OCP Global Summit 2026 | Meta | Granthana Rangaswamy, Siamak Amiralizadeh, Xu Wang | Integrated Optics at Scale: From 51.2T Validation to 102.4T

1. What This Talk Is: A Close-Out Report, Not a Tech Demo

Meta optical engineer Siamak Amiralizadeh said this was the close-out of Meta's 51.2T CPO evaluation. Interim results had been presented at ECOC 2025 and elsewhere; this talk pulled everything together and laid out the next step: a 102.4T integrated-optics product.

Meta lists four reasons for integrated optics: power, density, disaggregation and reliability. The power chart is the most direct: at 100G/lane, per 800G, a retimed pluggable draws about 15 W, LPO about 8.3 W and CPO about 5.4 W (STT reading of the box-plot medians), and the slide labels the savings as 45% and 64%.

But power is not where this talk's value lies. Meta wrote it on the slide itself: the main hurdle for integrated optics is unknown RAS (reliability, availability, serviceability) at scale. This talk is a RAS field report.

2. The Test Design: an A/B Test Against Meta's Own Minipack3

Meta ran pluggables and CPO in the same lab as an A/B test, with the pluggable-based Minipack3 as the control group.

Group	Operating temperature	400G ports	400G port-hours
FRO, 2x400G FR4	40°C	1,792	8M
CPO 128x400G FR4, Phase 1	40°C	8,960	50M
CPO 128x400G FR4, Phase 2	15–35°C	38,400	175M

Phase 1 was deliberately run under stressed high-temperature conditions to surface failure modes; Phase 2 used updated build standards and normal operating temperatures to produce higher-confidence interruption statistics. At 128 400G ports per system, that is roughly 70 systems in Phase 1 and 300 in Phase 2 (STT estimate). In total Meta logged more than 1.8 million system hours.

The criterion matters: in the backend AI network, a single uncorrectable codeword (UCW) counts as an interruption. That is stricter than the frontend network, because one interruption stops a training job.

3. Where Failures Come From: CPO's Problem Is Concentrated in One Component

▲ Multiplying each failure category's share by the annualized interruption rate: FRO's 1.22% is spread across DSP firmware, transients and the optical engine, while more than 80% of CPO Phase 2's 0.2% comes from a single failure mode, the ELS (Image source: Simple Tech Trend original; data from Meta, 2026 OCP Global Summit presentation)

The failure mix of the three groups:

• FRO: DSP firmware 36%, transient/unknown 37%, optical engine hardware 18%, fiber/connector 9%.

• CPO Phase 1: ELS 85%, fiber/connector 11%, optical engine hardware 4%.

• CPO Phase 2: ELS 83%, fiber/connector 17%, optical engine hardware 0%.

Three points. First, the ELS failures were root-caused to one SMT component in the laser-driver circuit, not the laser chip itself, and Meta has fixed and verified it through a design change. Second, transient interruptions never appeared in CPO; these are short bursts of UCWs that never repeat, account for about one third of FRO failures, and are extremely hard to root-cause. Third, fiber/connector failures appear in all three groups; Meta says dirty connectors remain unsolved and it is working on expanded-beam optics and an optical backplane.

Another counterintuitive finding: CPO Phase 1 had an average pre-FEC BER of about 9e-8 versus about 1e-9 for FRO, nearly two orders of magnitude worse, yet CPO had fewer interruptions. Meta's conclusion: field failures are sudden error bursts, not gradual BER degradation, so pre-FEC BER cannot serve alone as a link-health metric. That is a warning for any optics qualification flow built on BER thresholds.

4. Interruption Rates: 6x Is Measured, 30x Needs a Discount

▲ Meta's 51.2T CPO RAS results: FRO annualized link interruption rate of 1.22%, CPO Phase 2 at 0.2%, 0.04% excluding ELS component failures, and zero non-serviceable interruptions (Image source: Meta | 2026 OCP Global Summit presentation)

Group	Assumption	Annual link interruption rate (ALIR)	MTBI (hours)
FRO	All interruptions	1.22%	0.72M
CPO Phase 1	All interruptions	0.47%	1.85M
CPO Phase 2	All interruptions	0.2%	4.5M
CPO Phase 2	Excluding ELS SMT component failures	0.04%	25M
CPO Phase 2	Non-serviceable interruptions	0	175M

The ratios: about 2.6x in Phase 1 (about 3x as stated on stage), about 6x in Phase 2, and about 30x excluding the ELS mode. Meta attributes this to fewer interfaces, system co-design, system-level screening on the manufacturing line, and less human intervention.

STT's caveat is statistical. Multiplying the rate by port-years, the FRO control group's 8M port-hours are only about 913 port-years, or about 11 interruptions; CPO Phase 2 excluding ELS is about 8 (about 7 from the MTBI). Using a Poisson approximation, STT puts the 95% confidence interval for the 30x figure at roughly 12x to 75x (and 6x at roughly 3x to 12x). The 175M hours for non-serviceable interruptions is also only a lower bound from zero events. Meta itself is cautious: results from one generation and one vendor do not mean all integrated optics are more reliable than pluggables.

5. 102.4T: Turning the Failure List into Design Specs

Granthana Rangaswamy presented the 102.4T integrated-optics switch, and nearly every design decision maps to a failure from Section 3:

• ELS was about 80% of failures → field-replaceable laser modules, compliant with OIF ELSFP to enable multi-sourcing.

• Fiber array units (FAU) can break only during temperature cycling → factory-assembled and pre-routed, untouched once in the data center; data ports use detachable fiber connectors.

