OCP Global Summit 2026 | Credo | Chris Collins | Standardized Transceiver Telemetry

1. The problem is not that optics fail, it is that failures can't be traced

AI clusters fail differently from traditional data centers. Collins highlighted three shifts on his slides:

• Scale: Optical transceivers are deployed at unprecedented volume, so even rare issues become daily operational events.

• Access: Bare-metal servers mean operators can neither physically inspect nor directly query the module on the GPU server side.

• Intermittency: Link flaps, LOL/LOS (loss of lock / loss of signal) events and marginal optical paths can vanish before debugging even begins.

His scenario: a link flaps at 2 a.m., and by the time an engineer arrives every LED is green again and the evidence is gone. Ten thousand or a hundred thousand GPUs sitting idle cost far more than any single transceiver.

This extends Credo's ZeroFlap narrative; we broke down its optical revenue in Earnings Highlights: Credo (CRDO) | Q1 FY2027. What is different this time: Credo took its own differentiating feature to OCP and turned it into a spec anyone can implement.

2. Breaking down the spec: a link is three segments, not one fiber

The spec's first conceptual shift is to stop treating a link as a single fiber path. Switch to local module is an electrical interface, module to module is the optical path, and remote module to host NIC is another electrical interface. A degraded optical signal and a bad host electrical signal need completely different fixes, so the three segments are scored separately.

▲ A link is split into three segments: switch electrical, optical, and host electrical, each producing its own health evidence (Image source: Credo | 2026 OCP Global Summit presentation (p. 3))

On that basis, the workstream defined six core capabilities:

Capability	What it does
Mission-mode segmented monitoring	Tracks host and optical receivers continuously while traffic is live
Link health reporting	Normalizes complex telemetry into a single health indicator
Remote telemetry harvesting	Reads far-end module data from the switch, no server login needed
Persistent event logging	Writes time-stamped events to non-volatile module memory
Host telemetry access	A common cross-vendor host interface built on CMIS CDB
Remote firmware management	Firmware operations on local and remote modules

Telemetry metrics fall into three buckets. Signal quality: SNR, eye height, PAM4 LTP, MPI (multipath interference). Error and recovery: pre-FEC BER, FEC histogram, uncorrectable codewords, LOS/LOL, uptime. Context and environment: case temperature, laser temperature, voltage, Rx optical power, link-partner serial number. Most of these numbers already exist in today's modules; the spec's value is defining how they are combined, read and kept consistent across vendors.

3. Breaking down the health score: follow the worst metric, not the average

The workstream first considered averaging, but an average hides a single low score. Credo's slide example: optical 94, electrical 92, errors 91, thermal 89, remote 93. The average is 92 and looks acceptable, yet the thing that needs attention is the thermal score of 89.

▲ The five category scores average 92, but the spec takes the lowest category, 89, as the link health score so a single degradation is not masked by the average (Image source: Simple Tech Trend; data from Credo's 2026 OCP Global Summit presentation)

So the spec "follows the worst": the health score equals the lowest category score, reported together with the category that dragged it down. Operator language shifts from "this port is bad" to "the remote optical receiver is degraded due to MPI" or "the host electrical eye is marginal". Only when location and cause arrive together can you decide whether to clean a fiber, reseat a module or replace a NIC.

The host interface is deliberately simple: built on CMIS CDB, a single command returns identity, health score, full telemetry, operating conditions, alarms and flags.

4. Breaking down event logging: keeping the 2 a.m. flap on record

The hardest problems to debug are the ones that heal themselves. The spec's answer is to give every module a "time machine":

• When plugged in, the module is provisioned with a system time shared across the whole fabric

• When any metric crosses a programmable threshold, the event is written to non-volatile module memory, e.g. reset, LOL/LOS, SNR threshold, FEC threshold, MPI threshold, optical power, temperature, firmware operations

• After a flap or failure, the logs are retrieved, the timeline rebuilt and correlated with other events across the fabric

▲ Event logging flow: provision system time, record link events, then retrieve the log after a failure and rebuild the timeline (Image source: Credo | 2026 OCP Global Summit presentation (p. 8))

The cost is small: Collins said in the Q&A that storing 250 to 500 events takes only a few hundred KB of memory. Remote harvesting runs in-band without interrupting traffic, and the document describing it was published this week.

5. The trial: 48 modules, two hidden faults on "healthy" links

Credo ran a trial with its own ZeroFlap optics: an Arista switch with 16 ports of 400G QSFP-DD DR4 (2x200G), connected to 16 servers, each with two 200G QSFP DR2 modules into NVIDIA CX7 NICs, 48 modules in total, all compliant with the OCP spec.

