OCP Global Summit 2026 | TeraHop | Rang-Chen (Ryan) Yu | Open Optical Solutions for AI Scaling

1. Why TeraHop Is the One Talking About "Open"

AI networking is no longer one network but four: scale-up, scale-out, scale-across, plus the scale-in front-end network (GPU to storage) recently framed by NVIDIA. Yu opened by saying the rack used to be the computer; now the whole data center is the computer.

TeraHop's slides list four bottlenecks, each with a number attached:

• Bandwidth: scale-up bandwidth is roughly 10x scale-out, about 10-20 Tbps per xPU.

• Density: neither ASIC shoreline nor faceplate density is sufficient. Citing an Arista illustration, a 512-xPU cluster built on today's OSFP needs 4 switch racks, as many as the compute racks.

• Fiber: citing Meta, a single AI data center already holds more than 10 million fibers, beyond cable-tray capacity, with more than 4 million fibers between data centers.

• Power: energy is the foundation of the AI data center, and every solution must clear the power test first.

The core message: CPO and NPO have mostly been proprietary, vertically integrated single-vendor solutions, and TeraHop wants to turn them into multi-vendor, interchangeable off-the-shelf parts. Ciena made the same case at ECOC, which we covered in ECOC 2026: Ciena Turns the 6.4T Optical Engine Back into a Shelf Part with Open CPX; this time another founding member fills in the data from its own products.

For TeraHop, open standards are not charity but a hedge: help define all three routes and you never have to bet on which one wins.

2. Breaking Down XPO: TeraHop Delivers the First 12.8T Module

The XPO headline specs are clear on the slide: 8 x 1.6T = 12.8T capacity and 4x faceplate density, making it the first pluggable with integrated liquid cooling. It supports FRO, RTLR and LPO architectures, optical PMDs spanning DRn, FRn, OCI, Coherent-Lite and ZR, and more than 150 companies have joined the MSA. Yu added that the goal is a 200 Tb/s switch in 1 OU.

TeraHop showed test data from what it calls the industry's first 12.8T XPO:

• All 64 transmitter eyes are open, and pre-FEC BER curves for all 8 x 1.6T-DR8 links reach around 1E-12.

• At 0.3 LPM, CW laser temperature tracks coolant inlet temperature: about 36-41°C at a 25°C inlet, 45-50°C at 35°C, and 53-57°C at 45°C.

• Module power stays at about 200 W across 45°C, 50°C and 55°C coolant and 0.3-0.5 LPM flow rates.

▲ TeraHop 12.8T XPO test results: 64-lane eye diagrams, pre-FEC BER for 8 × 1.6T-DR8, and laser temperature and roughly 200 W module power at different coolant temperatures and flow rates (Image source: TeraHop | 2026 OCP Global Summit presentation (08:41))

Doing the math: 12.8 Tb/s at about 200 W is roughly 15.6 pJ/bit (STT calculation). That is not low, but XPO was never about being the most efficient; it is about raising faceplate bandwidth 8x while keeping optics pluggable. Liquid cooling matters most for the laser: with 45°C coolant, the laser still stays below 57°C, which is a precondition for reliability.

3. Breaking Down Open CPX: One Socket, Four Placements

The formal name of the Open CPX MSA spec is "Co-packaged and Near-Package Connector and Optics Module", and Rev 1.0 was published on 2026/9/16. Founding members, in alphabetical order, are Ciena, Coherent, Marvell, Molex, Samtec and TeraHop; early contributors include Accton, Intel, Nexthop, Qualcomm and TE. Yu said more than 50 companies have now joined.

The design combines two dimensions:

• Connector: Type 1 and Type 2.

• Laser: both ILM (integrated laser module) and ELM (external laser module) are supported.

