OCP Global Summit 2026 | Luxshare-Tech | Robbie Xu | Liquid-Cooled OSFP 2x8 Cage for 1.6T Interconnects

1. Why Luxshare-Tech Is the One Telling This Story

1.6T optical module power keeps climbing. Robbie Xu put numbers on it: to reach longer distances and better bit error rates, module power has gone from 30 W to 40 W, and 45 W in extreme cases. Add 224G SerDes, high-density 2x8 cages and limited front-panel airflow, and air-cooled heat sinks are running out of road.

Luxshare-Tech's slides break the problem into six items: system thermal budget, compact integration, thermal uniformity, robust liquid sealing, electrical/thermal/mechanical co-design, and one that reads very much like a connector maker: "replicate and deliver OSFP 224G+ liquid-cooling solutions within 3 months".

Cooling is no longer the thermal vendor's job alone: once the cold plate has to be pressed together with the cage, whoever makes the cage gets the cold plate.

That is why it took the stage: Luxshare-Tech stresses that both the cold plate and the cage are designed and manufactured in-house, with simulation and automated production also kept internal.

2. Architecture: The Cold Plate Lives on the Cage, Not the Module

The 1.6T OSFP cold plate assembly is a 64-port belly-to-belly configuration:

• Top floating cold plate plus bottom floating cold plate, clamping the upper and lower rows of the 2x8 cage; each port's plate floats independently, linked by copper tubes for pressure resistance

• Single-side inlet and outlet, connected through MQD03 fluid quick disconnects, cutting internal tubing and joints

• The top plate has no extra fastening structure; clips provide the downward pressure so it stays in contact with the module for its whole life

• One-step cage and cold plate press assembly, marked by Luxshare-Tech as patent pending

▲ A 64-port belly-to-belly 1.6T OSFP liquid-cooled assembly with upper and lower floating cold plates, MQD03 quick disconnects and light pipes (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation (04:53))

The hard numbers on the spec page: module power 40 W, max module case temperature 70°C or below, total pressure drop across 16 ports 35 kPa or below, flow 1.15 L/min or below, coolant up to 25% glycol, inlet temperature 40°C or below. Footprint is 200 x 110.7 mm and tested samples stand 37.5 ± 0.5 mm tall (the 34 mm heard in the transcript does not match the slide; we follow the slide).

▲ Dimensions and specs of the OSFP 224G 2×8 cage and cold plate assembly: 40 W, ≤70°C, ≤35 kPa, 1.15 L/min (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation (05:53))

The value is straightforward: switch makers only swap the cage and do not wait for a new module form factor. As we covered in 51.2T NPO Switch Thermal Cooling, once the front panel is packed with optics, airflow becomes the first bottleneck; a liquid-cooled cage is the least disruptive way around it.

3. Thermal Data: A 40 W Spec, a 32 W Report Card

Both simulation and measurement used 16 x 32 W = 512 W, not the 40 W in the spec:

Item	Simulation	Measured (10 samples)
Heat load	512 W	16 x 32 W
Flow	1.15 L/min	1.15 L/min
Inlet temperature	40°C	40.8-42.7°C
Hottest port	66.35°C	62.6°C
Coolest port	—	48.4°C
Pressure drop	34.18 kPa	Not disclosed

Three things deserve a second look:

• Port-to-port spread within one sample reaches 12.6°C (#006: 61.0 vs 48.4°C). The speaker called it very uniform, yet 5 of 10 samples show spreads above 10°C.

• The 34.18 kPa pressure drop sits right against the 35 kPa limit, so more flow would require a new channel design.

• The coolant warms only about 4°C in testing. Based on water's specific heat, 512 W fully absorbed should raise it about 6.4°C, so some heat went into the fixture or air, or the test power was defined differently; the slides do not say.

▲ Across 10 samples at 32 W, the hottest port reads 55.2–62.6°C; STT's linear extrapolation to 40 W puts the hottest port at about 67–68°C, only 2–3°C below the 70°C limit (Image source: Simple Tech Trend original, data from Luxshare-Tech's 2026 OCP Global Summit presentation)

Take the hottest sample, #003: at 32 W it runs 19.9°C above inlet; scaling with power to 40 W gives about 24.9°C, plus a 42.7°C inlet equals 67.6°C. This is STT's simplified estimate that ignores flow and thermal resistance non-linearity, but the direction is clear: 40 W is the ceiling this design can reach, not a design point with margin. If 1.6T modules reach the 45 W the speaker mentioned, this cage will not be enough.

