OCP Global Summit 2026 | Luxshare-Tech | Andrew Kim, Jinhua Chen | Copper Interconnect Innovations for Realizing 400G Electrical

3 minutes ago

7 min read

1. Why this talk matters: the bottleneck moved from the chip edge to outside the box

Over the past year, the 448G electrical debate has centered almost entirely on the chip edge: BGA breakout, package routing and PCB loss, with CPC seen as the shortcut around all three. As we wrote in ECOC2026: Fast-and-Narrow vs. Slow-and-Wide, and the Dividing Line Is 448G, 448G is where copper and optics divide the work.

This 16-minute technical talk skipped the vision and showed measurement curves, but the conclusion was clear: Luxshare argues that twinax is the body of the 448G channel and connectors are the main limit, and it splits connectors into two classes: chip-to-twinax (CPC/NPC or an active front-panel module) and twinax-to-twinax (front-panel DAC and cable backplane).

Good twinax makes up about 99% of the link length; what is left to solve sits entirely at the junctions.

Kim put it bluntly: if any single connector has a resonance somewhere around 80 GHz, the whole link is lost. Every component has to be clean.

Why keep pushing copper? Luxshare's reasons are latency and power; the speaker called copper's advantage on both "huge", and the slides state that copper is needed for shorter link segments to keep power down (STT note: passive copper needs no DSP or laser). It closed by asking connector vendors to give mechanical drawings and SI characteristics to the OCP server project so that 400G (and 256G) per-lane links can be adopted directly by system builders. In other words, this is a race to get your own channel model into the reference design before the spec is final.

▲ Luxshare-Tech splits the 448G copper channel into two kinds of connectors: the chip-side CPC and the front-panel/backplane connectors outside the chassis (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation)

2. Breaking down the channel: the twinax itself is not the problem

The slides show Optamax 448 31 AWG, 85-ohm twinax whose measurement (2 m minus 1 m de-embedded) and simulation overlap all the way to 110 GHz; by STT's reading of the chart, loss at 90 GHz is about 17–18 dB per metre. Luxshare argues it beats PCB for two reasons:

Lower loss per unit length
Fewer impedance discontinuities: PCB trace bends and via layer changes create impedance discontinuities and mode conversion; cable has none of these

This is not a PCB-friendly argument: at 448G, PCB should only cover the few centimetres next to the chip, and long runs go to cable.

The cable comes in two gauges: thinner, easier-to-bend 31 AWG inside the box, and thicker, lower-loss 27 AWG outside the box or on the backplane. Where the gauge changes, you need a twinax-to-twinax connector, the subject of Section 4. In other words, the more the cable is optimized segment by segment, the more junctions there are, and every extra junction is one more place a resonance can appear.

3. The CPC is approaching the PAM4 threshold

Koolio is Luxshare's elastomeric CPC. It was shown at 200G at OCP in 2024 and at 400G for the first time in 2025; this year's improvements are all inside the connector, so customer boards do not change. Key specs:

Stubless interface: the elastomer is a silicone base embedded with a metal pattern that aligns the connector to the substrate footprint
360-degree shielding, with a differential-pair pitch of 1.9 mm × 1.7 mm
A 300 mm CPC-to-CPC cable assembly shows smooth insertion loss up to the 110 GHz VNA limit with no resonance; return loss is generally below -10 dB
Five re-matings, with a min-max spread of 0.51 dB
Crosstalk simulation (new 2026Q3 design): peak pushed from 110 GHz to 120 GHz and about 15 dB lower

▲ The Koolio elastomer CPC: stubless interface, 360-degree shielding and a 1.9 mm × 1.7 mm differential pair pitch (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation)

Two caveats: crosstalk is simulated, and the speaker said board limitations prevented measurement this time; measurement stops at 110 GHz, just below PAM4's 112 GHz, and the resonance-free 120+ GHz is a slide claim, not yet measured. But given that insertion loss, ILD and return loss within 110 GHz are all on the record, the CPC segment is most likely no longer the bottleneck for 448G PAM4 (STT reading).

4. The new bottleneck: twinax-to-twinax mating outside the box

The real novelty is the twinax-to-twinax mating structure, promoted for only about four months, with the backplane prototype built only a month ago. Two demos:

Demo	Channel	Total twinax	Slide conclusion
Cable backplane	330 mm 31AWG + VH-VR connector + 940 mm 27AWG + VR-VH connector + 330 mm 31AWG	1,600 mm	Clean to 100 GHz, resonance only at 110 GHz
Front-panel LR	Koolio CPC + 240 mm 31AWG + front-panel connector + 525 mm 27AWG + front-panel connector + 240 mm 31AWG + Koolio CPC	1,005 mm	Clean to 100 GHz, no major resonance; preliminary sample

First, what "clean" means: both demo slides show only insertion-loss (SDD21) curves, and "clean" means a smooth curve with no obvious resonance dip; no return loss, ILD or crosstalk data was shown for the external connectors, and these are whole-link measurements, not the connector alone. Put each segment's clean bandwidth next to the Nyquist frequencies of the two modulations, and the answer is plain:

