OCP Global Summit 2026 | Amphenol | Brian Kirk | Co-Designing High-Speed Copper and Optical Interconnects

1. Why Amphenol Is the One Saying This

At the same OCP Global Summit, Credo took the stage to argue that optics must beat copper on reliability, power and cost all at once. Amphenol's Brian Kirk offered a different version: copper versus optics has never been a black-and-white choice.

The speaker runs both copper and optics for Amphenol's high-speed products and has done this for thirty years. He offered one comparison: Amphenol ships about 40 million differential pairs per day, while the industry's annual volume of 800G/1.6T optical transceivers is on the order of 40 million. In other words, copper still dwarfs optics in volume, and the question was never which one replaces the other, but whether capacity can be shifted between the two.

We covered Amphenol's financials in Earnings Highlights: Amphenol (APH) | Q2 FY2026. What makes this talk more interesting is the technical stance: it is not selling "copper is better", it is selling "design both together".

Optics and copper are not replacing each other; they are competing for space in the same link budget.

2. Breaking Down the Copper Link Budget: The Bottleneck Is the PCB, Not the Cable

The first signal is stated plainly on the slide: beyond 102.4T, faceplate I/O density gates scaling more than the switch ASIC itself. 102.4T is 64 x 1.6T OSFP; 204.8T needs 128 OSFPs, or 16 x 12.8T XPO.

When people see the big cable cartridge at the back of a rack, the instinct is to assume that is copper's limit. Amphenol used a real 224G PAM4 link (53 GHz Nyquist, twinax at 14 dB/m) to show otherwise:

Path	Today: via host PCB	Launch at the package
Switch side	Package + BGA 8 dB, host PCB 10 dB	Package 3 dB, package launch 1 dB
Connectors	RAF 1.5 dB x 2	RAF 1.5 dB x 2
Cable	Cartridge 0.35 m, 6.2 dB	Twinax 0.8 m, 11 dB
GPU side	Host PCB 4 dB, package + BGA 3 dB	Package launch 1 dB, package 3 dB
Total / remaining	34.2 dB, 5.8 dB left	22 dB, 18 dB left

▲ The 40 dB budget of a 224G copper link: routing through the host PCB uses 34.2 dB and the cartridge only 6.2 dB; breaking out from the package uses just 22 dB, leaving 18 dB to trade for 3.8 m of reach (Image source: Simple Tech Trend original, data from Amphenol's 2026 OCP Global Summit presentation)

Put simply: remove the host PCB and the whole budget goes back to the cable. The remaining 18 dB, spent on a next-generation dielectric at 6 dB/m, translates into 3.8 m. Amphenol's conclusion is that copper can keep linking two racks, and only the rest needs to go to optics.

This is also why Kirk says "almost all of our customers" are looking at co-packaged copper. Amphenol also showed a 448G front-I/O feasibility measurement: the electrical path stays relatively linear up to 110 GHz, covering the 87 GHz Nyquist of 448G PAM6.

3. Breaking Down the Optical Link Budget: Linear Optics Is Defined by the Electrical Channel Too

The line in this talk that optical vendors should note most is: linear ports, copper or optical, are defined by the electrical channel performance.

Amphenol splits ports into four types: fully retimed Tx/Rx offers the best interoperability; half-retimed (RTLR, LRO) is a compromise; non-retimed LPO/NPO/CPO offers the lowest optical power and latency; passive copper offers the lowest cost, power and latency. The last two both remove the DSP, which pushes the link's loss pressure back onto the electrical side.

Comparing NPO with front-panel XPO:

• NPO: I/O connector 2 dB + module PCB 7 dB, an electrical channel of 9 dB in total.

• NPC/CPC to XPO: adds a 2 dB CPC connector and 8 dB of 30 cm twinax, 19 dB in total.

▲ NPO electrical channel 9 dB versus front-panel XPO 19 dB; reliability and serviceability favor pluggables, while the electrical channel and bandwidth density favor NPO (Image source: Amphenol | 2026 OCP Global Summit presentation (06:45))

Amphenol's verdict is that "RAS (reliability, availability, serviceability) favors pluggables, but the gap is shrinking". Another set of numbers is even more direct: in a 64-lane XPO-LPO development (3 nm switch host, 15 dB host plus cable), a single lane runs at 1e-10 to 1e-12 BER, but with all 64 lanes on it drops to about 1e-7, and the gap is crosstalk. Improving cables and connectors recovers 0.8 to 1.5 orders of magnitude of BER, more than improving Tx- or Rx-side package routing.

4. CPO's Other Budget: Five Times the Connections of Copper Inside the Box

CPO discussions usually stop at the optical engine and laser reliability. Amphenol opened up a physical CPO switch mock-up: internal and external lasers (ELSFP), integrated FAUs, midboard MMC (one set for single-mode and one for polarization-maintaining fiber), plus FSC144 and MMC front panels. Kirk estimates that inside an optical box there are, in many cases, five times as many interconnects as in a straight copper solution.

▲ A physical CPO switch prototype: internal and external lasers, integrated FAU, midplane MMC, ELSFP cages and front panel, with far more connection points than a copper solution (Image source: Amphenol | 2026 OCP Global Summit presentation (08:25))

The optical budget has to be counted piece by piece too. Using the OCI MSA 500 m single-mode link model, the end-to-end optical budget is 2.5 dB; between two NPOs there are six to eight separable connections at about 0.3 dB each, so six of them consume 1.8 dB and leave only 0.25 dB for the fiber. Connector choice is decisive: V-groove 0.1 dB, MPO 0.35 dB, expanded beam (EBO) 0.7 dB.

