ECOC 2026 | A 409.6T CPO Switch Has 4,096 Single Points of Failure: The EBO MSA Wants to Eliminate Connector Contamination by Going Contact-Free

2 days ago

9 min read

Richard Ward, co-chair of the EBO MSA, wasn't talking about a new technology at the ECOC 2026 Market Focus. Expanded Beam Optics (EBO) dates back to the 1970s and has spent fifty years in military and then heavy-industrial use. What he talked about was why this old technology has suddenly become a data-center necessity. The reasons lie in two numbers: Oracle calculates that a single 409.6 Tbps CPO switch has 4,096 potential single points of failure; Renesas calculates that for a 51.2T switch card in a DR architecture, first-pass board-level fiber assembly yield is only 23.0%. And the operational reality in the field: roughly 80% of network issues are cleaning-related, and roughly 95% of maintenance needs are for cleaning. EBO's answer is simple — no contact, no contamination.

1. A 50-year-old technology: why is it only now entering the data center?

Let's start with the principle, because it determines every benefit that follows.

EBO's mechanism is almost plain in its simplicity: light leaving the fiber is collimated by a lens into a parallel beam, crosses an air gap, and is focused back into a fiber on the other side. The example on the slide was a 3M single-mode expanded-beam ceramic ferrule: light enters from a 9 µm fiber, is collimated by a total internal reflection (TIR) mirror, and the beam expands to about 80 µm; the ferrule surface carries an anti-reflection (AR) coating to reduce loss and back-reflection; the receiving ferrule on the other side uses the same mirror to focus the beam back into the fiber.

It looks like a small polygonal block of glass, with light emerging from the middle.

The technology emerged in the 1970s for harsh military environments — high resistance to contamination was its original selling point. Over the following decades it moved into heavy industry. But it never made it into the data center, because the form factor and the cost were wrong.

Now things have changed — and what changed isn't EBO, it's the data center.

2. Problem one: 4,096 single points of failure

This slide comes from Oracle's perspective, and the math is straightforward:

A 409.6 Tbps CPO switch, configured as 512 ports of 800G DR4, with each port using an 8-fiber physical-contact connector. 512 × 8 = 4,096 potential box-level single points of failure.

And the key line: a single contaminated physical-contact connector can get the entire switch RMA'd. This is a common failure mode in data centers — transceivers or fibers being replaced frequently because of end-face damage.

The slide also showed three micrographs of contaminated end faces: dust and debris, oily fingerprints, and deep scratches.

The conclusion was blunt: expanded-beam connector technology eliminates this failure mode because it makes no physical contact at all, and is therefore also immune to dust, debris and other contaminants.

Oracle's 4,096 single-point-of-failure calculation, the 80%/95% cleaning share of data-center maintenance, and Renesas' 51.2T switch-card yield comparison (DR 23.0% vs FR4 67.8%). Source: Simple Tech Trend | Data: Richard Ward, EBO MSA — ECOC 2026 Market Focus

3. Problem two: the yield math no longer adds up

The second slide comes from Jeff Hutchins of Renesas. The speaker said he particularly likes this one because it takes a system-level view.

Chassis used to be "electrical on the board, optical modules plugged into the front panel": if fibers got dirty you could clean them, and you could clean the optics. But as NPO spreads, optical connectors will move inside the chassis — and in that position, you can't reach them to clean.

Then comes the sobering table. Take a 51.2T Ethernet switch card with 100G lanes as an example:

Tx fiber connections: DR (PSM) 512; FR4 (CWDM) 128
Rx fiber connections: DR (PSM) 512; FR4 (CWDM) 128
Laser PIC fiber connections: DR (PSM) 64; FR4 (CWDM) 32
Total fiber connections: DR (PSM) 1,088; FR4 (CWDM) 288
Assumed 3σ first-pass fiber connection yield: DR (PSM) 99.865%; FR4 (CWDM) 99.865%
First-pass board-level fiber assembly yield: DR (PSM) 23.0%; FR4 (CWDM) 67.8%

A single-connection yield of 99.865% sounds perfect, but compounded across 1,088 connections it drops to 23%. That is the unforgiving math of a system with this many links.

