OCP Global Summit 2026 | Samtec | Tom Palkert | Open CPX Lives or Dies on a Sub-1 dB Socket

5 hours ago

7 min read

1. Why Listen to a Connector Maker on CPO

We have already covered Open CPX from two angles. At ECOC last month, Ciena framed it as an open ecosystem: optical engines turned from custom parts back into multi-vendor standard components, see ECOC 2026 | Ciena Uses Open CPX to Turn the 6.4T Optical Engine Back into a Shelf Part. At OCP this week, Palkert himself noted that the CPX MSA had come up in many talks.

Samtec's talk was different. Palkert is a system engineer, and he said on stage that he "spends a lot of time looking at what that channel looks like." The slide lists the collaborators as Samtec, Ciena, Coherent, TeraHop, Marvell and Molex: system, module and connector vendors at one table.

His positioning is blunt: CPX is not pluggable, and not strictly integrated optics either; it sits in between. The same socket can go on the package, next to the package, or on the main board.

Optical engines talk in pJ per bit; connector makers talk in dB per channel segment. The second decides whether the first holds.

2. Breaking Down the Loss Budget: 15 dB vs 32 dB

Whether a linear (no-DSP) interface works depends on total electrical channel loss. Samtec splits the CPX channel into three segments, with these limits on the slide:

Segment	Test points	Max loss
Host die to socket	TP0d to TP1	Below 15 dB
Module channel	Socket to module die	Below 3 dB
Of which: socket plus connector	Mated	Below 1 dB
Of which: connector to module die	Module substrate	Below 2 dB
Die to die total	TP0d to TP1d	Below 18 dB

The comparison is a front-panel pluggable, whose host channel is about 32 dB. CPX cuts the host channel to 15 dB, saving about 17 dB (STT calculation), by removing three things: BGA escape, the front-panel pluggable connector, and AC coupling capacitors. Effective return loss (ERL) also improves from 8.5 dB for pluggables to 10 dB.

▲ A pluggable host channel is about 32 dB; CPX die-to-die is under 18 dB, of which socket plus connector account for under 1 dB (Image source: Simple Tech Trend original; data from Samtec, 2026 OCP Global Summit presentation)

The ratio is what matters most here: socket plus connector take less than 6% of the 18 dB budget (STT calculation). The connector maker's job in this architecture is to be nearly invisible in the loss budget.

▲ Samtec's CPX channel loss figures: host channel of 15 dB versus 32 dB for pluggables, and ERL of 10 dB versus 8.5 dB (Image source: Samtec | 2026 OCP Global Summit presentation)

3. Breaking Down Power: Where the 65% Comes From

Only a budget that holds delivers the power dividend. The slide cites an OFC 2026 paper (M4B.6): CPO/NPO saves 65% versus DSP pluggables. By technology:

Silicon photonics: 3.8 pJ/bit
VCSEL: 1 to 1.9 pJ/bit; Palkert added verbally that one vendor that day claimed 1 pJ/bit for a VCSEL solution

The chain is clear: short channel → low loss → linear interface works → DSP removed → power saved. Conversely, if any segment exceeds its loss budget, the DSP comes back and the 65% no longer holds.

Samtec also admitted a lesson: putting the socket directly on the package gives great electrical performance but not enough bandwidth density. That led to an NPO substrate design that fans out on a lower-loss, cheaper material than the package substrate, enlarging the beachfront and improving cooling while keeping the linear interface within 15 to 18 dB. The substrate drawing comes from Broadcom's OCP 2025 presentation, and the signal integrity curves from Broadcom's ECOC 2026 measurements.

4. Breaking Down Reliability: Borrowed Data and Homework Still Due

The slide offers three pieces of reliability evidence:

Meta (OCP 2025): 15M CPO device-hours with MTBF above 15M hours for unserviceable failures; 2M device-hours for 2×400G pluggables with MTBF of 0.55M hours. Numerically that is more than a 27x gap (STT calculation: 15 / 0.55; the CPO figure counts only unserviceable failures, so definitions may differ and this is an order-of-magnitude reference only).
Meta/Broadcom (OFC 2026): 36M device-hours, with 75% of ports at FEC bin below 4.
Oracle (ECOC 2026): named only; Palkert said he did not have approval to show the data.

▲ Samtec cites Meta and Broadcom CPO field data and lists four reasons for fewer link flaps (Image source: Samtec | 2026 OCP Global Summit presentation)

Samtec gives four reasons: simpler mechanics than pluggables, fewer DSPs in the path, shorter host channels, and fewer trace discontinuities (vias, BGA escape). By Samtec's argument CPX inherits these advantages, but all of this data comes from CPO soldered onto the package, a system we covered in 2026 OCP APAC Summit | Broadcom | CPO's Third-Generation Report Card.

During Q&A an audience member asked directly about field deployment. Palkert answered that Samtec makes the connector and cannot speak for end customers; hyperscalers are deploying, but he was not sure about large volumes: "Meta's got their data; as far as I know, we don't have that kind of hours of test data."