▲ Clean bandwidth of each segment of a 448G copper channel: twinax and CPC measured resonance-free to 110 GHz (CPC claimed above 120 GHz), while links with external connectors stop at 100 GHz, enough for PAM6 but not PAM4 (Image source: Simple Tech Trend original; data from Luxshare-Tech 2026 OCP Global Summit presentation)

Kim's spoken account had two layers: the backplane link is "definitely clean out to 90 GHz" and "a little turbulent past 90", though he still called it clean to 110 GHz; the front panel is "clean out to at least 90, 100 GHz". That is why the slide concludes PAM6 viable in the near term, PAM4 within reach, and not the other way around. In the 400G/lane module connector panel at the same summit, Luxshare-Tech's two-piece interface for OSFP-size modules simulated at about 111 GHz, likewise with almost no PAM4 headroom, pointing the same way.

How far is 12 GHz? Against PAM4's 112 GHz, it is roughly a tenth of the bandwidth. For reference, Koolio went from its first 400G showing in 2025 to resonance-free measurement at 110 GHz this year in one year, changing only the connector internals. Kim also said the first-generation prototype "bodes well for what this is going to look like a year from now". STT's timeline breakdown: the external connector needs at least one more generation to reach PAM4, so the answer will come around OCP 2027.

5. One mating structure, three products

The mating structure is not a new design; it is ported from Luxshare's existing backplane right-angle connectors:

Cable backplane: a vertical header (cartridge) mated to a vertical receptacle (to the CPC); the first 448G samples are built
Front-panel DAC: the plug carries 18 A power contacts and 4 low-speed pins, and the speaker said it fits the same lane density as an OSFP connector
Paddle card: the same structure built into the plug's paddle card serves AOC, AEC and ACC

▲ Front-panel LR demo link: two Koolio CPCs, two front-panel connectors and 525 mm of 27AWG external twinax, clean to 100 GHz (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation)

▲ Composition and length of the two 448G demo links: 1.6 m for the backplane and 1.0 m for the front panel, with external segments of only 940 mm and 525 mm (Image source: Simple Tech Trend original; data from Luxshare-Tech 2026 OCP Global Summit presentation)

What this strategy means: beyond the CPC, passive copper, active copper and optical cables share one outside-the-box interface. STT reading: whichever mating structure is first adopted in the OCP server project's reference design gains the early lead on 448G external junctions.

6. Risks and counterarguments

"Clean" is not a spec. The slides use a qualitative description, with no insertion-loss, ILD or crosstalk margin against an OIF or IEEE mask. The loss budget itself is the 448G challenge; the speaker also said that "even when you have clean channels, they're quite lossy at these high frequencies".
Sample stage. The front-panel connector is a preliminary sample with the process still being refined for tolerance; the backplane prototype is a month old. There is no yield or consistency data yet.
Limited CPC re-mating data: five re-matings is a quick lab check, not a production lifetime.
PAM4 could arrive first. If a second-generation external connector gains 12 GHz or more in one step, PAM4 copper could arrive alongside PAM6, and STT's call would be wrong.
Copper's reach is shrinking. External runs beyond 1 m are hard, which pushes more links to AEC or optics; see After Copper Runs Out for AI: Seven Paths for Scale-up Optical Interconnect; the opposing case is in 2026 OCP APAC Summit: SemiAnalysis, Scale Up Sophistry.

7. What it means for Taiwan's supply chain

The talk named no Taiwanese companies; what follows is read only from publicly known industry positions.
Luxshare is a direct competitor to Taiwanese cable and connector makers. This talk showed twinax (Optamax), CPC (Koolio), backplane and front-panel connectors together, meaning one company can propose the entire passive channel. Taiwan's high-speed cable assembler BizLink and Hon Hai's Foxconn Interconnect Technology (FIT) face this "sell the whole channel" competition at 448G.
PCB's role gets squeezed toward the chip. If long runs move fully to twinax, high-end PCB value at 448G concentrates around the package and CPC rather than long traces across the whole board.
A PAM6 start is a timing signal for SerDes and test: the slides say OIF and IEEE formally started only in the past year, and connector makers are now delivering 100 GHz samples; STT estimates the volume window for 448G passive copper more likely falls after 2027.

8. Summary

The value of Luxshare's talk is not the claim that 448G copper works, but that it measured where the bottleneck sits: twinax and CPC are resonance-free to 110 GHz, approaching PAM4's 112 GHz, while the twinax-to-twinax connector outside the box stops at 100 GHz. That 12 GHz gap is why first-generation 448G passive copper picks PAM6.

STT's call: first-generation 448G passive copper will be PAM6, for in-rack and adjacent-box links with external runs under 1 m. Three things to watch: whether a second-generation external connector's clean bandwidth crosses 112 GHz, the modulation decisions in OIF CEI-448G and IEEE 400GPL, and whose mating structure the OCP server project adopts.

448G copper is won or lost not on the cable, but at that one connector outside the box.

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