OCP Global Summit 2026 | Flex, Lightmatter | Janos Veres, Bijan Nowroozi | CPO Manufacturing Starts After Reflow: Detachable FAUs and Open Reference Designs

1. Why an EMS Company Is Talking About Silicon Photonics

Nowroozi had already presented the "tray contract" behind these designs in a separate talk the day before (10/13). This session pairs him with Flex, and the angle is different: Flex is a large EMS (electronics manufacturing services) provider and Veres runs assembly and test. His question is not how fast the optical engine is, but how to build, test and repair one thousand or ten thousand of these.

The second slide states the position plainly: "Architecture can close on paper. Manufacturing readiness cannot." Starting in January, the two companies worked with roughly 19 to 20 collaborators on an ORv3, DC-MHS and MGX base and actually built a CPO tray. The reason for choosing CPO was practical: Open CPX and XPO are still in development, and CPO is the one they could physically assemble today.

This talk is not trying to persuade anyone to pick CPO. It records where things break when you actually build one.

2. Tolerances: Six Orders of Magnitude, One Handoff

The first challenge is scale. The slide splits the supply chain into three stages: the foundry works in sub-micron tolerances, the OSAT in microns, and system assembly in tens of microns to meters, spanning six orders of magnitude. The speakers' analogy: a hair's width held across a football field.

The order of operations makes it harder. CPO assembly has five stations: package, precision BGA attach (with warpage), optical-electrical connection, fiber routing and system integration. Under the third station the slide adds one line: FAU connections must occur after reflow. In other words, the package leaves the OSAT without fiber, and the fiber is attached at the assembly line.

▲ The five CPO assembly stations: packaging, BGA attach, optical-electrical connection, fiber routing and system integration; the FAU must be connected after reflow (Image source: Flex, Lightmatter | 2026 OCP Global Summit presentation)

Veres was direct: from the assembler's point of view, passive alignment and detachable fiber units are an "absolute must", because attaching fibers individually simply does not work on a volume line. That is the core of the talk: OSAT and assembly need a clear handoff point, and that point is the dFAU.

3. The dFAU: Trading Fiber Density for Serviceability

The slide scores two passive-alignment approaches. A single-row, fixed V-groove FAU passes only on "product performance" and "package edge efficiency". A vertical, multi-row, detachable wafer-level optical coupler (grating or cavity based) is a Yes on yield, UPH (units per hour), known-good optical engine, scalability across package architectures and serviceability. Its only "partial" is package edge, which has to be paid back with fiber density.

The "known-good" row matters most. Once a fixed FAU is glued on, the optical engine cannot be verified on its own; a detachable unit lets the engine be tested first and connected later. The speakers also mentioned an unexpected benefit: foundries want to use these vertical couplers to test undiced wafers and screen known-good dies.

Whether the package edge is sufficient depends on the information density per fiber. Using OIF EEI switch-side edge densities as the baseline (CPO usable at 0.9 Tb/s/dir/mm), the slide calculates the edge left for alignment and retention structures per 96-fiber dFAU:

• 0.2 Tb/s per fiber (DR): pitch binds first, no margin

• 0.4 Tb/s per fiber (OCI, DR-BiDi, where this vehicle sits): about 5 mm

• 0.8 Tb/s per fiber (OCI Gen 2): about 22 mm

• 3.2 Tb/s per fiber: about 124 mm

▲ The higher the information density per fiber, the more room detachable FAUs have for alignment and retention; 8, 4 and 2 dFAUs correspond to 768, 384 and 192 data fibers (Image source: Flex, Lightmatter | 2026 OCP Global Summit presentation)

There are three configurations: 8 dFAUs around a switch, 2 around an XPU:

Configuration	Data fibers	Bandwidth (Tb/s/dir)	Positioning
8 dFAUs	768	154–614	Most bandwidth
4 dFAUs	384	77–307	Easier to manufacture
2 dFAUs	192	38–154	Simplest to service

The takeaway: at today's 0.4 Tb/s per fiber, detachability barely fits; the room only opens up at OCI Gen 2 and 0.8 Tb/s per fiber. The pace of per-lane speed upgrades directly sets the dFAU design margin. We covered the OIF EEI edge-density framework in OIF Drew an Official Map for AI Interconnect; for the beachfront shortage, see also ECOC 2026 | A 1024-Lane Switch Stuck at 0.07 T/mm.

4. Yield: 99.5% vs 95% Is Not a 4.5-Point Gap

The measured data comes in three sets:

Vehicle	Structure	Mean insertion loss	FAU count and yield
Product/Platform Kit (PPK)	2D MCP with V-grooves, 56G NRZ to 224G PAM4	1 dB	181 units, 99.5% (below 2 dB); package test 99.5% (500+ units)
Multi-reticle assembly test vehicle	4,000 mm² 3D chip-on-wafer, edge attach	1.5 dB	62 units, 95% (below 2 dB)
Single-reticle assembly test vehicle	Full-reticle 3D stack	Below 1.5 dB	Above 99% (simulated)

▲ Passive-alignment results on three vehicles: the platform kit reached 99.5% yield over 181 FAUs, the multi-reticle vehicle 95% over 62, and single-reticle values are simulated (Image source: Flex, Lightmatter | 2026 OCP Global Summit presentation)

The slide says yields are "similar to socketed high-rate electrical connectors", but read it carefully: the set closest to a future AI switch is the 4,000 mm² multi-reticle vehicle, at 95% on 62 units. The 99.5% comes from the simpler 2D platform kit, and the single-reticle 99% is still a simulation.