• Dirty fiber → 100% liquid cooling, with cold plates on the ASIC, optical engines and ELSFPs; the stated reason is dust-contamination risk, not just thermals.

• Specs: 1.5OU, 64 data ports (16x4 MMC16 connectors), 16 ELSFP ports; TP2/TP3 compliance for interoperability with existing retimed pluggables; CMIS 5.3; 500 m reach spec, capable of 2 km; 60% optics power saving versus retimed pluggables.

▲ Meta's 102.4T integrated-optics switch architecture: 512 lanes of 200G, 64 MMC16 data ports and 16 ELSFP laser ports connected to the optical engines through FAUs, 100% liquid-cooled (Image source: Simple Tech Trend original; data from Meta, 2026 OCP Global Summit presentation)

▲ Front and rear views of the 102.4T integrated-optics switch POC: 64 MMC16 data ports on the top row of the front panel, 16 ELSFP laser ports below, and 48V DC and liquid-cooling connectors at the rear (Image source: Meta | 2026 OCP Global Summit presentation)

The contrast with IceCube, the pluggable 102.4T switch Meta presented in a separate talk at the same summit, is clear: that box is 2OU with 340 m of internal flyover copper and about 85% liquid cooling; the integrated-optics version is 1.5OU, and Rangaswamy said CPO can "completely eliminate the BGA, vias, PCB routing and flyover cables," buying back signal margin at 212.5 Gb/s per lane.

But the board does not get simpler. Optical engines need many power rails, so the layer count stays high; only the high-speed material grade may come down. Power integrity and placing power components relative to the optical engines are the major challenges Meta named.

6. CPO or NPO: at 200G It Is a Supply-Chain Question, at 400G a Physics Question

Meta's comparison of the two paths is candid:

Aspect	CPO	NPO
Optical engine location	Directly on the ASIC package substrate	Near the ASIC, linearly driven by the SerDes
102.4T power	Baseline	Comparable to CPO (200G/lane)
Scaling to 400G	Fewest interfaces, most scalable	Linear drive degrades at 448G, retimers may be needed
RAS	Partially proven at 51.2T	New mechanics and failure modes, to be proven
Supply chain	Tied to ASIC vendor, higher risk	Modular interposer, easier multi-vendor

Amiralizadeh's strongest line: Meta has run multi-vendor pluggables for a decade, and "moving to a single vendor in the CPO case is a huge risk for us." That matches the scoring logic in ECOC 2026: Huawei Scored CPO vs NPO, and NPO Led 4:2: in the 200G generation, NPO's edge is not physics but procurement.

The Q&A also revealed Meta's red line: no retimers in integrated optics, because once serviceability is sacrificed and a retimer is added back, little of the power and cost return remains. A speaker added in the Q&A that at 400G/lane a pluggable OSFP-style form factor is very hard to achieve: "it's either NPO or CPO." For CPO's reliability progress, compare 2026 OCP APAC Summit: Broadcom's Third-Generation CPO Report Card.

7. Risks and Counterarguments

• Small sample: see Section 4; the 30x figure is sensitive to a single event.

• "Excluding ELS" is a post-hoc condition: how the fixed ELS performs in the 102.4T generation needs new data.

• One vendor, one generation: Meta says the result cannot be extrapolated; if 102.4T changes vendor or moves to NPO, RAS must be proven again from scratch.

• The fiber problem is not solved: its share rose from 11% to 17%, but its annual rate actually fell from about 0.05% to 0.03% (STT estimate); the issue is that it becomes the only failure category left once ELS is excluded, and expanded-beam optics and optical backplanes are still at the demo stage.

• No 102.4T timeline: the slides only say "next target" and show POC photos, with no production or deployment date.

8. What It Means for Taiwan's Supply Chain

Meta named no suppliers, so the following covers only what the disclosed component specs imply:

• The external laser module is CPO's largest failure source and also the clearest multi-sourcing entry point. Meta requires OIF-compliant ELSFP precisely to buy from multiple vendors; with the root cause in an SMT component of the laser-driver circuit, the module's circuit design and PCBA quality matter as much as the laser chip itself. For Taiwanese laser-chip, module-packaging and PCBA makers looking to enter, reliability data will persuade more than spec sheets. For laser power and architecture trade-offs, see CPO Lasers Don't Have to Chase 400 mW: NPO with Integrated Lasers.

• Fiber assembly value shifts to the factory. FAUs are pre-assembled and pre-routed at the factory, and the front panel moves to high-density MMC16; fiber-assembly yield and temperature-cycling reliability become part of the switch manufacturing flow.

• PCB demand changes shape: layer counts stay high, but high-speed material grades may come down, and power-integrity design becomes more valuable.

• In this design, liquid cooling goes from optional to required: ELSFPs get cold plates too, for dust protection and not just heat.

9. Conclusion

What Meta's close-out report really rewrites is CPO's risk list. For two years the most common argument against CPO has been "the optical engine is soldered to the package, so one failure scraps the whole switch"; Meta's data is 175M port-hours with zero optical-engine failures. What remains is one laser-driver component and a bundle of fiber, both of which can be handled through design and process.

So the 102.4T contest is no longer "is CPO reliable enough" but whether ELSFP can be multi-sourced, whether FAUs can be mass-produced, and whether CPO ties Meta to a single vendor. At 200G/lane those questions favor NPO; at 400G/lane physics tips the balance back to CPO. Over the next 12 months, watch three things: a second volume ELSFP supplier, the path of the first 102.4T deployment, and the event counts and confidence intervals in the JLT paper.

51.2T proved the optical engine doesn't break; 102.4T has to prove that nobody gets locked in.

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