OCP Global Summit 2026 | Meta | Amin Yekani, Qing Wang | Operating LPO at Hyperscale: Fleet Reliability Data

14 hours ago

7 min read

1. Three Years, Three Steps: Meta Finally Puts LPO to the Test

Everyone can recite the LPO (Linear Pluggable Optics) pitch: remove the retimer DSP from the module and you save power, heat and cost. Meta's slide puts it more bluntly: at AI scale, the per-module DSP is a "repeated tax" multiplied across millions of links.

The cost is just as clear: there is no DSP safety net in the module. Equalization falls on the switch-side SerDes, and the triage capability that used to come from the DSP disappears with it. So the industry has really been asking one question: can LPO hold up in large-scale production?

Meta broke the journey into three years:

Year	Phase	What was done
OFC/OCP 2024	Definition	Defined Meta's LPO architecture and module spec, informed by the LPO MSA
OCP 2025	Productization	Qualification, host and module calibration, diagnostics, path to production
OCP 2026	Reliability at scale	Live production data plus a controlled lab to answer "can it beat retimed"

Most of the spec table has not changed in three years: the pre-FEC BER target stays below 3E-6, and the fiber budget stays at 500 m and 2.5 dB insertion loss. Only three rows moved: interoperability covers LPO-to-LPO only, with LPO-to-retimed still in progress; TP2 compliance now uses an LPO-specific spec that is tighter than the retimed lite spec; and host SerDes settings decoupling has been proven, but production still runs combined switch-plus-module calibration.

▲ Three link architectures, retimed, LRO and LPO: LPO removes the DSP at both Tx and Rx and is driven directly by the switch SerDes (Image source: Meta | 2026 OCP Global Summit presentation (p. 2))

2. Breaking Down the Method: Age, Network Layer, Link Mode

The most valuable part of the talk may not be the result but how Meta made the comparison fair. The slide title says it plainly: comparing linear optics to retimed is not a hookup-and-count. Meta controlled three variables:

Deployment age: new fleets start with more issues and settle over time. LPO was a young, limited deployment and retimed a mature fleet, so a direct comparison would penalize LPO.
Network layer: a direct fiber and a path through patch panels and patch cords have completely different failure modes, so the comparison must be made at the same layer.
Link mode: in 1x800G mode, eight lanes form one link, so one bad lane takes down the full 800G; in 8x100G mode only that 100G goes down. The blast radius differs, so both sides must use the same mode.

The production setup is 2x400G FR4 modules running in 1x800G mode on the DTSW-to-EDSW hop, with LPO on both ends, every other hop still retimed, and FBOSS as the network OS.

3. A New Yardstick: From Failure Rate to Interruption Rate

Meta deliberately avoided the familiar failure rate. Its reasoning: failure rate is polluted by thresholds. How many interruptions count as one failure? Triage and service workflows then add their own rules.

AI training is harsher. A large synchronous training job runs across many accelerators, and the upper layers can absorb only very shallow interruptions; most interruptions can send the job back to the last checkpoint. So Meta counts interruption rate instead: every uncorrected codeword counts.

The last line of that slide is the real point: as optics move toward the package (NPO, CPO), a failed link can no longer be pulled and swapped like a pluggable. Giving up serviceability only pays off if reliability climbs to match. This matches the logic we saw in OCP APAC Summit 2026 | Broadcom | CPO's Third-Generation Report Card: for CPO to be accepted, it must first show better link stability than pluggables.

▲ Failures were the old bar, interruptions are the new one: any link event can push AI training back to the last checkpoint (Image source: Meta | 2026 OCP Global Summit presentation (p. 7))

4. Breaking Down the Production Data: 2x, 2.6x, 8.4x

With retimed set to 1.0 per link, in the same link mode, live production showed:

Interruptions: LPO 2.0x fewer (full April-to-August record)
Service-affecting failures: LPO 2.6x fewer
Interruptions in the last 60 days: LPO 8.4x fewer

▲ Meta production data: LPO has 2.0x fewer interruptions and 2.6x fewer failures, and over the last 60 days the interruption gap widened to 8.4x (Image source: Simple Tech Trend; data from Meta's 2026 OCP Global Summit presentation)

8.4x is the headline number, but it needs to be read correctly. Meta explained that the full record compares a very young deployment with a very mature fleet; the last 60 days is the closest match in maturity. In other words, 2x is the conservative figure, and 8.4x is the trend once deployment issues fade.

The speakers' interpretation is worth noting: without a DSP, it is hard at first to tune performance up to retimed levels, but once you get there, there is one fewer class of parts that can fail. The slide also states that the two optics are on par at the module level, and the advantage is architectural.

