Co-Packaged Optics (CPO) 2026: The Complete Technical Guide

Co-packaged optics (CPO) moves the optical engine from a front-panel transceiver onto the switch chip package. In 2026 it crossed from demonstration to shipping product. This guide explains what CPO is, how a CPO switch is built, what it saves, what it costs, and where it fits. The point most coverage skips is the one that matters to anyone who buys optics: CPO removes the transceiver from the switch, not from the network. The host side still needs pluggable transceivers, in-rack links still use copper, and the laser moves into its own replaceable module. The guide is written for the engineers who specify, qualify and approve those parts.

1. What Is Co-Packaged Optics?

Co-packaged optics (CPO) places the optical engine, the part that converts electrical signals into light, on the switch chip package instead of in a removable transceiver on the front panel. The fiber connects at the chip. That one change is the whole idea, and it is easiest to see against the way switches are built today.

In a conventional pluggable switch, the switching chip (ASIC) sits in the middle of the board and the optics plug into the front panel. Between them run inches of circuit board, a connector and the module. At today's lane speeds that electrical path loses much of the signal, so a digital signal processor (DSP) inside every module restores it, and the DSP consumes power.

In a CPO switch, the optical engine sits millimeters from the ASIC. The electrical path almost disappears, the DSP is no longer needed, the laser moves to a separate replaceable module, and only fiber leaves the switch. The result is far less power per bit and more bandwidth in the same faceplate, which is what the largest AI fabrics have run out of.

The core difference: CPO moves the optical engine off the faceplate and into the ASIC package.

CPO Terms Used in This Guide

The term covers two different designs, and most of the confusion about CPO comes from mixing them up.

Two Kinds of Co-Packaged Optics: Switch CPO vs GPU-Side CPO

Everything in this guide concerns switch CPO unless it says otherwise.

2. CPO Switches and the Optics They Pair With

CPO switches are available for both InfiniBand and Ethernet fabrics: NVIDIA's Quantum-X Photonics (InfiniBand) and Spectrum-X Photonics (Ethernet) platforms, and Ethernet systems built on Broadcom's Tomahawk 5 Bailly and Tomahawk 6 Davisson silicon. The same switch generations also ship with OSFP cages for pluggable modules, so CPO is a choice made per build, not a requirement of the platform. None of these switches removes the need for transceivers. They move the transceiver from the switch to the server side.

CPO moves the optical engines into the switch package, while pluggable transceivers remain on the host side.

What each kind of switch means for the optics you buy:

• InfiniBand CPO switches connect to server NICs that carry 800G OSFP pluggable transceivers. The fabric decides the NIC and the host optic that pairs with it (see InfiniBand vs Ethernet for AI clusters).
• Ethernet CPO switches connect to servers that carry 400G or 800G DR4 and FR4 pluggable transceivers in OSFP, QSFP112 or QSFP-DD. 1.6T modules (DR8, 2×FR4) sit on switch-to-switch links today, not at the NIC.
• On the switch itself, the front panel carries fiber connectors, either MPO or the smaller MMC, and replaceable laser modules. There are no transceiver cages.
• Short links are unchanged. Inside the rack, DAC, ACC and AEC copper cables serve GPU-to-GPU and server-to-leaf links, and AOC serves slightly longer in-row runs.
• Cooling depends on the system. NVIDIA's CPO switches are liquid-cooled; some Broadcom-based 51.2T systems are air-cooled. Plan the switch row before the optics.

3. CPO vs Pluggable, LPO, NPO and XPO

CPO is one end of a spectrum, not a single choice. Each step moves the optics closer to the ASIC, shortens the electrical path and lowers power, and each step gives up some of the serviceability that makes a pluggable module convenient.

As the optics move toward the ASIC, power per bit falls and field serviceability changes.

Switch Optics Compared: Power, Density, Cooling and Serviceability

CPO moves both power and density. XPO moves density. LPO moves power. Serviceability falls as the optics move closer to the chip.