▲ Test setup: 16 switch ports of 400G DR4 connected to 16 servers with 2×200G DR2, in three optical attenuation groups of 0 dB, 5 dB and 10 dB (Image source: Credo | 2026 OCP Global Summit presentation (p. 10))

At the startup pre-check every link reported link up, yet telemetry caught two issues:

• Switch port 12 reported high BER. Reading local and remote data from the switch side alone pinpointed a low electrical eye on the server module's host interface, health score 40; pulling, cleaning and reseating it fixed the issue.

• Port 18 had a health score of 60 due to MPI, which usually indicates reflections in the optical path. This typically goes undetected at install time and could take down a training run two weeks or two months later; this time a fiber cleaning fixed it.

Optical attenuation was then added on purpose. Scores tracked attenuation almost one-to-one:

Attenuation	Ports	Health score	Remote Rx power	Status
0 dB	9/17/18/19/20	80	approx. +0.3 to +1.9 dBm	All good
5 dB	1/3/5/6/10/11/12	Mostly 60 (two at 80, two at 40)	approx. -6.1 to -3.1 dBm	Flag or eye-height warnings
10 dB	25/26/27/28	40	approx. -9.6 to -8.3 dBm	Alarm state, 1 link down

▲ Across 32 server-side optical receive measurements on 16 ports, health scores split into three tiers of 80, 60 and 40 by received optical power, yet link status flagged only 1 link as abnormal (Image source: Simple Tech Trend; data from Credo's 2026 OCP Global Summit presentation)

The key is the last column: only 1 link in the 10 dB group went down; every other link reported link up. Looking at link status alone, operators could not tell these three link qualities apart.

6. Who benefits, and whose moat gets leveled

• Hyperscalers are the biggest winners. Without being tied to one supplier's proprietary tools, they can screen an entire fabric with one command set; the slides claim an automated workflow can process a whole fabric in minutes.

• DSP vendors are the ones actually implementing the spec. Segmented monitoring, eye-height and MPI measurement and the in-band remote channel all depend on the DSP and firmware. Credo is a DSP vendor, a module maker and workstream co-chair at once, giving it the best starting position.

• Module makers' differentiation gets redefined. Once telemetry is standardized, "we have diagnostics" stops being a selling point; the contest becomes firmware quality, threshold calibration and speed to compliance.

The spec explicitly does not depend on data rate or optical PMD and applies to 400G, 800G and 1.6T; Credo's newly announced 1.6T ZeroFlap modules are already compliant. In 2026 OCP APAC Summit | Broadcom | CPO's Third-Generation Report Card we saw the CPO camp argue reliability with field data; the pluggable camp's answer this time is to make reliability a measurable, auditable interface.

7. Risks and counterarguments

• Small sample, in-house product: 48 modules, all Credo ZeroFlap, and the attenuation test mainly validates Rx optical power, the easiest dimension to measure.

• No prediction accuracy: Asked directly in the Q&A how far ahead failures can be predicted and with what success rate, Collins said exact numbers are hard to give, only that it reveals events that were previously invisible.

• Discrete scores, programmable thresholds: Only three tiers, 80, 60 and 40, appeared in the trial; if each vendor calibrates its own thresholds, whether a 60 means the same across vendors, and whether spec 1.0 can enforce that, remains to be seen.

• Installed modules won't upgrade themselves: Existing telemetry can cover part of this, but remote harvesting and event logging need new firmware, so deployment volume depends on module refresh cycles.

• The author is also a seller: Credo is co-chair and currently the only vendor publicly declaring compliance.

8. What it means for Taiwan's supply chain

• For optical module makers, it is a new compliance ticket. Built on CMIS CDB and rate-agnostic, the spec could be required for every pluggable generation from 400G to 1.6T. For module makers serving hyperscalers, the race is how fast firmware teams can build in all six capabilities and keep them consistent across platforms.

• For production lines and test, it adds new test items. Segmented health scores, event logs and remote harvesting all have to be verified in factory and system validation; the "link up but only 60 points" links in the trial are exactly what traditional BER and link-status testing would pass.

• Neither the slides nor the transcript named any Taiwanese company in the workstream; whether Taiwanese vendors join can be tracked through future compliance lists. This matches the view in ECOC 2026 | No Technology Revolution in the Next Decade, Only a Volume War: as specs converge, the winners are decided by production yield and delivery speed.

9. Summary

This talk brought no new physics and no pJ/bit figures, but it tackles the most underrated line item in AI cluster ROI: time to fault isolation. Segmented scoring gives problems a location, worst-metric scoring keeps them from being averaged away, remote harvesting ends the bare-metal blind spot, and event logs preserve evidence of self-healing flaps.

Separate the timelines. Spec 1.0 is published and Credo's 400G to 1.6T ZeroFlap modules are compliant today; but turning a spec into a procurement requirement will take at least one product generation. Three things to watch over the next 12 months: a second module or DSP vendor declaring compliance, native support in mainstream switch NOSes, and hyperscalers writing it into RFQs.

Once telemetry becomes a standard, optics competition shifts from "will it fail" to "when it fails, can it explain why".

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