OCP APAC Summit 2026 | SENKO | Chengting Chen | Detachable Fiber Interfaces: Whether CPO Reaches Volume Comes Down to 0.15 dB

2 days ago

9 min read

For the past two years, the CPO (Co-Packaged Optics) conversation has revolved around pJ/bit, bandwidth and lasers. But what actually blocks volume production is a far less glamorous question: once a fiber is attached to the chip, can you ever take it off again? And if you do, is the loss the same when you put it back?
SENKO's answer at the OCP APAC Summit 2026 is SEAT (SENKO Elastic Averaging Technology): instead of traditional kinematic coupling, where a few contact points dictate position, dozens of compliant bumps share the error. Measured results: repeatability <0.15 dB (3σ) and interchangeability <0.5 dB (3σ).
The more important part is the second half: SENKO ties together the CudoForm metal stamping process it acquired in 2022, a PIC active alignment machine co-developed with Aerotech and Santec, and a wafer-level packaging path. This is not an optical design problem; it is a manufacturing problem — and manufacturing is exactly where Taiwan's supply chain is strongest, and where it should be most alert.

1. Why a single connector deserves a whole article

Once optics move inside the package, manufacturing difficulty doesn't rise linearly; it jumps to a different class.

Speaker Chengting Chen laid out three issues right at the start: tighter alignment tolerances, more complex assembly flows, and serviceability becoming a real problem. The whole industry talks about the first two; the third is the heart of this talk, and the piece the market most often overlooks.

In the pluggable era, if an optical engine failed you pulled it and swapped in a new one; nobody needed to discuss it. With CPO, the optical engine is sealed onto the same substrate as the ASIC, and the fiber is glued directly to the photonic chip (PIC). The question then becomes: if a CPO substrate worth tens of thousands of dollars has to be scrapped because of one fiber or one FAU (Fiber Array Unit), the business simply doesn't work financially.

We broke down this bottleneck in full in Fiber-to-chip coupling: the least glamorous step in CPO, and the one blocking everyone. SENKO's talk essentially took that article's question into the lab and put the data on the table.

2. What SEAT changes: from "three points set the position" to "twenty bumps share the load"

This is the technical turn most worth writing down from the whole session.

Traditional precision mechanisms align using kinematic coupling: the minimum number of contact points (usually three) fully constrains all six degrees of freedom, giving a unique position with no over-constraint. It is the textbook-elegant solution, but it has a fatal weakness — position is determined entirely by those few points, so any local defect, burr or tiny wear on one of them transfers 100% of the error straight onto the optical axis.

SENKO went the opposite way: elastic averaging. Rather than three points, about twenty compliant bumps are spread across the contact surface so that positioning error is "averaged out" over the whole interface. Even if two or three bumps are off, the elastic deformation of the rest absorbs the deviation, and the final position converges on the statistical mean of the whole set.

An analogy: kinematic coupling is like a tripod — stable, but raise one leg by a millimeter and the whole camera tilts. Elastic averaging is like a board resting on dozens of springs — if one spring is off, the board barely moves.

Instead of making every contact point perfectly precise, let the "not precise enough" cancel out statistically.

SEAT's physical structure is simple, just two parts: a receptacle bonded to the PIC and a cover plate bonded to the MPC; mating the two completes the alignment. SENKO's MPC (Metallic PIC Coupler) uses an integrated metal aspheric mirror to turn the optical path rather than forcing alignment via grating or edge coupling. It supports 8 to 36 channels, with 42 channels in development, and is compatible with both grating-coupler and edge-coupling PIC designs.

One easily skipped point with big commercial significance: SEAT is connector agnostic. The same receptacle/cover plate interface can take different FAU designs; the speaker's example was 42 channels, but it could also be an 84-channel dual-row. In plain terms: a customer using 42 channels today can move to 84 channels next year without requalifying the interface. For an early-stage market whose specs change every year, that matters far more than shaving another 0.05 dB of loss.

3. Two datasets are the real headline: repeatability and interchangeability

The speaker himself called this the "most important result" of the work. I agree.