Each approach in plain terms:

• Pluggable (DSP) is today's default. The module slots into the front panel with a DSP inside to restore the signal after the long electrical run. It is hot-swappable in minutes and sourced from any vendor. The cost is power, roughly 15–25 W at 800G.
• Linear pluggable optics (LPO) keeps the same module and form factor but removes the DSP; the switch ASIC drives the optic directly. Power roughly halves while hot-swap and multi-vendor sourcing remain. The trade is tighter host compatibility, so LPO is qualified against the specific switch and NIC. LRO keeps a DSP on the transmit side only.
• XPO is a liquid-cooled pluggable module that carries eight 1.6T links (64 lanes, 12.8 Tb/s) behind one cold plate rated for up to 400 W, at about four times the faceplate density of OSFP. It keeps the pluggable service model at the price of a liquid loop to the switch row and a new switch design. A second high-density pluggable design, USFP, backed by NVIDIA and Cisco and built for 400G electrical lanes, is also in development. Neither ships yet.
• NPO (near-package optics) moves the optical engine off the faceplate to a socket beside the ASIC. The electrical path shortens sharply and the engine stays replaceable in the field. The Open CPX MSA defines a common socket for CPO and NPO modules.
• CPO (co-packaged optics) puts the engine on the ASIC package itself. It has the lowest power, roughly 4–5 W at 800G, and the highest density, but repair is board-level rather than a module swap. The laser module stays replaceable.

For most networks the practical path is pluggable, then LPO, with XPO or CPO where faceplate density or power per bit becomes the limit.

4. Inside a CPO Switch

You do not need to design the silicon to evaluate it, but it helps to know what sits where. A co-packaged switch has five parts that matter.

A simplified CPO switch: optical engines sit beside the ASIC, fiber exits the package, and the laser stays in a replaceable front-panel module.

• Switch ASIC. The switching chip. It drives the optical engines over a path only millimeters long, so it needs simpler, lower-power circuitry than a pluggable design.
• Optical engines. The photonics that convert between electrical signals and light, beside the ASIC. They do the job a pluggable module used to do, without a DSP, because the short path no longer needs one.
• Fiber shuffle. The internal routing that gathers the fibers from the engines and organizes them toward the faceplate.
• Faceplate connectors. Where single-mode fiber leaves the switch, as MPO or MMC connectors. There are no transceiver cages.
• External laser source (ELS). The laser, kept out of the package in a front-panel module. It is the most failure-prone part, so keeping it replaceable matters.

5. Optics Used in a CPO System

This is the question that matters most to anyone who specifies optics, and it is where CPO is most misunderstood. CPO removes the transceiver at the switch. It does not remove optics from the network. A CPO fabric still contains pluggable optics and copper; they move to where the switch no longer carries them.

CPO removes the switch-side module. The optic you specify, qualify and order moves to the host and NIC side.

Three things are true of every CPO deployment:

• The host and NIC side still uses pluggable transceivers. CPO switches present single-mode fiber on the faceplate and connect to servers whose network cards carry pluggable optics, typically 400G and 800G FR4 (1310 nm, 2 km, duplex LC) and DR4 (500 m, MPO-12).
• Short links still run on copper. Inside the rack, GPU-to-GPU and adjacent-rack connections use DAC, ACC and AEC cables, the lowest-power choice up to a few meters (see the 800G interconnect selection guide).
• The switch needs fiber and lasers. The faceplate uses dense single-mode fiber and replaceable laser modules. Check whether the switch presents MPO or the smaller MMC connector before ordering trunks and patch cords; the fiber plant has to match the faceplate.

Quick reference: which optic serves each position in the fabric.

Net effect for buyers: CPO does not shrink the importance of choosing the right transceiver. It relocates it. The qualification work moves to the host and NIC side, where reach, connector, fiber and compatibility still have to be right.

6. Benefits of Co-Packaged Optics

The case for CPO is led by power, and the saving is large enough to justify rebuilding the switch. Three things drive it: the short electrical path lets the switch use simpler circuitry, removing the DSP removes a major power draw, and sharing a few lasers across many engines is more efficient than a laser in every module.

Representative power per 800G port at three boundaries: optics only, the whole switch, and the full cluster.

How Much Power Does CPO Save?

Most of a pluggable link's power goes to driving the long electrical path and the DSP at each end. Co-packaging shortens the path to millimeters. In NVIDIA's figures for a 1.6T link, channel loss falls from about 22 dB to about 4 dB and per-link power from about 30 W to about 9 W. Feeding many engines from a few shared lasers removes roughly three-quarters of the laser count on top of that. Independent measurements put the optical power saving at roughly two-thirds to three-quarters per port. Treat the exact multipliers as vendor figures against favorable baselines; the direction is corroborated.

