OCP Global Summit 2026 | Dell | Jian Chen | Beyond Copper: Optical Opportunities for AI Scale-up Fabrics

1. Why Listen to Dell on Optics: a Buyer, Not a Seller

Most of the CPO, NPO and XPO talks at this OCP came from module makers, connector vendors or chip companies, each pushing its own packaging format. Dell sits somewhere else: it is a system integrator that puts other people's optics into its own racks, sells them to customers, and is on the hook for repairs when something breaks.

So the point of this 12-minute talk is how a buyer writes its purchasing criteria. Discussing the optical integration continuum, Jian Chen said no technology optimizes reach, power, density, cost, manufacturability and serviceability at the same time. Our reading is that, for a system vendor, the last three weigh more than module makers tend to assume.

Module makers ask how many watts they can save; system vendors ask how long they will be down when it breaks.

2. The Copper Bottleneck: a Few Meters at 200G, Under 1 m at 400G

Dell's view of copper is conservative and specific:

• 200 Gb/s per lane: passive-copper reach is typically limited to a few meters, depending on implementation.

• 400 Gb/s per lane and above: electrical reach is expected to shrink to under 1 meter.

• Scale-up domains are moving from a single rack to multiple racks; Jian Chen added verbally, "even multi-row."

That is the system-design challenge: short-reach copper favors compact single-rack topologies, while larger domains need longer reach, lower power, higher bandwidth density and manageable cable routing.

But Dell does not conclude that optics should replace copper everywhere. He ran a thought experiment: replace every copper cable in an NVL72 backplane with fiber plus traditional fully retimed transceivers, and it quickly stops being a sensible solution, with unacceptable power, cost and reliability trade-offs. The conclusion is a hybrid transition: copper where reach is still practical, because copper remains the industry's most reliable and cheapest medium, and optics where distance, density or power make copper inefficient.

▲ Dell's optical integration spectrum: from DSP pluggables and LPO pluggables to near-package NPX and co-packaged CPX, with optics moving ever closer to the ASIC (Image source: Dell | 2026 OCP Global Summit presentation)

Note the taxonomy: the slide defines NPX as "near-packaged optics or copper" and CPX as "co-packaged optics or copper." Our reading: to a system vendor, the spot next to the package is first an interface; whether it carries light or copper comes later.

3. Unpacking CPO: Five Costs, Three Unrelated to Physics

Dell accepts the benefits of CPO: shorter high-speed electrical channels, lower electrical I/O and retimer power, higher bandwidth density, and the potential for larger scale-up domains. The column that matters is the trade-offs:

• More complex co-design of package, thermals and board

• New manufacturing and test requirements

• Reduced field serviceability

• A larger "blast radius"

• Potential component and vendor lock-in

▲ Dell lists four benefits of CPO and five system-level costs, of which serviceability, blast radius and vendor lock-in are operational (Image source: Dell | 2026 OCP Global Summit presentation)

Jian Chen put the blast radius plainly: in today's pluggable world, when a link fails you swap one module; when a CPO link fails, you may need a chassis-level intervention, which lengthens mean time to repair (MTTR) and lowers overall system availability.

Jian Chen never used the term FRU. We translate his argument into one quantity: when a link fails, how big is the field-replaceable unit (FRU)? Each step the optics take toward the ASIC shortens the electrical channel, and makes the thing you replace one size bigger.

▲ The closer optics move to the ASIC, the shorter the electrical channel, but the unit to replace when a link fails grows from one module to the whole chassis; XPO and Open CPX both aim to keep things removable in between (STT's summary of Dell's argument) (Image source: Simple Tech Trend original; data from Dell, 2026 OCP Global Summit presentation)

4. Two Disaggregation Paths: XPO Holds the Faceplate, Open CPX Holds the Package Edge

Dell's answer is "optical disaggregation," and it names two paths.

XPO (eXtra-dense Pluggable Optics): Jian Chen said the MSA was formed in March 2026 around OFC, mainly promoted by Arista. It is still a pluggable module; fly-over cables separate the optics from the switch ASIC or XPU, it is liquid-cooled by design, and it can host low-power LPO as well as high-power coherent DWDM optics such as ZR/ZR+.

The slide cites Arista's numbers:

• XPO module 64 × 200G = 12.8T; OSFP module 8 × 200G = 1.6T

• 8x bandwidth and 4x front-panel density

Put the two together: bandwidth per module is 8x, front-panel density only 4x. If 4x means total bandwidth over the same faceplate, we estimate one XPO takes about 2x the faceplate of an OSFP (8 ÷ 4); this is our inference, not on the slide. XPO is not a smaller module; it packs more lanes into a shell that is still pluggable, then uses liquid cooling to carry the power.

