OCP Global Summit 2026 | Amphenol, FIT, Luxshare-Tech, TE | Sam Kocsis, Terry Little, Robbie Xu, Matthew Ho | 400G/Lane Optics Module Connectors

2 days ago

8 min read

1. Why This Session Matters: OSFP Has No Road to 448G

This was a set of lightning talks followed by a panel, moderated by Andy (Arista co-founder Andy Bechtolsheim, acknowledged on FIT's slides). Four connector makers took turns on stage with a single question: how should the electrical interface of a 400G-class (448G per lane) optical module mate.

TE's Matthew Ho framed it bluntly at the start: the existing OSFP specification has no clear path to 448G. Amphenol's Sam Kocsis ranked the bottlenecks beyond 224G in two: first the ASIC package and BGA breakout, second the pluggable I/O itself. The first is being solved with cabled hosts and package co-design; the second is what this session was about.

XPO (eXtra-dense Pluggable Optics) is the shared backdrop. We covered its position between CPO and NPO in OCP APAC Summit 2026 | CPO / NPO / XPO panel. Amphenol's slides list the XPO spec: 64 high-speed lanes, a single 48 V supply, integrated liquid cooling, MPO-16 compatible interfaces, and under 90 W for a linear module. FIT's Little said on stage that the XPO MSA started in March with about 25 members and now has more than 150.

At 224G the question was whether XPO could replace OSFP; at 448G it becomes whether the card edge survives at all.

2. Breaking Down the Bandwidth: PAM6 Needs 90 GHz, PAM4 Needs 112 GHz

FIT laid out the arithmetic on its slide: at 448 Gb/s with PAM6 (5 bits over 2 symbols, 2.5 bits per symbol), the symbol rate is 448 / 2.5 = 179.2 GBd and the Nyquist frequency is 89.6 GHz; with PAM4 the Nyquist is about 112 GHz. FEC and line coding push the real numbers slightly higher.

Amphenol overlaid the insertion loss of every pluggable generation on one chart: 10G, 28G, 56G, 112G, 224G. Each generation pushes the roll-off further right, and the 224G example drops to -10 dB around 80 GHz. Amphenol's target for 448G is 130 GHz, and the slide's conclusion is a single line: IO bandwidth drives modulation.

That line is the key to the session. As we wrote in ECOC 2026 | Fast-narrow vs slow-wide: the dividing line is 448G, IEEE 802.3dv has not yet chosen PAM4 or PAM6 for 448G. Following the logic through (STT reading): if connectors can reach beyond 112 GHz, PAM4 has a chance; if not, PAM6 is the only option.

▲ Insertion-loss roll-off points of four 448G connector designs (STT chart-reading estimates) against the PAM6/PAM4 Nyquist thresholds: all four clear PAM6, but headroom above 112 GHz is only 0–14 GHz (Image source: Simple Tech Trend original; data from Amphenol, FIT, Luxshare-Tech, TE 2026 OCP Global Summit presentation)

Design	Interface type	Roll-off (STT chart reading)	Headroom above 112 GHz	Data type
224G pluggable (Amphenol example)	Card edge	~78 GHz	Insufficient	Shipping product
FIT 448G XPO	Card edge, new contact	~118 GHz	~5%	Simulation
Amphenol XPO 448G	Card edge, tighter tolerances	~123 GHz	~10%	Projected curve, booth demo
Luxshare-Tech 448G OSFP	Two-piece backplane-style	~111 GHz	Near zero	Simulation, samples
TE MegaLane	Two-piece, stubless	~126 GHz	~13%	Simulation, TE cites measurements

Roll-off points are approximate STT readings of the presenters' charts (about ±3 GHz); test conditions differ, so compare relative positions only.