Does CPO Need Liquid Cooling?

Not by itself. Removing the module DSPs removes the faceplate's largest heat source, which is why a first-generation 51.2T CPO switch ships air-cooled in four rack units. Most higher-density CPO switches are liquid-cooled because their designers spent the freed faceplate on density, not on staying on air. CPO changes how much heat there is to remove; whether it leaves on air or liquid is a density decision for the switch row (see the 800G thermal and power planning guide).

Does CPO Increase Switch Density?

Yes, but density is no longer exclusive to CPO. Clearing the transceiver cages frees faceplate area, and a 102.4T CPO switch fits in two rack units. A liquid-cooled pluggable switch of the same capacity can also reach two rack units, and XPO claims about four times OSFP density in a module you can still pull. What stays exclusive to CPO is the power per bit and the shared lasers.

CPO Benefits Compared With Pluggable Optics

Every benefit traces to one cause: the short electrical path removes the DSP, the loss and the per-module laser at once.

7. CPO Challenges and Trade-offs

CPO is not free. An honest evaluation weighs each gain against what it costs to live with, and the magnitude is what decides the call.

Can You Repair a CPO Switch in the Field?

A failed pluggable is a two-minute swap by any technician, with a spare from any vendor. When the optic is on the package, the unit of replacement becomes the line card or the whole switch sled, so the spare you stock is a sled holding hundreds of fibers rather than a module, and shipping one across regions is slow and costly. The laser module is the exception: it is designed to be replaced from the front panel. Confirm your vendor's service policy before you count on that, because some vendors reserve the laser swap for their own service staff.

What Happens When an Optical Engine Fails?

Early reliability data favors CPO. In Meta's laboratory testing of Broadcom CPO, one million port-hours ran without a link flap, because removing the pluggable connector removes a common failure point. The concern is the failure domain. One engine serves several ports, so a failure takes down several links instead of one, and in a synchronized training job a stalled link can idle a large slice of the cluster. Operators therefore plan 5–10% spare ports and route around a dead lane rather than rush to repair it.

Does CPO Lock You Into One Vendor?

At the switch tier, yes. With pluggables you mix optics vendors and change reaches independently of the switch. In a CPO system the optical engines are qualified as part of the platform, so that leverage disappears at that tier. The host side keeps its flexibility, which is why buyers retain multi-vendor sourcing there.

Is CPO More Expensive Than Pluggable Optics?

Per port, first-generation CPO still costs more than high-volume pluggables, and the gap narrows only as volume grows. TrendForce estimates co-packaged optics at about 0.5% of AI data center optical modules in 2026, rising toward 35% by 2030. CPO starts where the power wall is hardest and the operator is most sophisticated. Pluggable and LPO remain the default everywhere else.

XPO changes the serviceability trade rather than removing it. One XPO module carries eight 1.6T links, so a swap darkens eight links at once, but the repair unit is a module and minutes rather than a sled and days.

8. CPO Applications and Use Cases

CPO arrives first where density and power pressure are greatest and spreads outward. In 2026 that means the leaf and spine networks of large AI training and inference clusters at hyperscalers and neoclouds. Where buyers will meet it, and what it means for their optics:

• AI training and inference fabrics. The leading use. CPO leaf and spine switches pair with pluggable optics on every server and copper inside every rack.
• Cloud and colocation Ethernet. Hyperscalers first. Broader adoption waits on serviceability and multi-vendor sourcing.
• HPC and research. Follows the AI fabric playbook where halls are power-constrained.
• Telecom transport. Uses a different, coherent form of co-packaging on a slower roadmap.
• Enterprise and edge. Pluggable and LPO remain the answer. CPO is rarely justified at this scale.

The adoption roadmap runs top-down: CPO leads at the top of the largest fabrics and thins out toward enterprise, where pluggable optics stay the right call.

9. CPO in AI Data Centers

Almost all CPO deployment in 2026 is in AI data centers, so it is worth being precise about where in an AI fabric it sits, because that decides which optics you buy. An AI fabric has two networks, and CPO sits in only one of them today.