The same optical engine can be placed four ways:

Use case	Optical engine location	Connector
NPO for xPU scale-up	Host PCB	Type 1 or Type 2
NPO for switch	Host PCB, surrounding the switch ASIC	Type 1 or Type 2
CPO for switch	ASIC substrate, high-speed signals on substrate, power and I2C on host PCB	Type 1
IOA (Integrated Optical Attachment)	HDI board attached to the ASIC substrate via elastomer connector	Type 1 or Type 2

▲ Four uses of Open CPX: xPU NPO, switch NPO, switch CPO and the newly proposed IOA (Image source: TeraHop | 2026 OCP Global Summit presentation (10:25))

What this table says: NPO and CPO need not be an either-or choice; the only difference is whether the customer gives up substrate area for the connector. IOA is a new term that sits between the two, trading an HDI board for signal integrity.

TeraHop's own product is the Diablo-1 6.4T Open CPX optical engine: silicon photonics, integrated laser, Type 1 connectors. At the ECOC OIF interoperability demo it shared a board with Ciena's ELM version, co-packaged copper (CPC) and ELSFPs. Test data:

• OIF demo: BER across 8 lanes from 1.7E-11 to 5.3E-10, outer extinction ratio about 2.9-3.1 dB, TDECQ about 2.3-2.5 dB.

• Preliminary tests: ER 3 dB, TDECQ 2.1 dB, BER curves for 32 channels, with about 12 dB of insertion loss between ASIC and NPO.

That 12 dB condition matters: it means the NPO does not need to sit right next to the ASIC, leaving a meaningful electrical channel budget in between.

4. Breaking Down Integrated vs. External Laser: A 3 dB War

This was the most opinionated part of the talk. Using a 100 Tb/s switch surrounded by 16 6.4T modules as the example, Yu listed the benefits of an integrated laser:

• The NPO is close to but separate from the ASIC: separate cold plates keep the laser and other critical components at a modest temperature.

• About 5.5 pJ/bit overall, including the internal laser. Yu called it "probably one of the best benchmarks out there, including CPO".

• Simpler fiber management: plain single-mode fiber runs straight to the front panel, with no extra connector loss or failure point.

Then came four criticisms of ELSFP:

• Extra coupling and connector loss of more than 3 dB, meaning half the optical power is thrown away.

• Laser optical power must be more than 2x (200-300 mW+, versus about 100 mW inside), without field reliability data, and laser power dissipation rises about 4x.

• PM fiber is required, adding loss and cost, and PM fiber is not deployed in high volume today.

• Blind-mate connectors add another loss and failure point, and dust can cause failures.

▲ TeraHop's argument: an external laser loses 3 dB or more, pushing laser power from about 100 mW to 200–300 mW and raising laser heat dissipation about 4x (Image source: Simple Tech Trend original, data from TeraHop's 2026 OCP Global Summit presentation)

This matches our earlier view in CPO Lasers Don't Have to Chase 400 mW: NPO Integrated Lasers: once NPO brings the laser back into the module, laser power requirements fall from "high power" back to "medium power". Note that this is the integrated-laser camp's argument; the external-laser camp's case is field replaceability and distance from heat sources, which TeraHop did not address.

5. Breaking Down Reliability: 100 Billion Device-Hours Is the Moat

Yu asked the audience to remember one number: 100B device-hours, "big B, not M". The slide says TeraHop has shipped more than 25 million silicon photonics transceivers, with more than 100 billion PIC device-hours and FIT < 0.01.

The same chart includes several reference points (read from the chart by STT, not official figures):

• TeraHop 1.6T-DR8 OSFP (silicon photonics plus CW laser): about 0.2-0.3 FIT.

• 6.4T NPO-ILS (equivalent to 4 x 1.6T OSFP): about 1 FIT.

• 12.8T XPO (equivalent to 8 x 1.6T OSFP): about 2 FIT.

• CW laser vendors shown: Vendor A FIT < 0.03, Vendor E FIT < 0.01; Intel's integrated laser on PIC about 0.1.

▲ TeraHop plots FIT against cumulative device-hours: PICs exceed 100 billion device-hours with FIT below 0.01 (Image source: TeraHop | 2026 OCP Global Summit presentation (16:45))

The logic is multiplication: multiply channel count by 4x or 8x and FIT multiplies too; if the base is low enough, the result is still acceptable. Yu was blunt: customers do not want a $30,000-$50,000 XPU taken down by a failing optical part next to it, and when vendors talk FIT, ask for field deployment data, not short lab tests.