First: the repeatability test. The same FAU and the same PIC assembly, mated and measured over and over, answering a very plain question — is it the same optical result every time you reconnect? The figure given verbally on stage was insertion-loss variation of 0.1 to 0.2 dB across multiple channels and cycles; the rigorous version in the official abstract is <0.15 dB (3σ).

Second, and harder: the interchangeability test. This time it's not reconnecting the same unit, but swapping in two completely different FAU cables paired with different PIC assemblies, in a mix-and-match matrix. The official figure is <0.5 dB (3σ).

You have to read the two numbers together to see the difference:

Metric: Repeatability; official figure (3σ): <0.15 dB; question answered: is the same unit the same after unplugging and replugging?; who cares: data center operations

Metric: Interchangeability; official figure (3σ): <0.5 dB; question answered: is a different unit the same?; who cares: production lines and spares inventory

Repeatability decides whether CPO can be serviced; interchangeability decides whether CPO can be mass-produced. Fail the first and data centers won't deploy it; fail the second and factories must pair every PIC with a dedicated FAU that only fits that PIC, abandoning economies of scale entirely. The 0.5 dB interchangeability figure is more than three times wider than repeatability, which honestly reflects that unit-to-unit consistency is inherently much harder than reproducibility on a single unit.

A reading trap worth flagging: the "0.1 dB" said on stage and the abstract's "<0.15 dB (3σ)" are not the same thing. The former is typical observed variation; the latter is a statistical upper bound including three standard deviations. When reading a datasheet, a loss figure without a σ attached has limited value.

4. It's not just selling a component — it's selling an alignment machine

This is, in my view, the most underrated part of the talk.

Since 2024, SENKO has been working with Aerotech (precision motion control) and Santec (optical test and measurement) on a PIC active alignment machine. Three key innovations:

Synchronized six-axis active alignment: instead of scanning axis by axis, all six degrees of freedom are searched simultaneously, cutting alignment time.
Parallel signal monitoring: a loopback architecture monitors multiple signal paths at once rather than measuring them one by one.
Customizable objective function: customers can weight different loopback paths themselves — for example, sacrificing a weak path to optimize the whole — instead of treating every channel equally.

The third point is the machine's real differentiator. Aligning a multi-channel FAU is fundamentally a multi-objective optimization problem — tune channel 1 to its best and channel 16 may drop. Letting customers define what "best" means effectively hands the definition of yield back to the production line.

The speaker also walked through the full process sequence: incoming inspection, dry alignment, dispensing, wet alignment, UV cure, post-cure bake, post-thermal inspection, and finally the repeated-mating repeatability test. The most troublesome step is cure drift: adhesive shrinks as it cures and, combined with thermal stress, the position shifts — worst and most time-consuming during the bake. SENKO validated across different PICs using 7 loopbacks precisely to show this drift is controllable.

A connector company starting to sell an "alignment process" is a more important signal than the connector itself. Early CPO know-how lives not in the parts but in the process parameters — whoever holds the process window holds pricing power in the next round. It's the same overall trend we observed in Can the fiber be swapped? The most underrated hurdle in CPO volume production, and the fifteen companies racing for it.

5. Stamping and wafer-level: turning precision optics into scale manufacturing

In July 2022 SENKO acquired US-based CudoForm — a maker of high-precision metal micro-optics whose core capabilities are metal stamping plus beam shaping and steering.

At the time the deal didn't obviously make sense; four years on, the roadmap is very clear: move precision optical alignment — a craft that used to require active alignment and unit-by-unit fine tuning — toward stamped parts that are accurate by construction. Stamping has extremely low marginal cost and its consistency is set by the tooling, which is exactly what elastic averaging needs: dozens of bumps that are statistically consistent, rather than each one being perfect.

One level up is the wafer-level packaging path. At ECOC 2025 SENKO and GlobalFoundries already demonstrated it together: GF etches trenches, a broadband spot size converter and embedded microlenses on the wafer, and the SEAT receptacle is bonded directly onto the photonic die, achieving passive, repeatable alignment and supporting both wafer-level and die-level test.