▲ NPO-to-NPO optical link: six separable connection points take 1.8 dB of the 2.5 dB budget, so connector choice decides whether 12,288 fiber lanes can close (Image source: Amphenol | 2026 OCP Global Summit presentation (13:33))

That is also why XPO was proposed: no external laser, far fewer optical interconnects inside the box, and a balance between density and serviceability. We discussed the same direction in ECOC2026 | Ciena's Open CPX Turns the 6.4T Optical Engine Back into a Shelf Part: CPO's challenge increasingly looks like a connector and assembly problem, not just an optical engine problem.

5. 256 GPUs, One Hop: Copper Inside the Rack, Fiber Between Racks

Amphenol used a concrete architecture to put both budgets together: 256 GPUs with 64 lanes each, connected to 64 switch ASICs with 256 lanes each, 16,384 lanes in total, with a single hop between any two GPUs.

Split into four racks, each with 64 GPUs and 16 switch ASICs:

• Each GPU: 16 copper lanes to its own rack, 48 fiber lanes to the other three racks.

• Each switch ASIC: 64 copper lanes and 192 fiber lanes.

• Whole system: 4,096 copper lanes and 12,288 fiber lanes, 25% copper, 75% fiber.

▲ A 256-GPU single-hop network: 4,096 in-rack lanes on copper and 12,288 cross-rack lanes on optics, where the optical link's six connection points consume 1.8 dB of the 2.5 dB budget (Image source: Simple Tech Trend original, data from Amphenol's 2026 OCP Global Summit presentation)

If copper reach really extends to two racks, Kirk says a 50/50 split is also possible. This matches the view in After Copper Runs Out for AI: Seven Paths for Scale-Up Optical Interconnect: copper will hold the rack, and optics fights beyond it.

Amphenol also laid out a clear timeline. It wants OCP to drive a reference definition for passive interconnect from package to faceplate to backplane:

• Q4 2026: scope and reference channels

• 1H 2027: first contributed measurements and test-point proposal

• 1Q 2028: specification consensus and first deployment

6. Risks and Counterarguments

• The budget numbers are Amphenol's reference values: the 40 dB total budget and the 6 dB/m next-gen dielectric are its design assumptions, and 3.8 m only counts once the new cable is in volume production.

• Longer copper means harder-working SerDes: removing the DSP from passive copper assumes the SerDes can handle a channel of 22 dB or more; Credo attacks precisely on SerDes power, and Amphenol did not address that power.

• 448G is still at the feasibility stage: the 110 GHz measurement is on initial parts, PAM4 versus PAM6 has not converged, and Kirk himself said many backplane customers are still asking about PAM6.

• There is no consensus on blind-mate optical backplanes: MPO, MMC and self-cleaning ferrules each have supporters, and float and cleanliness specs are undefined, which is exactly Amphenol's fourth question to the community.

• The standards timeline is only a proposal: a 1Q 2028 spec consensus needs switch, GPU and module vendors to sign on, and so far only Amphenol is driving it.

7. What It Means for Taiwan's Supply Chain

• The loss budget is moving from the PCB to cables and connectors. With a package launch, the host PCB goes from a 14 dB line item to a segment that can be skipped. For Taiwan's cable assembly houses, twinax flyovers, RAF backplane harnesses and co-packaged copper assembly are incremental business; for high-end host boards, this is a long-term variable worth tracking, at least on scale-up high-speed channels.

• Fiber lanes outnumber copper lanes three to one. With 12,288 cross-rack fiber lanes and six or more separable connections on each, demand for connectors and fiber arrays will grow faster than for modules themselves. Taiwan remains mostly in contract manufacturing and assembly for high-fiber-count blind-mate connectors and FAUs, while spec-setting power sits with connector majors such as Amphenol and Senko.

• External lasers are still in the CPO mock-up. Amphenol's CPO mock-up carries both internal lasers and an ELSFP external-laser cage, meaning both laser architectures will coexist in the near term, an open window for Taiwan's laser and light-source module suppliers.

• The most important move is to join the OCP workstream. Reference channels are due in Q4 2026 and measurements in 1H 2027; if Taiwanese vendors want to shape the test-point definitions, the window is the next two quarters.

8. Conclusion

Credo and Amphenol told two opposite stories at the same conference. Credo says optics must win on all three of copper's home turfs; Amphenol says copper's limit is not the cable but the host board, and with a package launch copper still has 3 m of room, enough to hold the rack or even two racks.

The deeper message is: once the DSP is removed, copper and optics become the same problem. LPO, NPO and passive copper are all defined by the electrical channel, while about 70% of the optical budget between NPOs is eaten by connections. Whoever defines the passive interconnect from package to backplane holds the spec power for next-generation scale-up. Three things to watch over the next 18 months: the content of the Q4 2026 reference channels, when the 6 dB/m next-gen dielectric reaches volume, and whether blind-mate optical backplanes settle on MPO or MMC.

The copper-versus-optics war has no loser, except whoever does not understand the link budget.

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