5. Breaking Down the Lab Data: Zero Failures and 46% Less Power

Production is complex, so Meta isolated module behavior in the lab: about 1,000 LPO modules in optical loopback on 51T MiniPack3 switches, in a continuous soak with fan speed held fixed throughout.

Power stability: 8 lanes over 104 days, optical power within 0.25 dB peak-to-peak and per-lane drift within 0.04 dB
FEC load: typical FEC bin of 2 for both linear and retimed, worst case 5 for LPO versus 7 for retimed, a tighter tail and wide margin to codeword loss
Pre-FEC BER median of about 1E-9
Zero module-attributable failures over 2.96 million link-hours
At the same fan speed, 46% lower power per module and 17.5°C cooler

▲ Controlled lab: median pre-FEC BER around 1E-9, 46% lower module power, 17.5°C lower temperature and zero module failures (Image source: Meta | 2026 OCP Global Summit presentation (p. 9))

Production and lab, two different methods, point in the same direction, which is why Meta says the result can be trusted.

6. What This Data Means for the LPO, NPO and CPO Debate

For two years the strongest argument against LPO was that nobody had proven it in large-scale production. That argument is now much weaker. Meta did not show a lab demo; it showed 4,000 modules running for four months on a production network OS.

The deeper message, though, is that LPO succeeds by moving link responsibility from the module to the host. Equalization, calibration and diagnostics all sit on the switch side, and TP2 has to be tightened. That is exactly the road NPO and CPO must take. In NPO Is Not a Stepping Stone to CPO we argued that NPO's value lies in keeping part of the serviceability; Meta has now added a quantitative condition to that debate: every step closer to the package requires pushing the interruption rate down another notch.

Meta's conclusion was also restrained: the 100G/lane experience is methodology, not a guarantee for 200G/lane. Asked about 200G/lane data in Q&A, the speakers said the work is ongoing, there are no numbers to share yet, and LPO at 200G/lane is "an option."

▲ Scope of Meta's LPO evidence: at 100G/lane the two methods agree; 200G/lane, LPO–retimed interoperability and NPO remain to be validated (Image source: Simple Tech Trend; data from Meta's 2026 OCP Global Summit presentation)

7. Risks and Counterarguments

One hop, four months: LPO sits only on the DTSW-to-EDSW hop, and four months is still short relative to optical module lifetime; long-term aging data is not yet available.
Combined calibration does not copy over directly: production uses combined switch-plus-module SerDes calibration, which must be redone for a different switch or module. Decoupling has been proven but is not yet in production.
The 46% saving is at the module: LPO pushes equalization onto the host SerDes, and Meta did not give the net system-level power saving in this talk.
Lab conditions are favorable: the loopback test ran idle traffic, and an audience member noted that comparing at the same fan speed lets LPO run cooler; comparing reliability at the same case temperature could change the result.
Interoperability is still the bottleneck: LPO-to-retimed interop is still in progress, and the link management challenges of mixed multi-vendor links are not covered by this data.

8. What It Means for Taiwan's Supply Chain

For module makers, the bar shifts from "can build it" to "can tune it in." Meta's LPO production relies on a tighter TP2 spec, combined switch-plus-module calibration and new Driver and TIA diagnostics. For Taiwanese module makers, the ability to calibrate and validate jointly with the switch side matters more than standalone module specs in getting onto a hyperscaler's list.
Value moves from the DSP to analog components and light sources. Without a DSP, signal quality depends directly on the linearity and stability of the Driver, TIA and laser. Data like 0.04 dB of optical power drift over 104 days will become the new acceptance language for light source and optical subassembly suppliers.
Reliability metrics will rewrite laser datasheets. Acceptance based on interruption rate means laser and module makers must provide not just MTBF but long-term stability data that maps to link events. The same applies to the light-source role discussed in Is Nvidia's Shift from CPO to NPO Bad News for the Laser Supply Chain?.
200G/lane is the next hurdle. Meta has not committed to 200G/lane LPO, so Taiwanese suppliers planning 1.6T linear products should prepare LRO and retimed options in parallel rather than betting on a single path.

9. Summary

This talk marks a milestone in the LPO debate: at 100G/lane, with Meta's scale and method, LPO not only holds up but is more reliable than retimed. 2x is the conservative figure and 8.4x is the trend after the fleet matures; both point to the same thing: one fewer class of components means one fewer class of failures.

Meta's bigger contribution, though, is the new yardstick. Moving reliability from failure rate to interruption rate sets a higher bar for every optical approach that moves toward the package (NPO, CPO). Over the next 12 months, watch three things: production data for 200G/lane LPO or LRO, progress on LPO-to-retimed interop, and when decoupled SerDes calibration enters production.

LPO's win is not a win for optics but for link management: whoever can tune the link from the host side gets to remove the DSP.

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