An AI data center can have an extraordinary amount of computing power and still spend too much time waiting. Thousands of accelerators exchange data as they work on the same model. When communication becomes the bottleneck, adding more processors does not necessarily deliver a proportional increase in useful work.

That puts the network under pressure. It needs to move more data, maintain reliable connections, and do both within a finite power and cooling budget.

Co-packaged optics, or CPO, addresses a small but increasingly expensive part of that problem: the electrical connection between a network switch chip and its optical transceivers. It moves optical conversion onto the chip’s package, shortening the distance that high-speed electrical signals travel.

The idea sounds like a packaging detail. At data-center scale, packaging helps determine how much bandwidth fits in a system, how much electricity it consumes, and how technicians maintain it.

What are co-packaged optics?

In a conventional optical switch connection, data leaves the switch’s application-specific integrated circuit, or ASIC, as an electrical signal. It travels across a circuit board to a pluggable optical transceiver at the front panel. The transceiver converts that signal into light for transmission through fiber. Incoming traffic follows the reverse process.

CPO places optical engines beside the switch ASIC on a shared package. Electrical signals travel a much shorter distance before becoming optical signals. Fiber carries the data onward from there.

Two terms often appear together here, but they describe different things. Silicon photonics describes a technology for integrating optical functions on chips. Co-packaged optics describes how optical engines and processing silicon are packaged together. Silicon photonics can also appear inside conventional pluggable modules.

The switch still processes packets electronically. CPO does not turn it into an all-optical computer, and it does not replace Ethernet or InfiniBand. It changes the physical implementation of the connection.

Why the electrical path becomes a problem

As electrical signaling rates increase, circuit-board traces and connectors become harder to work with. Signal attenuation, reflections, and interference leave less margin for distinguishing one transmitted symbol from another. Designers use increasingly capable transmitters, receivers, equalization, and signal processing to recover the data.

That engineering consumes power and produces heat. Shortening the electrical path reduces the burden on those circuits. It also allows designers to arrange optical connections around a package instead of relying entirely on electrical routes to front-panel modules. Cisco’s CPO demonstration illustrates both the power opportunity and the challenge of making optical components small enough for package integration.

The benefit comes from making the connection easier to drive efficiently. An optical link does not automatically fix congestion, poor topology, oversubscribed uplinks, or a slow application.

Why data-center operators care

Power savings accumulate across thousands of connections

NVIDIA’s January 2026 technical description claims a fivefold reduction in power per 1.6 Tb/s port compared with its pluggable-interconnect baseline. That is a vendor comparison for a particular implementation and measurement boundary. It is not a claim that an entire data center consumes five times less electricity.

For a sense of scale, consider a hypothetical deployment with 10,000 optical endpoints. A reduction of 10 watts per endpoint equals 100 kilowatts of continuous electrical load, or 876 megawatt-hours over a year of continuous operation. An endpoint means one end of a link, so a connection with two upgraded ends counts twice.

Those figures are arithmetic, not a product benchmark. They exclude cooling effects and assume the full reduction persists all year. They show why a seemingly small per-port improvement attracts attention in a large facility.

Bandwidth density affects the whole network design

A switch’s useful capacity depends on its connections as well as its internal silicon. More bandwidth leaving each package can support denser systems and give architects more flexibility in constructing a network.

The Ethernet Alliance’s 2026 roadmap includes 800G interconnects and emerging 1.6 Tb/s Ethernet, alongside improvements in optical and copper technology. These advances make electrical efficiency and optical packaging increasingly relevant. Ethernet Alliance’s 2026 roadmap

Keep three numbers separate when reading announcements: lane speed, port speed, and total switch capacity. A 102.4 Tb/s switch does not provide 102.4 Tb/s to every server. Its capacity is distributed among ports, each of which can combine multiple signaling lanes.

Reliability includes the time it takes to recover

Fewer electrical interfaces can remove potential failure points, but tighter integration changes what happens when something fails. The questions become practical: Which component is replaceable? How many links share it? How long does replacement take? Can traffic continue on another path?

Research on CPO integration identifies packaging, thermal behavior, and serviceability as important determinants of deployment, alongside photonic device performance.

A reliable data center needs both low failure rates and manageable recovery procedures.

The tradeoffs: optics move closer, maintenance changes

A front-panel pluggable module has an operational advantage: a technician can replace it without accessing the switch package. CPO makes the optical engine part of a more integrated assembly. Depending on the design, an engine failure can require replacement of a larger unit.

Thermal design also becomes more demanding when temperature-sensitive optical components sit near a hot ASIC. Manufacturing needs precise assembly and effective testing before expensive components become part of the final package. These are system-design challenges, not details that disappear once an optical link works in a laboratory.

One useful design choice is an external laser source. The laser supplies continuous light to an optical engine, where modulation encodes the data. Keeping the light source separate can make it independently replaceable.