3. The Card-Edge Camp: Amphenol and FIT Keep the XPO Form Factor

Amphenol's approach is to repeat what XPO already did once. Its slide lists five directions: reuse the mechanical and thermal infrastructure, preserve the card edge interface, tighten tolerance loops, further reduce pad pitch, and offset high-speed lanes to limit crosstalk. Kocsis said XPO is already slightly better than OSFP at 224G precisely because of tighter tolerances and smaller geometries; 448G needs another round. Amphenol ran a live demo at its booth, and the XPO 448G curve on its slide holds until about 122 GHz before rolling off.

▲ Insertion loss of Amphenol's pluggable I/O generations compared; the XPO 448G curve does not roll off until above about 120 GHz (Image source: Amphenol | 2026 OCP Global Summit presentation)

FIT's approach keeps the shape and replaces the contact. The tip of a traditional cantilever-beam contact leaves a stub that causes insertion-loss roll-off at high frequency. FIT's 448G connector keeps the same card-edge form factor at 0.56 mm pitch but uses a new "compliant medium" contact; details are undisclosed and prototypes are still being built. The simulation results:

Mating impedance rises from 81.93 Ω at 224G to 89.00 Ω, and cable-solder impedance from 85.04 Ω to 92.00 Ω
Insertion loss at 80 GHz improves from -1.32 dB to -0.89 dB, and return loss from -10.42 dB to -15.26 dB
About -2 dB at 112 GHz, then sliding to about -6 dB without falling off a cliff

When the moderator asked whether the same contact works for OSFP, Little said yes: OSFP, QSFP and other card-edge styles. The card-edge camp is not only betting on XPO, it is trying to extend the life of the whole card-edge family.

4. The Two-Piece Camp: Moving Backplane Connectors into Pluggable Modules

Luxshare-Tech's logic is "if the backplane can do it, bring it to the front panel". Its NEXUS 448G backplane connector shows -0.8 dB at 53 GHz and -1.8 dB at 90 GHz, with power-sum crosstalk (PSXT) below -80 dB from 0 to 120 GHz; a measured CPC-to-backplane-to-CPC link is clean to 90 GHz with no resonance until about 110 GHz. It puts the same two-piece interface on both the module side and receptacle side of a traditional paddle card, creating the "448G OSFP Airchannel" (labeled 448G HSIO OSFP on slide 27): close to OSFP in size, simulated bandwidth beyond 110 GHz, a receptacle with 8 pairs in each of 2 rows, and an open question of whether it could become a candidate interface for OSFP or OSFP-XD.

▲ Luxshare-Tech moves a backplane-style two-piece interface onto an OSFP-size module, with simulated bandwidth above 110 GHz (Image source: Luxshare-Tech | 2026 OCP Global Summit presentation)

TE goes furthest and abandons the card edge outright. AdrenaLINE MegaLane solders a metal receptacle frame to the top and bottom of the paddle card, fully shielding every differential pair, and lands the host twinax right next to the contacts, with no stub. The simulation (including 0.5 mm trace, module card pads, connector and 2 mm cable) is flat to about 126 GHz, with crosstalk around -60 dB out to 130 GHz.

Two other design choices may matter more to system makers than the bandwidth:

Scalable: the same interface comes in x8, x16, x32 and x64; it was originally designed for x64
Power isolated from high speed: power and low-speed signals use a separate connector; cold plates can sit on both sides, and TE shows two modules stacked within 1RU

▲ TE MegaLane module applications: scalable from x8 to x64, power separated from high-speed, and two modules stacked between top, middle and bottom cold plates (Image source: TE Connectivity | 2026 OCP Global Summit presentation)

TE's slides add one more line: USFP adopts MegaLane (Ho said as its electrical interface), with a link to an NVIDIA news release about a new pluggable optics form factor from NVIDIA and Cisco. In other words, the two-piece camp already has a system-level form factor attached; the slide does not state the USFP lane rate, and the card-edge camp did not present a system-level 448G adoption on stage.