This is TeraHop's second cut at external lasers: integrated lasers have field data, while the 200-300 mW lasers that ELSFP needs do not yet.

6. Breaking Down SDM4 MCF: Fiber Count Divided by 4

The third MSA addresses the fiber itself. The SDM4 MCF MSA launched on 2026/3/11 with founding members AFL, Corning, Sumitomo and TeraHop, and Release 1.0 was published on 2026/9/17. It defines a 2x2 four-core fiber with a 125 µm diameter, each core compatible with conventional single-mode fiber, targeting short-reach, high-density O-band links in AI data centers and campuses.

Every benefit comes in "4x":

• Fiber count and interface terminations are reduced 4x.

• Cable size, weight and pathways (cable trays, underground ducts) shrink accordingly, cutting construction time and bringing AI capacity online faster.

TeraHop already has an 800G-MCF native transceiver, shown with Corning's MCF at OFC 2026 and in the OIF interoperability demo at ECOC 2026.

▲ TeraHop's three MSAs each target one bottleneck: XPO for bandwidth density, Open CPX for power density, and SDM4 MCF for fiber count (Image source: Simple Tech Trend original, data from TeraHop's 2026 OCP Global Summit presentation)

Fiber count has long been the most underrated bottleneck in optical interconnect, as we wrote in What Killed Slow-and-Wide Wasn't Physics, It Was 60,000 Fibers. The significance of SDM4 MCF: transceiver makers and fiber makers are writing the spec together for the first time, rather than each working alone.

7. Risks and Counterarguments

• 5.5 pJ/bit lacks measurement conditions: the slide does not say whether a DSP is included or at what rate and reach it was measured, so it may not share a baseline with other vendors' CPO numbers.

• FIT reference points are chart readings: the 1.6T, 6.4T and 12.8T FIT values are points on a chart; NPO and XPO have not shipped in volume, so those two points are likely component FIT multiplied by channel count, not field data.

• The ELSFP critique is one-sided: the advantages of external lasers (field replaceability, distance from ASIC heat) were not discussed.

• XPO's 200 W is heavy: each module draws 200 W and must be liquid-cooled, requiring coolant at the front panel, a major change for existing air-cooled switches.

• MCF requires changing the whole chain: fiber, connectors, modules, splicing and test equipment all have to follow; Release 1.0 has just been published, and adoption depends on whether hyperscalers go first.

• No volume production timeline: Diablo-1 is at the demo and preliminary-data stage, and the talk gave no shipment date.

8. What It Means for Taiwan's Supply Chain

• Taiwanese companies are on the Open CPX list. Among the early contributor logos are Accton, Alpha Networks and Lotes. NPO gives switch makers back control of optical engine integration, and once the connector spec is public, Taiwanese connector makers have a chance to enter an optoelectronic connector market long held by a few vendors.

• No Taiwanese company is in SDM4 MCF. The founding members are AFL, Corning, Sumitomo and TeraHop, and multi-core fiber and related connectors are currently outside Taiwan's footprint.

• This debate will decide CW laser specs. If integrated-laser NPO wins, demand tilts toward medium-power, high-reliability lasers around 100 mW; if ELSFP wins, it needs high-power lasers of 200-300 mW or more. Taiwanese laser makers should prepare both product lines and prioritize field reliability data, because TeraHop has made field data the entry bar.

9. Conclusion

The core of TeraHop's talk is not any single product but a betting structure: XPO defends pluggables, Open CPX enters NPO and CPO, SDM4 MCF handles fiber, and TeraHop helped launch all three. Its product argument converges on one point: integrated lasers backed by 100 billion device-hours of field data, positioned against the external-laser route.

Keep the timelines separate. What arrived in September 2026 is two specifications (Open CPX Rev 1.0 and SDM4 MCF Release 1.0), not volume production. Over the next 12 months, watch three things: when Diablo-1 gets customer adoption and a shipment timeline, the measurement conditions behind 5.5 pJ/bit in a full system, and when the first hyperscaler publicly deploys SDM4 MCF.

Which packaging form wins is still open, but TeraHop has already made reliability data the entry ticket to this race.

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