▲ One XPO module carries 12.8T, 8x an OSFP, at 4x the faceplate density, implying each module takes about 2x the faceplate area (Image source: Simple Tech Trend original; data from Dell, 2026 OCP Global Summit presentation)

Open CPX: also launched around OFC in March 2026; Jian Chen said version 1.0 of the spec was published last month. It defines a pluggable optical engine connected through a standardized socket and electrical connector, rather than permanently attaching optics to the electronic package. It keeps most of CPO's benefits (short electrical channels, lower loss and retiming needs, high density near the package) while gaining multi-vendor interoperability and independent evolution of optics and electronics.

▲ Open CPX uses a standard socket to host pluggable optical engines in Type 1 and Type 2 placements, and the socket can also take copper cables (Image source: Dell | 2026 OCP Global Summit presentation; figure from the Open CPX MSA)

Item	CPO	XPO	Open CPX
Optics location	Permanently on the package	Front panel, fly-over cables to the ASIC	Standard socket next to the package
Replaced on failure	May need chassis-level intervention	One pluggable module	Optical engine in the socket
Cooling	Co-designed with package and board	Direct liquid cooling in the module	Not stated on slides
Spec status	Not listed on slides	MSA formed 2026/3	MSA launched 2026/3, v1.0 published

The split is clean: XPO handles "the faceplate is full," Open CPX handles "the electrical channel is too long," and both protect the same thing: when optics fail, you replace a module, not a machine. It is the direction we saw in ECOC2026 | Ciena Turns 6.4T Optical Engines Back Into Shelf Parts With Open CPX, only this time endorsed by a buyer.

5. Dell's Bet: Both MSAs, Plus a Silicon Photonics Project

Dell's commitment slide has three lines:

• Active member of both the XPO and Open CPX MSAs

• Founding member of OCP's Open Silicon Photonics for AI Systems, being developed under the OCP Server project

• Support for open, disaggregated optical solutions that preserve interoperability and serviceability

He added verbally that Dell is an early adopter of LPO with large-scale field deployment experience, but gave no deployment numbers.

Joining both MSAs looks like hedging, but our reading is that the logic is consistent: Dell does not care what shape the optics end up in, only whether they can be pulled out of the machine. Asked in Q&A about a new form-factor MSA announced that week, Jian Chen answered the same way: too early to say, more options are welcome, what matters is openness and interoperability.

Our reading: the CPO, NPO and XPO trade-offs (see 2026 OCP APAC Summit | CPO / NPO / XPO Panel) gain one more dimension in OEM hands: who owns the repair boundary. Hyperscalers that design their own racks and send their own technicians can absorb CPO's chassis-level interventions; OEMs that sell systems to others and provide after-sales service have to keep the repair boundary at module level.

6. Risks and Counterarguments

• A position without numbers: no pJ/bit, watts or cost in the whole talk, and Dell's "large-scale" LPO deployment is not quantified. The value of serviceability cannot be put on the same scale as CPO's power advantage.

• CPO reliability may shrink the repair problem: if soldered optical engines fail rarely enough, chassis-level interventions are rare too, and the blast-radius argument weakens. Dell offered no failure-rate data to counter this.

• Two MSAs fragment the ecosystem: XPO, Open CPX and the new form factor raised in Q&A; the more options, the smaller each one's volume, and the slower costs may fall.

• System vendors have limited leverage: GPU and switch-silicon vendors set scale-up architecture; if mainstream GPU platforms adopt CPO directly, OEM choices narrow.

7. What It Means for Taiwan's Supply Chain

• Value shifts to the "removable interface." If Dell's criterion becomes the OEM consensus, sockets, electrical connectors, fly-over cables and liquid-cooled module housings become new key parts. In a separate talk at the same OCP (not Dell's), Hon Hai's Foxconn Interconnect Technology (FIT) presented a liquid-cooled XPO module architecture, so Hon Hai's connector arm is already active on this path.

• System assembly and test flows will be rewritten. Moving from "swap a module" to "swap the optical engine in a socket" redraws the line between factory test and field repair; Taiwanese contract manufacturers assembling systems for OEMs may need to build up optical validation capability.

• Open CPX sockets also take copper (the Type 1 drawing shows a copper cable connector), and with Dell arguing for a hybrid transition, copper cable makers will not be pushed out of scale-up overnight.

8. Conclusion

The most valuable thing about this talk is that it is the buyer's voice. Module makers use pJ/bit to convince each other; Dell uses MTTR and blast radius to convince customers, and it is the latter who sign the purchase orders.

Our call: in the 400G/lane generation, scale-up optics in the OEM channel will run on two tracks, "XPO at the faceplate, Open CPX at the package edge," with soldered CPO left to the few hyperscalers that can own their repair boundary. Over the next 12 months, watch three things: results of Open CPX multi-vendor interoperability testing, the production timeline for liquid-cooled XPO modules, and Dell's first product built on either MSA.

For a system vendor, the best optics are not the most power-efficient ones, but the ones easiest to replace when they break.

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