The Optical Internetworking Forum’s ELSFP specification defines a pluggable external-laser form factor for this purpose. Its January 2025 revision describes field-replaceable modules that feed co-packaged optical transceivers. That improves serviceability for the laser; it does not mean every optical engine is also hot-swappable.

The product roadmap: what is actually here?

As of September 27, 2026, CPO includes shipping and production announcements. The evidence supports a more precise statement than either “it is still a lab experiment” or “every data center is switching.”

Technology or platform	Public milestone	What that tells an operator
Broadcom Tomahawk 6–Davisson	Broadcom’s October 8, 2025 announcement reports shipping its third-generation CPO Ethernet switch, with 102.4 Tb/s aggregate capacity and 200 Gb/s channels.	Commercial CPO silicon exists. The announcement alone does not establish deployment volume or availability of every partner system.
NVIDIA Spectrum-X Ethernet Photonics	NVIDIA’s May 31, 2026 announcement describes the CPO switches as in production, within the Vera Rubin platform rollout.	Manufacturing is progressing beyond a future-only roadmap. Installation schedules and system availability still require product-specific confirmation.
Intel optical compute interconnect	Intel’s June 2024 demonstration shows an optical I/O chiplet co-packaged with a CPU, supporting up to 4 Tb/s bidirectional transfer.	Optical integration also targets compute packages. A demonstration is not evidence of broad commercial deployment.

Sources: Broadcom’s Davisson announcement, NVIDIA’s production announcement, and Intel’s optical I/O demonstration.

These milestones also point to two distinct applications. Scale-out connects servers and racks through a network. Scale-up connects accelerators more tightly within a computing domain. CPO on a network switch and optical I/O on a processor package address related physical challenges, but they have different protocols, packaging requirements, and adoption schedules.

The standards roadmap: several pieces advance together

There is no single specification that makes every CPO system interchangeable. Three layers deserve attention:

Layer	Relevant work	Why it matters
Ethernet link technology	IEEE P802.3dj covers 200 Gb/s, 400 Gb/s, 800 Gb/s, and 1.6 Tb/s Ethernet work. Its public record includes draft-ballot activity in September 2026.	Defines link behavior and interfaces. It does not require a particular CPO package.
Co-packaged module interfaces	OIF’s 3.2 Tb/s CPO module implementation agreement provides a defined module framework.	Helps align component interfaces and integration. It is not proof that all vendors’ packages are interchangeable.
External laser supply	OIF’s ELSFP agreement defines an external, replaceable laser module.	Supports a serviceable optical supply component and a more consistent supplier ecosystem.

Sources: IEEE P802.3dj project, IEEE draft-ballot record, OIF CPO module agreement, and OIF ELSFP agreement.

Standards progress, component shipments, and operational readiness are separate milestones. A procurement plan needs evidence for all three.

CPO shares the future with better pluggable optics

Pluggable optics continue to improve. Linear pluggable optics, or LPO, reduces module signal-processing complexity while retaining a removable front-panel form factor. The tradeoff includes tighter requirements on the host electrical channel and its qualification.

Cisco’s February 2026 portfolio announcement includes both 800G LPO and 1.6T OSFP pluggable optics. That is a useful reminder that the industry’s roadmap contains competing approaches to power, density, and maintenance. Cisco’s systems and optics announcement, Cisco’s technical overview of optical implementation choices

My assessment is that coexistence remains the most useful planning assumption. A large AI fabric can justify an integrated optical design while another part of the same facility benefits from conventional pluggables or short copper connections. The appropriate choice depends on reach, power, density, serviceability, and cost.

A practical adoption roadmap for data centers

The following is a planning framework, not a vendor delivery forecast.

Now: establish the baseline. Measure actual network power, link failures, application throughput, and time lost to communication. Record fiber distances, connector types, and cooling constraints. Compare CPO against the best suitable pluggable alternative at the same capacity and reach.

Next 12–24 months: qualify a bounded deployment where the economics justify it. A new AI cluster or a network expansion offers a useful evaluation boundary. Test application behavior under load, optical telemetry, failure isolation, replacement procedures, and recovery time. Include external lasers, cooling, spares, and support in the cost comparison.

2028 and beyond: expand according to operational evidence. Watch for repeatable manufacturing, broader system choices, compatible components, and field reliability data. Track optical I/O near CPUs and accelerators separately from switch CPO. Treat wider adoption as conditional on those results rather than assigning a universal retirement date to pluggable optics.

The purchasing question is specific: Does this architecture deliver more useful application work within our power, space, cost, and recovery constraints?

CPO earns attention because moving the conversion point by a few centimeters can change the economics of thousands of connections. The long-term opportunity is a data center that spends less energy moving information and puts more of its infrastructure to productive use. Its success depends as much on reliable manufacturing and maintainable systems as on the optical technology itself.