Yield compounds at the package level. STT calculation, assuming each dFAU attach is independent:

▲ Per-dFAU yield compounded to the package level: with 8 dFAUs at 95%, the chance that all pass first time is only about 66%; at 99.5% it is about 96% (Image source: Simple Tech Trend original; data from Flex and Lightmatter, 2026 OCP Global Summit presentation)

• 8 dFAUs at 95% each: about 66% pass first time

• 8 dFAUs at 99% each: about 92%

• 8 dFAUs at 99.5% each: about 96%

This is exactly why detachability matters: when a fixed FAU fails, the whole package may be scrapped; a detachable one can be reseated or swapped. Detachability turns yield loss from scrap into rework, but rework eats UPH, so per-unit yield still has to approach 99.5% for volume economics to work.

5. Test and Qualification: Schedule Risk Sits at L4 and L10

Flex's other focus is modular test. The tray breaks into about six modules: gang-connected FAUs, blind-connect fiber trays, a configurable front panel and blind-connect modular lasers (the Lightmatter Guide DR on page 9, confirmed in Q&A as external lasers). The speakers noted that NVIDIA's keynote on assembly time points the same way: final assembly is just putting six modules together.

Modularity lets each subsystem be measured on its own (FAU insertion effectiveness, optical path, laser thermal effects, failure-chain tracing), including deliberate fault injection on the line. Each finding at a boundary becomes a row of the tray contract.

▲ The L0–L11 build flow from wafer to rack, with schedule risk concentrated at L4 JEDEC qualification and L10 UL/NEBS certification (Image source: Simple Tech Trend original; data from Flex and Lightmatter, 2026 OCP Global Summit presentation)

The backup slide splits the flow into L0 to L11: L4 is multi-chip module assembly plus known-good-die screening (OSAT/foundry), L7 is tray build and routing, L10 is rack build with power and cooling, and L11 is burn-in and validation. Every level has a compliance gate, and the slide is blunt: compliance gates do not define function, they gate shipment; schedule risk concentrates at L4 JEDEC qualification and L10 UL/NEBS certification.

Modularity also has a commercial upside: the same design can switch between FPO, NPO and CPO without a redesign, keeping SKU counts down. It follows the same logic as ECOC 2026 | Ciena Uses Open CPX to Turn the 6.4T Optical Engine Back into Shelf Stock: make optics a replaceable module.

6. Risks and Counterarguments

• Sample sizes are small: the most important vehicle, multi-reticle, has only 62 FAUs, and the digital-twin lifetime analysis at a 10-million-unit scale is still a model.

• Edge margin is thin: at 0.4 Tb/s per fiber there is only about 5 mm left; if per-lane speed upgrades slip, the dFAU mechanical design has no room.

• No reliability numbers: the "Reliability at Every Scale" slide lists test items only (HTOL, temperature and humidity, FIT lifecycle) with no results.

• 2027 HVM conflicts with the spec timeline: specification drafts run through 2027, and the L4 and L10 gates are where schedule risk sits, so volume-grade trays landing in 2027 are unlikely.

7. What It Means for Taiwan's Supply Chain

• Foxconn and QCT are on the collaborator list. The open community on the slide includes Lightmatter, Flex, Celestica, Hyve Solutions, Corning, Keysight, Foxconn, QCT, Dell Technologies and Qualcomm. With two Taiwanese system makers at the table during spec drafting, Taiwan has a chance to help define CPO tray assembly and test rules, not just take orders.

• Value is moving to "after reflow". FAU connection, fiber trays, laser blind-mate and modular test all happen at the assembly end, which is home ground for EMS providers and server ODMs. Whoever first builds dFAU insertion, measurement and rework capability gets the first ticket to CPO tray assembly.

• The OSAT's job is known-good. L4, "MCM assembly plus known-good die", is the OSAT/foundry gate. The slides name no specific OSAT, but detachable FAUs let optical engines be verified before shipment, which adds screening and test work.

• External lasers are moving to high-density modules. In Q&A the speakers said that reaching roughly 200T-class bandwidth in 1RU leaves no room for enough ELSFPs on the front panel, so they use high-density, blind-mate laser modules, while stressing this does not replace ELSFP. For Taiwan's CW laser supply chain, there may be more than one form factor, so both ELSFP and high-density modules need tracking.

8. Conclusion

This talk moves the CPO debate from "which architecture" to "where is the handoff". Flex and Lightmatter's answer: the OSAT delivers a known-good package, the assembler attaches fiber after reflow with detachable, passively aligned FAUs, and lasers and fiber trays become blind-mate modules.

But the timeline needs unpacking. The whitepaper is public, specification drafts run through 2027, and multi-reticle FAU yield is 95% on 62 units. STT's view: what 2027 can deliver is reference designs and base specifications, not volume-grade CPO trays. Over the next 12 months, watch three things: whether multi-reticle dFAU yield reaches 99% on samples of several hundred or more, the timing of OCI Gen 2 at 0.8 Tb/s per fiber, and when the MHS Interconnect-for-AI workstream publishes its first base specification.

Whether CPO reaches volume production is not decided inside the optical engine, but in that one insertion after reflow.

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