The same analysis gave another key number: to keep the rework rate under 2% in the 51.2T FR4 scenario, the first-mate insertion-loss failure probability of every fiber connection must be as low as roughly 7×10⁻⁵. In the slide's own words: the acceptable rework rate determines the acceptable first-mate insertion-loss limit.

Here the speaker offered a measured disagreement of his own: the analysis is well done, and includes tail analysis for both beginning-of-life and end-of-life, but it assumes standard (physical-contact) connectors — and he believes EBO rewrites the whole calculation.

4. The reality in the field: 80% of problems are really just dirt

This slide has only two lines of text, but it is the underlying motivation for the whole talk:

About 80% of network issues are cleaning-related
About 95% of maintenance needs are for cleaning

Then the yellow conclusion banner: systems that have adopted EBO have demonstrated roughly an 85% reduction in installation and cleaning time.

The speaker also raised a rarely discussed risk. WDM optical power has now reached the 25 dBm class; if a single strand of lint lands on a fiber end, it is exposed to more than 300 mW of power — at that point it's no longer a cleaning problem, it's a burn-out. Because EBO's beam stays inside the housing and the end face is never exposed, it is immune to this kind of lint.

The same logic yields two more benefits: improved eye safety (the beam is fully enclosed in the housing during operation), and no end face that an operator can touch — and a large share of the industry's problems are precisely operator error and contamination.

5. The data: how good can EBO actually get?

This is the slide most worth citing from the entire talk, because it isn't a claim — it's measured.

Insertion loss (random-mate testing): 552 connector pairs, 8,832 data points. The result: 97% of channels (IEC basis) below 0.55 dB; 99.2% of channels below 0.7 dB; mean insertion loss 0.32 dB.

The speaker's way of reading the data is worth learning: "I actually don't care about the average." What he cares about is the tail of the distribution — the earlier slide was about 4,096 single points of failure, and here, across 8,832 tests, the tail is very short. When your system has thousands of connection points, success is never decided by the average; it's decided by the 99th percentile.

Return loss: 16-fiber single-mode, 9,600 channels. Mean return loss 66.7 dB, 99% of channels better than 55 dB.

Repeated mating: 12-fiber EBO connector, 1,000 re-matings with no cleaning at all, insertion-loss change less than ±0.1 dB.

Alignment tolerance: with a ferrule-to-ferrule offset of ±10 µm in the x direction, insertion loss changes by only about 0.35 dB. The comparison is brutal — a 9 µm fiber offset by 9 µm is effectively a complete disconnect.

Mechanical: under standard vibration testing, less than 0.02 dB; and because there's no contact, mating force can be reduced by about 20×. The importance of this is easy to underestimate: only with 20× lower mating force does it become possible to build larger connectors with far more parallel channels, which is exactly what AI data centers need most right now.

There's no free lunch — the extra optics do carry an insertion-loss penalty, but the speaker's verdict was "it's actually fine."

EBO measured data: insertion loss (552 pairs / 8,832 points), return loss (9,600 channels), 1,000 cleaning-free re-matings, alignment tolerance and vibration testing. Source: Simple Tech Trend | Data: Richard Ward, EBO MSA — ECOC 2026 Market Focus

6. The MSA itself: 57 members, and a roadmap Taiwanese suppliers should watch

The EBO MSA was founded in March 2026, and its purpose is stated plainly: to meet customer demand for interoperable EBO connectors and bring supply-chain stability through specifications.

Leadership: the chair is Mark Filer of Oracle; co-chair and administration is Richard Ward of 3M / Xscape Photonics; editors come from Amphenol, Sumitomo and Senko.

The membership breakdown was the most information-dense slide of the talk: 57 members in total — 9 "end users" and 48 "suppliers".