• Scale-up, the GPU-to-GPU fabric inside the rack. NVLink-class links between GPUs remain copper in current systems. Co-packaged optics for GPU-to-GPU links is a later step.
• Scale-out, the leaf and spine network between servers. This is where every CPO switch available today sits. The switch side becomes fiber; the server side keeps pluggable optics. It is also where multi-vendor sourcing and field serviceability matter most, which is why CPO adoption here is selective.

Where Does CPO Sit in Hopper, Blackwell and Rubin Clusters?

The GPU generation sets the NIC speed, the NIC speed sets the host optic, and the switch decides whether CPO is in the picture at all.

CPO by GPU Generation: Hopper, Blackwell and Rubin Clusters

Through Vera Rubin, the module at the NIC stays 800G. CPO changes only the switch end of the link.

NVIDIA has said the generation after Vera Rubin adds co-packaged optics to the NVLink fabric for links between racks; links inside the rack stay copper. Until then, co-packaged optics in an NVIDIA cluster means the leaf and spine switch, and the part you qualify is the 800G module at the NIC. AMD and other Ethernet clusters follow the same pattern: copper scale-up inside the rack, 400G to 800G Ethernet scale-out, pluggable optics at the NIC, and CPO available only on the switch.

The headline is that AI data centers are adopting CPO. The buyer's reality is narrower: CPO is a scale-out switch technology, and most of an AI fabric's optics remain pluggable transceivers and copper on the host side. For the platform-by-platform picture of which optic each NVIDIA system takes, see the 800G transceiver compatibility guide for NVIDIA platforms, and the OSFP IHS vs RHS selection guide for getting the heat-sink type right before you order.

10. Example: An 800G CPO Fabric

Consider one leaf switch with 128 ports of 800G connecting to GPU servers, built two ways.

A CPO leaf build: the switch side changes, the host side keeps its pluggable optics.

In the pluggable build, every switch port and every server port carries an 800G transceiver, so each link has a module at each end: fully serviceable and multi-vendor, with a hot, dense faceplate. In the CPO build, the switch faceplate carries only fiber and replaceable laser modules. The host side is unchanged: each server still needs its pluggable optic, a 400G or 800G FR4 or DR4, and in-rack links still run on copper. A move to CPO changes where you buy optics, not whether you buy them. The host-side transceiver becomes the part to qualify most carefully.

11. Operating a CPO Network

CPO changes three day-to-day practices, and all three are worth planning before the first switch arrives.

• Commissioning. Much of a CPO system is assembled and tuned at the factory, so deployment depends more on vendor coordination and acceptance testing than on bench work in your own lab.
• Monitoring. A degraded lane cannot be swapped on its own, so per-lane telemetry (laser current, photodiode temperature, engine status) has to feed your monitoring, and troubleshooting shifts toward catching problems early.
• Spares. Because a failed engine takes down more than one port, provision 5–10% spare ports and keep a board-level RMA workflow. Stock laser modules only if your vendor allows customer replacement; NVIDIA's documentation for its InfiniBand CPO switch instructs operators not to service the laser module themselves.

Migration planning should cover the fiber plant as well: trunks, patch panels and the connector type at the CPO faceplate, MPO or MMC. The host-side optics, by contrast, are operated exactly as they are today: hot-swappable, monitored through the same tooling, and sourced from the vendor of your choice.

12. Selecting Optics for CPO

Whether the next switch is pluggable, LPO, XPO or co-packaged, the decision comes down to matching each link to the right part. Start with the pressure: is faceplate power or density the limit? If not, stay on pluggable or LPO. If density alone is the limit and the switch row can go liquid, XPO solves it without leaving pluggable modules. Then serviceability: can you absorb board-level repair and a larger failure domain? Then sourcing: can you accept single-vendor optics at that tier? CPO is the answer only when all three hold.

A simple way to decide what belongs on each link.

Once the part is chosen, four checks apply to every link. They matter more in a CPO fabric, where the switch-side DSP is gone and there is less margin to mask a marginal connection:

• Confirm reach against the real distance. DR4 at 500 m and FR4 at 2 km are not interchangeable.
• Match the connector and fiber to the port, and inspect every end face.
• Enable the correct host FEC.
• Qualify any LPO against the specific switch and NIC, because its behavior depends on the host.

A pre-deployment checklist for qualifying optics in a CPO fabric.

About this guide. Updated on October 15, 2026.