Connect the three and SENKO's full narrative is: stamped parts deliver cost and consistency, elastic averaging tolerates the remaining error, and wafer-level processing pushes alignment upstream into the fab. If this path works, competition in the FAU segment shifts from "whose alignment equipment is better" to "whose tooling and wafer process are more stable" — two completely different kinds of companies.

6. What the speaker admits is still unsolved

What made this talk unusual is that the speaker also covered what hasn't been achieved yet.

All results shown so far are die-level fiber attach. The next step, module-level attach — attaching fiber to a module that has already gone through advanced packaging — the speaker flatly called "more challenging," listing four specific obstacles:

Smaller die-to-die gaps, compressing the working space
Restricted UV curing angles: the adhesive can be dispensed, but the light can't reach it
Harder dispensing, with less space and tighter tolerances
Larger substrates, amplifying accumulated error from warpage and CTE mismatch

These four are really the old problems of advanced packaging, now colliding with optical-grade tolerances. It echoes what we argued in Why CPO is an advanced packaging war: CPO's hardest problems keep migrating from the optics side to the packaging side.

As for the Q&A question "how far can this detachable connector scale in volume?", the speaker's answer was quite conservative: based on current repeatability and interchangeability data, SEAT has a shot at becoming a standard interface and so he believes it can reach volume production — but he gave no capacity figures, yield figures or customer names. In the same session he also acknowledged that "overall CPO volume isn't ramping that fast yet," so the near-term goal is getting the alignment machine to market faster and shortening assembly cycle time.

Translation: technical feasibility is proven; commercial ramp hasn't started. That is the most honest description of where the entire CPO supply chain stands today.

7. What it means for Taiwan's supply chain

Three concrete signals.

First, the competitive yardstick for the FAU segment is being rewritten. The old contest was who had the lowest insertion loss; the next one is about three things: is it the same after a hundred matings (repeatability), is it the same with a different unit (interchangeability), and can specs change without requalifying the interface (connector agnostic). For Taiwan's FAU and optical coupling players — FOCI, Largan, AuthenX and peers — the real target isn't another decimal place of single-point loss, but whether they can produce 3σ statistical data and dare to put it in a datasheet. We previously broke down FOCI's six core FAU patents and AuthenX widening alignment tolerance to ±18 µm; the two take different approaches but are tackling the same problem.

Second, this is a manufacturing problem, not a design problem — good news and bad news for Taiwan. The good news: tooling, stamping, precision assembly and process-window control are already Taiwan's supply-chain strengths. The bad news: SENKO has already filled that gap with one acquisition (CudoForm), and did so back in 2022 — making this call three to four years ahead of most Taiwanese peers.

Third, the active alignment equipment segment is being locked up by international partnerships. Aerotech plus Santec plus SENKO effectively combines motion control, optical metrology and connectors to package the process window into a deliverable machine. If Taiwan has no equivalent integrated equipment solution on this line, purchasing power over key equipment for future CPO lines will sit in someone else's hands — a long-term impact far more serious than losing a few FAU orders.

Conclusion

On the surface SENKO's talk was about a connector; in reality it answered the most practical question in CPO commercialization: once optical components are sealed into the substrate, how do you rewrite the maintenance logic of an entire data center?

Its answer is to change the positioning philosophy from "pursue ultimate precision" to "tolerate error and cancel it statistically," then use stamping and wafer-level processing to turn that philosophy into a scalable manufacturing flow. The two figures — <0.15 dB (3σ) repeatability and <0.5 dB (3σ) interchangeability — have, for now, moved "detachable fiber interfaces" off the list of technical risks.

But be clear about what it hasn't answered yet: module-level attach is unsolved, no volume ramp curve was given, and no customers have been disclosed. Today's SEAT is a solution that has passed lab validation but not yet factory validation.

If I had to pick one indicator to track, I'd watch whether SEAT gets written into the specification documents of any standards body or major customer. Whether an interface spec wins ultimately depends not on its loss figures but on how many people are willing to design their own products around it. Until that happens, this remains an excellent technical showcase from a connector maker, not an industry inflection point.

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