The end users are: AMD, Arista, Cisco, HPE, Meta, Microsoft, nexthop ai, NVIDIA and Oracle.

Among the suppliers, Taiwanese companies have a very strong presence.

The speaker specifically explained why members are split into two categories: the MSA has to decide which connector specs to write first, with suppliers contributing designs and priorities set by vote. Listing end users and suppliers separately ensures both sides' votes carry equal weight — suppliers can see what end users want, and end users can see what suppliers have proposed.

Connector roadmap (18 form factors currently under discussion, four of which have made the main list):

128f (8×16f): design specification in progress
16f MPO: specification design under discussion
16f VSFF: planned
ELSFP: planned

7. The one box left unchecked: high power

An audience question went straight to the sore spot: what is the plan for expanded-beam connectors on insertion loss, polarization-maintaining (PM) compatibility, and high power?

The speaker's answer was honest: there is currently no production-grade solution for high power. The samples they're handling now are at a certain power level, and "people are telling us they need more." He gave no numbers, saying only that "power keeps going up," and committed to addressing it.

This is the risk most worth remembering from the talk, because external light-source power is central to the CPO architecture. If EBO can't keep up at high power, the ELSFP line of the roadmap will stall — and ELSFP is precisely where Taiwanese suppliers have the best opportunity.

Two further caveats:

First, this is the MSA's home-turf narrative. The speaker himself said, "Don't take my word for it — I'm very biased." EBO's insertion loss is indeed higher than that of the best physical-contact connectors (mean 0.32 dB); his argument is that in a system with thousands of connection points, the tail of the distribution matters far more than the mean. I think the argument holds, but it is a system-level argument, not a component-level win.

Second, cost was not mentioned at all. Every EBO ferrule needs a lens structure and an AR coating, which cannot come free. That question will only be answered once the 18 form factors converge into a few specs and volume production begins.

8. Conclusion

Placed in the context of ECOC 2026, this talk and Corning's are really two sides of the same story: Corning says 95% of CPO reliability risk is concentrated in the FAU; the EBO MSA says serviceability risk is concentrated in connector end faces. Both point to the same conclusion — CPO's volume-production bottleneck isn't the chip, it's the interface.

For the Taiwanese supply chain, three concrete takeaways:

First, Taiwanese companies are already at the table — and not just one or two. Ten of the 48 suppliers are Taiwanese. This isn't a game of fighting to get in; it's a game of having already gotten in and now competing for a voice in the specs. And the MSA's rule is that suppliers contribute designs and priorities are set by vote — whoever submits a design proposal first has the chance to turn their own structure into the standard.

Second, the 20× reduction in mating force is the real commercial lever of this technology. No contact and dirt immunity are defensive value; but cutting mating force to one-twentieth means parallel connectors with far higher channel counts than today become possible — 128f (8×16f) is the first product of that. For Taiwanese companies making connector structural parts, alignment mechanisms and precision molds, this is a window in which specs are being rewritten, and when specs are rewritten, existing market share doesn't count. For related context, see 2026 OCP APAC Summit | SENKO | Detachable fiber interfaces: whether CPO can scale comes down to that 0.15 dB.

Third, watch the ELSFP line. Right now it only says "planned," but if EBO solves the high-power problem and an expanded-beam version of ELSFP becomes viable, the connection interface of external light-source modules will be redefined — and that is exactly the learning curve Taiwan's ELSFP supply chain (Elaser, Luxnet and others) is climbing right now. For related progress, see Earnings highlights: Elaser (3450) | 2026 Q2 — EPS tripled in one quarter, but the real signal is an external light source not yet in volume production.

Verdict: EBO isn't a new technology; it's an old technology being repriced by the scale of AI data centers. When a single chassis has 4,096 connection points and 80% of failures come from contamination, "no contact" goes from an engineering preference to an economic necessity. Only two real questions remain: can high power be solved, and can cost come down?

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