▲ The reliability data Samtec cites comes from fixed CPO; pluggable CPX sockets have no public field hours yet (Image source: Simple Tech Trend original; data from Samtec, 2026 OCP Global Summit presentation)

5. Breaking Down the Mechanics: Two Reflows and One Removal Tool

The CPX mechanical design targets three old pluggable problems: no cage, no compression bar, and no module sliding over the thermal interface material (TIM). It presses down and latches, and the thermal interface drops on top.

The real key is the process. Palkert played a Samtec Si-Fly HD reflow video: 32-lane sockets are pick-and-placed onto the package substrate and reflowed; a stiffener is added, and the whole substrate is reflowed onto the main board, with the upper socket not allowed to shift during the second reflow. Many people believe a pluggable socket cannot be reflowed together with the package, and the video is a direct answer to that doubt.

Serviceability relies on a removal tool that grips the module front and back and lifts out a single unit, with no compression bar to lift and no need to pull the whole tray. Palkert verbally claimed 50 to 100 mating cycles.

6. Two Socket Types, One Tray Contract

Open CPX defines two connector and module types:

Item	Type-1	Type-2
Lanes	32 lanes	36 lanes
Power and management	Separate MP socket (10 mm), power via main board	Same socket as high-speed signals, one extra row of pins
Use case	On top of package, package supplies no power	Near package, substrate can supply power
Show demos (stated in Q&A, speaker hedged)	Marvell; Ciena external laser (speaker recalled at ECOC)	TeraHop integrated laser (speaker said "I believe")

At 200G/lane, 32 lanes is 6.4T and 36 lanes is 7.2T (STT calculation). On protocols, the spec says little; system work mainly references IEEE 802.3dj 200G/lane PAM4, and the slides also list OCI, PCIe/CXL and proprietary interfaces. On copper it supports DAC flyover, analog ACC, DSP-based AEC and RF conversion cables. Connectors come from Samtec, TE, Molex and other second sources, and the speaker said cross-vendor interop testing will "hopefully" happen soon.

▲ The two Open CPX socket types: Type-1 with separate power and 32 lanes, and Type-2 with integrated power and 36 lanes (Image source: Samtec | 2026 OCP Global Summit presentation)

At the system level, Palkert borrowed an OCP composable switch tray slide: a 204.8T-class switch with 512 × 400G or 1,024 × 200G ports, flyover copper and optics on one tray, and faceplate pluggables and CPO under one tray contract. That works out to 32 CPX positions of 32 lanes for one 204.8T switch (STT calculation).

7. Risks and Counterarguments

Is CPX really CPO: An audience member said the CPX socket is too large for true CPO. Palkert himself conceded that integrated optics will always be denser. CPX sits closer to NPO, consistent with our view in NPO Is Not a Stepping Stone to CPO.
Reliability is inferred, not measured: A pluggable socket adds a contact interface, and there is no public data on contact resistance or link flaps after 50 to 100 mating cycles.
The 65% is a cited value: It comes from an OFC paper comparing CPO/NPO overall, not a measurement of the CPX socket.
Compliance testing is not yet defined: Samtec itself lists "robust compliance test parameters and methods for high-density CPO/NPO" as an open problem, and multi-vendor interop is hard to establish before test methods are settled.
400G/lane is a roadmap item: The loss budget holds at 200G/lane; whether the socket can stay within 1 dB when the frequency doubles, the deck gives no number.

8. What It Means for Taiwan's Supply Chain

No Taiwanese vendor among the socket suppliers the speaker named. The named suppliers are Samtec, TE and Molex, plus unnamed second sources. A CPX socket has to hold mated loss below 1 dB and survive two reflows, so the technical bar is not low.
Value shifts toward substrates and assembly processes. The NPO substrate is a new fan-out board, and "socket reflowed to package, substrate reflowed to main board" is a new SMT flow. Yield and warpage control in these two steps will directly decide whether the socket can ramp reliably; the slides name no substrate or assembly vendor, and Taiwanese participation cannot be confirmed from public information.
Compliance testing is the next entry point. With test parameters and methods still open, demand for fixtures and test services may appear before volume production.

9. Conclusion

The value of this talk is that it pulls Open CPX away from the "open ecosystem" and "floor space" narratives and back to the most basic physical budget: an electrical channel below 18 dB die to die, in which the socket may take less than 1 dB. Hold that, and the linear interface works, the DSP goes away and power drops sharply; miss it, and CPX is just a more expensive pluggable.

Our view: CPX will first establish itself in 200G/lane NPO switches, but its reliability story currently rests on other people's data. Three things to watch over the next 12 months: results of multi-vendor interop testing, the first published field hours for CPX sockets, and when high-density compliance test methods are written into the spec.

CPO proved optics can stay up; CPX has to prove that when optics fail they can be swapped, and that the swapping itself does not make them fail more often.

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