no unit or year, and copper is extremely cheap in scale-up, so the same number of links moved to optics does not necessarily scale revenue proportionally.

• Fragmentation hurts module makers: the slide itself admits scale-up is "highly customized for every hyperscaler"; the more fragmented the specs, the smaller the scale advantage of pluggables.

• The routes will cannibalize each other: CPO, NPO, XPO and USFP are advancing at the same time, and on LightCounting's forecast chart every route ramps between 2027 and 2030, so who takes the main share of scale-up is still open.

• Product power figures are axis readings: the pJ/bit numbers in Section 6 were estimated from the slide's log-scale axis; Source Photonics did not provide a spec sheet.

• This was an opening, not a product launch: the session gave no production timeline or customer names for any Source Photonics product.

8. What It Means for Taiwan's Supply Chain

• At the scale-in tier, TSMC is already on the list. The slide lists TSMC COUPE CPO (with Ayar Labs) as a representative scale-in case; if scale-in becomes a fourth tier of optical demand as Frank Chang suggests, TSMC's advanced packaging is the first to benefit.

• Taiwan has few seats at the IPEC table. The only identifiable Taiwanese member on the slide is Foxconn; the rest are mostly Chinese companies and US equipment and chip vendors. Scale-up specs are taking shape at the MSA stage, and being absent means following the spec after the fact.

• If pluggables continue, manufacturing value continues. XPO and Open CPX are both pluggable form factors, so demand for module assembly, connectors and test will grow along with scale-up, but the talk named no Taiwanese supplier, and who wins volume depends on each MSA's actual supplier list.

• We covered the trade-offs among the three routes in OCP APAC Summit 2026: CPO / NPO / XPO Panel, and Taiwanese suppliers' entry points shift depending on which route wins.

9. Conclusion

The value of Frank Chang's opening is that it reframes scale-up optics from a "technology choice" into "market arithmetic": scale-up is 5-10x scale-out, and if 10-20% moves to optics, the optics market doubles again. That number explains why every optics conference in 2026 is arguing about scale-up.

But the arithmetic is easy and the specification problem is hard. The post-10G era took about a decade to converge on QSFP28; the post-1.6T era has more options and more customized customers. Source Photonics has staked its position with XPO and Open CPX: pluggables are not leaving. Over the next 12-18 months, three things are worth watching: the shipment timing of the first 12.8T XPO modules, whether Open CPX linear optical engines win hyperscaler scale-up designs, and whether USFP or XPO is adopted by a second system vendor first.

The winner in scale-up optics will not be whoever has the most power-efficient optical engine, but whoever first converges custom demand into one pluggable spec.

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