▲ 16-port thermal test results for 10 samples, upper cold plate T1–T8 and lower cold plate T9–T16, with a maximum of 62.6°C and a minimum of 48.4°C (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation (09:57))

4. Mechanics: After 100 Cycles, Only 0.6 N of Insertion Margin Left

A liquid-cooled cage adds one job a conventional cage does not have: the cold plate must keep pressing down on the module, or contact thermal resistance drifts. The price is plug force.

• First insertion: across 16 samples, insertion force was 28.5-41.3 N (limit 55 N) and extraction force 25.3-30.9 N (limit 45 N), all passing

• After 100 cycles: the highest insertion force among 8 ports was 54.4 N (port 8, first cycle 39.0 N), with port 4 at 52.9 N; the highest extraction force was 41.3 N (port 4, first cycle 27.9 N)

▲ After 100 insertion cycles, insertion force generally rises, with port 8 at 54.4 N close to the 55 N limit, and extraction force up by as much as 53% (Image source: Simple Tech Trend original, data from Luxshare-Tech's 2026 OCP Global Summit presentation)

Everything is marked Pass, but the trend points up. Data center modules are swapped far fewer than 100 times, so this will not block volume production; the real questions are whether the curve keeps climbing beyond 100 cycles, and how many ports exceed 55 N under production tolerances. The speaker named the same challenge himself: moving from samples to volume means every port's cold plate, cage and module contact must be consistent.

5. Two Liquid-Cooling Paths: Cold Plate on the Cage, or Inside the Module

At the same summit, FIT (Foxconn Interconnect Technology) presented XPO, which puts the cold plate inside the pluggable module itself; one XPO equals eight OSFPs. One table shows the difference:

Item	Luxshare-Tech liquid-cooled OSFP cage	FIT XPO
Cold plate location	On the cage, module unchanged	Inside the module
Heat load	40 W per 1.6T port (tested at 32 W)	200 W per module, about 25 W per 1.6T (simulated)
Flow	1.15 L/min for 16 ports	0.4 L/min per 1x2 cold plate (simulated)
Ecosystem	Reuses existing OSFP modules	New form factor, needs a new MSA

Our read: cage-level cooling takes DSP modules first, while XPO takes LPO and retimed modules first. XPO's 25 W budget is not enough for a 40 W DSP module, and cage cooling does not wait for anyone to change a spec. Luxshare-Tech is not betting on one side either: its booth also demos "224G OmniStack NPC to XPO Airchannel". For the trade-offs across XPO, NPO and CPO, see OCP APAC Summit 2026: CPO / NPO / XPO Panel.

6. Risks and Counterarguments

• 40 W was never measured: all thermal data is at 32 W, the 40 W spec is an extrapolation, and 45 W is outside the design.

• The uniformity claim is optimistic: a 12.6°C port-to-port spread within one sample means "consistent" needs a discount.

• The heat balance does not close: the coolant temperature rise is about one-third below theory, so how much heat the liquid actually carries is unclear.

• Durability covers only 100 cycles: insertion force is clearly trending up, and the distribution at higher counts and production tolerances is not shown.

• The limits lack an independent source: 55 N and 45 N are cited as "industry standard" on the slides, without naming the MSA clause.

7. What It Means for Taiwan's Supply Chain

• Cold plates are being folded into the connector BOM. Luxshare-Tech's pitch is co-designing the cold plate with the cage and pressing them together in one step. If switch makers accept this integrated part, I/O cold plate orders follow the cage instead of going separately to thermal vendors. For Taiwanese thermal makers such as AVC and Auras, whose cold plate business centers on GPUs, the thing to watch is whether they can pair up with cage makers rather than sell cold plates alone.

• Connector makers lead liquid-cooled I/O, one from Luxshare and one from Foxconn. At the same summit FIT (Foxconn group) pushed XPO and Luxshare-Tech pushed liquid-cooled OSFP cages; both paths are led by connector makers, and the Foxconn group already holds an early position in XPO mechanics.

• Module makers' power budgets are being repriced. Cage-level cooling lifts the ceiling for 1.6T DSP modules to 40 W, which is good news for module makers; beyond 40 W, they wait for the next cage. Compare with the 1.6T ramp in Earnings Highlights: Innolight 2026 H1.

8. Conclusion

The point of this talk is not one new cage; it is that liquid cooling has officially moved from the chip to the front panel, and connector makers are leading it. Luxshare-Tech showed that without changing the module, simply swapping the cage frees the 1.6T front panel from the airflow bottleneck.

But read the report card carefully. The thermal data is at 32 W, not 40 W, and after 100 cycles insertion force has only 0.6 N of margin. Over the next 12 months, watch three things: measured 40 W data, the first switch platform to adopt a liquid-cooled cage, and when the XPO MSA raises its power budget to DSP levels.

The cooling battlefield has moved from heat sinks to cages; the next bottleneck is whether every port can be pressed equally tight, and stay that way for life.

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