OCP Global Summit 4/6 | The Bar for Optics Shifts from pJ/bit to Recovery Time
OCP Global Summit paid series 4/6: optical link reliability — FIT, link flaps, fault domains and recovery time, and whose GPU jobs survive a link going down
Free section
Part 1/6 took apart Meta's 51.2T CPO field data, and 3/6 took apart the failure root causes of laser modules; both asked whether optics break. This piece asks: once the optics do break, does the job on the GPUs stop?
When people talk about optics entering AI clusters, the market mostly compares pJ/bit and FIT. For progress on CPO component reliability, see 2026 OCP APAC Summit | Broadcom | CPO's Third-Generation Report Card. A month later at the 2026 OCP Global Summit, the people talking about reliability were on the system side.
First, Microsoft swept a table: at a per-lane Flap-FIT of 10,000, a 1,024-GPU optical pod flaps on average once every 0.76 hours; even at 1,000 it lasts only 7.63 hours. The table notes these are not field data.
Second, the OCP SROI workstream set reliability targets for scale-up optics: at most 2 FIT per link end and zero link flaps per million device-hours, against 50 to 100 FIT for pluggables today.
Third, a 75,000-GPU OpenAI training job hit a T1 switch that silently stopped forwarding; about 25% of connections were affected and about 580,000 packets were dropped, and the speaker said the job did not stop.
STT's view: the bar for optics is shifting from pJ/bit to recovery time.
▲ Microsoft's sweep of per-lane Flap-FIT against pod size gives the mean time between flaps: at 1,024 GPUs and a Flap-FIT of 10,000, about one every 0.76 hours (the table is marked as not field data) (Image source: Simple Tech Trend original; data from Microsoft 2026 OCP Global Summit presentation)
Where copper and optics divide in scale-up: 2026 OCP APAC Summit | SemiAnalysis | Scale-Up Sophistry; high-temperature lifetime of VCSEL arrays: ECOC 2026 | 850nm Dies at 1,600 Hours, 1060nm Hits 5,000 Hours with Zero Failures.
In the paid section
Scale breakdown: four sets of scale numbers on one ruler, and how much flap budget a 1,000-GPU pod has left
Component breakdown: reliability numbers for LPO, CPO and copper, and why pre-FEC BER alone cannot predict link health
Flap breakdown: seven classes of burst-error sources, and what NPO removes, keeps and adds
Fault-domain breakdown: why the same failure can take out 100% or one-eighth of the bandwidth
Recovery-time breakdown: recovery times from hitless switching to checkpoint rollback, side by side
Telemetry breakdown: what OCP's first cross-vendor telemetry spec covers and what it misses
Testimony from the other side: what the companies selling CPO, copper and connectors each concede
Differences from our ECOC calls: one partial contradiction, two revisions, one reinforcement
Taiwan supply chain: Taiwanese names on the slides, where they sit, and the two real bottlenecks
Counterarguments and eight signals to watch
──────────────────────────────
That is the summary of this piece.
STT's full analysis — what FIT, link flaps, fault domains and recovery time each decide, and four differences from our ECOC series — is in the paid section, with 6 original charts.
👉 Subscribe to the STT paid section to read the full version:
──────────────────────────────
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments