OCP Global Summit 2026 | FIT | Terry Little | XPO: A High-Density, Liquid-Cooled Pluggable Optical Interconnect

1. Why FIT Is the One Talking About XPO

The XPO (eXtra-dense Pluggable Optics) MSA is championed by Arista co-founder Andy Bechtolsheim. Its goal is to put CPO-class density and pluggable serviceability into a single form factor. The MSA launched in March with about 25 members; by OCP it had passed 150.

In 2026 OCP APAC Summit | CPO / NPO / XPO Panel we covered where XPO sits at the system level. This talk is different: the speaker is Terry Little, Sr. Development Engineering Manager at FIT, and the subject is the parts you never see light in: connector, cage, cable, power board and cold plate.

That is exactly FIT's position. Optical module makers decide whether LPO, LRO or DSP goes inside, but whether XPO works at all depends on the 0.56 mm pitch card-edge connector, the cold plate and the twin-ax cable running straight to the ASIC. FIT was blunt on stage: it makes the connector, the mezzanine connectors and the cable assemblies in-house, and the cold plate is being tooled now.

XPO's optics can come from any supplier, but FIT wants to be first to build its mechanical skeleton.

2. Breaking Down the Architecture: Two Paddle Cards Around One Cold Plate

Here is how to read the XPO plug module exploded view:

• Two paddle cards (A and B), top and bottom, with an integrated cold plate in between; coolant enters and exits at the rear of the module, and MPO fiber connectors sit at the front.

• High-speed signals: the connector provides 128 differential pairs (64 Tx / 64 Rx), built from 8 cable wafers stacked in one housing. At 224G one module carries 12.8 Tbps; at 448G, 25.6 Tbps, in the same form factor.

• Power and sidebands merged: each of 8 OSFPs has its own Vcc and sideband contact group. FIT calls "64 extra pins" illogical and replaces them with one dedicated power and sideband PCB lying inside the cold plate, connected to the host through a "pass-thru" power connector (1.0 mm board on the module side, 1.57 mm on the host side), then fed to both paddle cards by two mezzanine connector pairs. The speaker said verbally that this power board is designed to handle 50 V.

• Host side is cabled directly: there is no PCB routing behind the connector; 30 or 32 AWG drainless twin-ax cable runs straight to the ASIC, and FIT uses resistance welding to hold impedance at 92 Ω.

Put simply: OSFP lines up 8 separate small boxes; XPO puts their high-speed, power, sideband and cooling into one box.

3. Breaking Down Density: 3x More Bandwidth per Millimeter of Panel

▲ For the same 12.8 Tbps, 8 OSFPs occupy a 190 × 16.5 mm panel area while one XPO needs only 60.79 × 21.10 mm, about 3.1x the bandwidth per millimeter of panel width (Image source: Simple Tech Trend original, data from FIT's 2026 OCP Global Summit presentation)

Plugging in the numbers from the slides:

Item	8 x OSFP (1.6T)	1 x XPO
Bandwidth	12.8 Tbps	12.8 Tbps
Panel footprint	190 x 16.5 mm	60.79 x 21.10 mm
Panel area	about 3,135 mm²	about 1,283 mm²
Bandwidth per mm of width	about 67 Gbps	about 211 Gbps
High-speed contact pitch	0.60 mm	0.56 mm

A single XPO is about 4.6 mm taller than a row of OSFPs, but the math has to cover the whole faceplate. OSFP switches must leave room between every row of ports for riding heat sinks or cold plates; FIT's comparison shows 64 OSFP ports in 2RU. XPO moves the cold plate into the module, and the same logic yields 128 XPO ports in 1OU, which FIT estimates as about 4x port density.

▲ 128 XPO ports in 1OU versus 64 OSFP ports in 2RU; with OSFP, the vertical space between rows is taken up by cold plates or heat sinks (Image source: FIT | 2026 OCP Global Summit presentation (05:17))

Note that the 128-port 1OU layout is marked on the slide as "being carefully considered"; mechanical, thermal and routing work is still under way. At 12.8 Tbps per port, that faceplate would exceed 1.6 Pb/s, far beyond any single switch ASIC. In other words, after XPO the faceplate is no longer the bottleneck; the bottleneck moves back to the ASIC and the rack's cooling budget.

4. Breaking Down Thermal and Mechanical: 200 W Works, 250 W Not Yet

Density comes at the cost of concentrated heat. FIT's published cold plate simulation is a 1x2 cage with a parallel manifold at 200 W:

• At a maximum flow rate of 0.4 L/min, the two paths show pressure drops of 34.63 kPa (5.02 psi) and 36.44 kPa (5.29 psi)

• Maximum module temperature is 64.49°C, under the 70°C spec

• The cold plate is C10200 soft copper; with 200 psi applied to all internal channel walls, maximum von Mises stress is 30.87 MPa against a 49.8 MPa yield strength, and maximum deformation is 1.136 µm

▲ 200 W thermal simulation of an XPO 1x2 cold plate in a parallel configuration: maximum module temperature is 64.49°C at 0.4 L/min (Image source: FIT | 2026 OCP Global Summit presentation (10:05))

200 W at 12.8 Tbps is about 15.6 pJ/bit, or a 25 W budget per 1.6T module; the speaker said this covers retimed and LRO-class designs. The DSP version has not been done yet, and anything above 250 W needs higher flow and micro-channel cold plates. As we saw in Tech Analysis | 51.2T NPO Switch Thermal Design, the real challenge in liquid cooling lies in manifolds, quick disconnects and service procedures, not the cold plate itself.

▲ At 200 W the temperature headroom is about 5.5°C, while stress under burst pressure is about 62% of yield strength, so the thermal margin is tighter than the mechanical margin (Image source: Simple Tech Trend original, data from FIT's 2026 OCP Global Summit presentation)

Our read: mechanical margin is comfortable, thermal margin is tight. The 5.5°C of headroom applies only at 200 W; DSP modules will need a new cold plate design.

5. Breaking Down Signal Integrity: 224G Passes, 448G Is Still Simulation

For the 224G generation, FIT has measured data:

• OFC 2026 live demo: 3.93 dB of loss at 53.125 GHz; subtracting 3.0 dB for the 300 mm loopback cable and dividing by the top and bottom interfaces gives about 0.465 dB per mated interface

• Bit error rate: using an EXFO BA-1600 and Wilder Tec's XPO 224G host compliance board (128DP), 8 differential pairs at 106.25 GBd PAM4 showed pre-FEC BER from 1.07E-12 to 7.25E-10 and zero post-FEC errors

448G is another story. The stub of the traditional cantilever beam contact causes insertion loss to collapse at high frequency; the 224G design falls below -10 dB near 90 GHz. FIT's second-revision 448G contact simulation:

Metric	224G XPO	448G XPO (simulated)
Insertion loss at 67 GHz	-0.94 dB	-0.59 dB
Insertion loss at 80 GHz	-1.32 dB	-0.89 dB
Return loss at 80 GHz	-10.42 dB	-15.26 dB
Impedance at mating	81.93 Ω	89.00 Ω

▲ Simulated insertion loss of the 448G XPO card-edge contact, with Nyquist markers at 89.6 GHz (PAM6) and 112 GHz (PAM4) (Image source: FIT | 2026 OCP Global Summit presentation (16:25))

FIT targets both modulations: PAM6 at 2.5 bits per symbol gives 179.2 GBd and an 89.6 GHz Nyquist; PAM4 needs 112 GHz. This echoes our conclusion in ECOC2026 | Three Hours of Fast-Narrow vs Slow-Wide, and the Dividing Line Is 448G: IEEE 802.3dv has not picked PAM4 or PAM6, so connector makers have to bet on both.

6. Risks and Counterarguments

• 448G is simulation only: the speaker himself called it the second simulation revision with prototypes in progress. Holding a 0.56 mm card edge to 112 GHz counts only once measured curves appear.

• Liquid cooling inside a pluggable module dilutes serviceability: with coolant inlets and outlets on the module, hot-swap now involves quick disconnects, drips and service procedures. Part of the "pluggable" advantage gets eaten by liquid cooling.

• 12.8 Tbps per module multiplies failure impact by 8: swapping one OSFP used to affect 1.6T; now one swap is 12.8T, so spares cost and inventory strategy must be recalculated.

• The DSP-class thermal design is unfinished: 200 W suits LPO, LRO and retimed modules; DSP modules may exceed 250 W, and FIT's answer is micro-channels, with no numbers yet.

• Self-measured data: the OFC demo used FIT's own host compliance board, two BER channels sit at the E-10 level, and each channel ran for 180 seconds, still a long way from production qualification.

7. What It Means for Taiwan's Supply Chain

• The Hon Hai (Foxconn) group gains a mechanical foothold in XPO. FIT is Hon Hai's connector and cable subsidiary; the connector, cage, mezzanine connectors, twin-ax cable assemblies and cold plate shown here are all FIT designs, with cables built in FIT's own factory using resistance welding. If XPO becomes the mainstream pluggable form factor for AI switches, FIT is the first to show a full set of parts and measured data.

• Optical module makers must learn liquid cooling. With the cold plate inside the module, mechanical and thermal design capability becomes a barrier, not just optics and DSP. For Taiwan's module and component makers focused on optical packaging, this is a new engineering challenge.

• Demand for cable-to-ASIC assemblies grows with it. Behind every XPO are 128 pairs of 30 to 32 AWG twin-ax, so cable assembly process capability and value scale with port count.

8. Conclusion

The significance of FIT's talk is not that XPO adds another optical option, but that it moves cooling from the switch into the module, lifting faceplate density to 3 to 4 times OSFP while folding 8 sets of power and sideband contacts into one board. The 224G connector, thermal and mechanical numbers already add up, which is why the MSA grew from 25 to 150 members this year.

But the timing has to be separated. 224G XPO is something you can discuss for volume production in 2026 to 2027; the 448G contact is still in its second simulation revision, and the DSP-class cold plate has not been done. Three things to watch over the next 12 months: measured insertion loss on 448G prototypes, a cold plate for DSP-class modules, and IEEE 802.3dv's choice between PAM4 and PAM6.

Once the faceplate stops being the bottleneck, XPO comes down to two things: how many watts the cold plate can absorb, and how many GHz the contact can hold.

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