5. Breaking Down the Cost: Backward Compatibility, Impedance and Tolerance

Putting all four side by side, the 448G connector trade-offs fit into one chart:

▲ Three ways to build a 448G pluggable electrical interface: 224G cantilever-beam card edge, improved 448G card edge, and a stubless two-piece interface, compared on bandwidth, compatibility and manufacturing challenges (Image source: Simple Tech Trend original; data from Amphenol, FIT, Luxshare-Tech, TE 2026 OCP Global Summit presentation)

Backward compatibility: the final panel question asked whether 448G XPO would be backward compatible. Kocsis said it may have to be sacrificed to reach 120–130 GHz, adding that deployments are customized enough to justify bifurcating by data rate; Little said the paddle card will definitely change but the connector can stay as compatible as possible, just as OSFP changed its paddle card at 224G.
Tolerance: an audience member questioned whether a card edge can hold mechanical tolerances at 448G. Little answered that XPO's latching "marries" the module to the connector so wiping distance can be controlled; Kocsis said tightening tolerances is uncomfortable every generation, but once performance is demonstrated, volume tolerances will follow.
Impedance: Luxshare-Tech is developing 92 Ω and 85 Ω in parallel and urged the industry to settle the mainstream impedance before 448G system design kicks off. FIT's 448G design moves toward 89–92 Ω, and TE's simulation sits near 92 Ω.

One panelist added a practical point: for front-panel optics PAM4 is best; for longer reach toward the back, PAM6 makes more sense. Connector bandwidth will end up choosing the front-panel modulation.

6. Risks and Counterarguments

Almost all simulation: FIT's prototype missed the show, Luxshare-Tech's OSFP interface is simulation plus samples, and TE cites measurements but showed only simulated curves on stage. A 5–10 GHz difference in roll-off could flip once measured.
Chart comparisons have caveats: TE's curve includes 2 mm of cable and the topologies differ, so roll-off points show trends only.
If PAM6 wins, the card edge need not fork: all four clear 89.6 GHz. If 802.3dv picks PAM6 for the front panel, every design has ample headroom, TE's bandwidth lead matters less, and the card edge need not give way.
Two-piece has cost and serviceability downsides: more parts, metal frames soldered to the paddle card, and new mating-cycle qualification, none with public numbers yet.
USFP and XPO may not be mutually exclusive: their customers and rack positions may differ, so a fork does not mean anyone loses.

7. What It Means for Taiwan's Supply Chain

FIT (Foxconn Interconnect Technology) is the most direct Taiwan-group position in this session. FIT is the Hon Hai group's connector arm; its slides introduced its own XPO platform, and it now brings a 448G card-edge contact. Whether the card edge survives to 448G directly affects how many generations FIT's XPO investment can carry. Watch for when FIT publishes prototype measurements.
The other three on stage are non-Taiwanese connector majors. Amphenol (in the card-edge camp with FIT) and TE are US-based, and Luxshare-Tech belongs to China's Luxshare group; no Taiwanese vendor presented in the two-piece camp.
Optical modules and paddle cards must prepare for a redesign. Whichever camp wins, the 448G paddle card changes: finer pad pitch, offset lanes, or soldered metal frames. Taiwan's optical module makers and their board supply chain should track both designs before 802.3dv is settled to avoid betting on the wrong side.

8. Summary

The value of this session is that four connector makers put the 448G pluggable interface options on one stage for the first time. The card-edge camp trades tolerance and contact design for compatibility; the two-piece camp uses a new interface to eliminate stubs and gain scalability. All four clear PAM6's 89.6 GHz, but headroom above the 112 GHz PAM4 threshold is only 0–14 GHz, and it does not split by camp.

Our call: the era of OSFP as the single front-panel standard ends at 448G. XPO will continue on the card edge with a new paddle card generation, while two-piece interfaces like USFP will land first in another set of systems. Over the next 12–18 months, watch three things: 802.3dv's modulation choice, measured curves from FIT's 448G prototype, and the 92 Ω vs 85 Ω impedance decision.

The 448G contest is not about the optics, but about whether the edge of that paddle card can hold to 112 GHz.

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