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

Co-packaged optics is the biggest change to switch design in a decade, and in 2026 it crossed from demo to shipping product. This guide explains what CPO is, the switches available now, how a CPO system is built, and the benefits and tradeoffs that decide where it fits. Its through-line is the one point most coverage skips: CPO doesn't remove optics from your network — it relocates them to the host side, which changes what you specify, qualify, and order. It's written for the engineers who make those calls.

1. What is co-packaged optics?

Co-packaged optics (CPO) puts the optical engine — the part that converts electrical signals into light — directly onto the switch chip's own package, instead of in a removable transceiver on the front panel. The fiber connects right at the chip. That one move is the whole idea, and it's 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 runs inches of circuit board, a connector, and the module — a long electrical path. At today's lane speeds that copper path loses a lot of signal, so a digital signal processor (DSP) inside every module has to clean the signal up, and that burns power.

In a CPO switch, the optical engine sits millimeters from the ASIC, so the electrical path almost disappears. The DSP is no longer needed, the laser is moved to a separate replaceable module, and only fiber leaves the box. The payoff: far less power per bit and much more bandwidth in the same faceplate — which is exactly 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.

A few terms recur throughout this guide. Here is the working vocabulary in one place:

In short: these four terms — optical engine, silicon photonics, ELS, and fiber shuffle — are the building blocks every CPO switch is described with.

2. Co-packaged optics switches in 2026

CPO is no longer a research project. As of mid-2026 there are co-packaged switches shipping, and named platforms with firm availability windows. What matters for planning isn't just the switch — it's the fabric it runs (InfiniBand or Ethernet) and the optics it pairs with on the host side.

Takeaway: co-packaged switches are real and shipping in 2026 across both Broadcom Ethernet and NVIDIA InfiniBand/Ethernet lines — but every one of them still pairs with pluggable optics on the host side.

Reading the landscape, three points matter for an optics buyer:

• Fabric splits the field. NVIDIA ships CPO in both flavors — Quantum-X for InfiniBand, Spectrum-X for Ethernet — while Broadcom's Tomahawk line is Ethernet. The fabric decides the NIC and the host optic that pairs with it. (See our guide on InfiniBand vs Ethernet for AI clusters.)
• The faceplate is fiber; the optics are on the host. Every one of these switches presents single-mode fiber, not transceivers. The transceivers you buy — 400G/800G/1.6T FR4 and DR4 in OSFP or QSFP-DD form factors — live on the server and NIC side.
• Short reach stays copper. Inside the rack, the same fabrics use DAC, ACC, and AEC cables for GPU-to-GPU links, and AOC for slightly longer in-row runs — CPO doesn't change that.

Not everyone is rushing: some major switch vendors remain cautious on the reliability of attaching a thousand-plus fibers inside the package, and favor linear pluggable optics for now. That split is a useful signal — CPO is real where the power problem is hardest, and a measured wait elsewhere.

3. CPO vs pluggable, LPO and NPO

CPO is best understood as 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 — while giving up some of the easy serviceability that makes pluggables so convenient.

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

The pattern down the table is the whole story: each step removes more of the DSP and cuts power, but trades away the simple hot-swap serviceability of a pluggable.

Linear pluggable optics (LPO) deserves attention because it captures much of CPO's power saving while keeping the familiar, hot-swappable, multi-vendor module — it simply removes the DSP and lets the switch drive the optic directly. For most networks the practical path is pluggable, then LPO, then selective CPO — not a jump straight to co-packaging.

Here is each approach in plain terms, from the most familiar to the most integrated:

• Pluggable (DSP) — today's default. A transceiver that slots into the front panel, with a DSP inside to clean up the long electrical run from the ASIC. Maximum flexibility: hot-swappable in minutes, sourced from any vendor. The cost is power — roughly 15–25 W at 800G.
• LPO (linear pluggable optics) — the power bridge. Same pluggable module and form factor, but the DSP is removed; the host switch ASIC drives the optic directly. Roughly half the power, while keeping hot-swap and multi-vendor sourcing. The trade is tighter host compatibility, since the signal conditioning now lives in the switch.
• LRO (linear receive optics) — the halfway step. A linear variant that keeps a DSP on the transmit side only and runs the receive side linear. It's a hedge for links where fully removing the DSP is too aggressive but some of the power saving is still wanted.
• NPO (near-package optics) — optics beside the chip. The optical engine moves off the faceplate to beside the ASIC on the same board, cutting the electrical path sharply but keeping the engine socketed, so it stays partly serviceable. Often a stepping stone to full co-packaging.
• CPO (co-packaged optics) — optics on the package. The engine sits on the ASIC's own package, millimeters away. Lowest power (~4–5 W at 800G) and highest density, but repair is board-level rather than a module swap. The laser stays replaceable as a separate front-panel module.

4. Inside a CPO switch

You don't 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 hug 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 now only millimeters long, so it needs far simpler, lower-power circuitry than a pluggable design.
• Optical engines — the photonics that convert between electrical signals and light, sitting right beside the ASIC. They do the job a pluggable module used to do, but without a DSP, because the short path no longer needs one.
• Fiber shuffle — the internal routing that gathers the many fibers coming off the engines and organizes them toward the faceplate.
• Faceplate connectors — where single-mode fiber leaves the switch. Note what's missing: no transceiver cages, just fiber.
• External laser source (ELS) — the laser, kept out of the package in a front-panel module you can replace like a pluggable. It's the most failure-prone part, so keeping it serviceable matters.

5. Optics used in a CPO system

This is the question that matters most to anyone who specifies optics, and it's 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 plenty of pluggable optics and copper — they move to where the switch no longer carries them.

CPO removes the switch-side module — but the optic you spec, qualify, and order moves to the host / NIC side.

Three things are true of a 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/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, most reliable choice up to a few meters. (See the 800G interconnect selection guide for DAC vs AEC vs AOC vs optical.)
• The switch needs fiber and lasers. The faceplate uses dense single-mode fiber and replaceable ELS laser modules.

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

6. Benefits of co-packaged optics

The case for CPO is led by power, and the savings are large enough to justify rebuilding the switch. Three things drive it: the short electrical path lets the switch use simpler, lower-power 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. Across a full switch, this is the difference between staying on air cooling and needing liquid.

The mechanism behind the numbers is worth understanding, because it's what makes the saving real rather than marketing:

• The electrical path is the villain. Most of a pluggable link's power goes to fighting electrical loss, which forces a power-hungry DSP at each end. Co-packaging shrinks the path to millimeters, so that loss largely disappears and the DSP isn't needed — in NVIDIA's figures for a 1.6T link, channel loss drops from roughly 22 dB to about 4 dB and per-link power from about 30W to about 9W.
• Shared lasers compound the saving. Feeding many engines from a few lasers, instead of one per module, removes roughly three-quarters of the laser count on top of the DSP saving.
• At switch scale it changes the cooling. Across a 102.4T switch this is the difference between a faceplate you can air-cool and one you can't; in a 400,000-GPU build, vendor figures put the interconnect saving in the tens of megawatts. (More on this in our 800G thermal and power planning guide.)
• Density is a bonus. Clearing the transceiver cages frees faceplate area, letting a high-radix switch flatten a three-layer network toward two.

Treat the specific multipliers as vendor figures against favorable baselines — but the direction, roughly two-thirds to three-quarters less optical power per port, is independently corroborated.

Every benefit here traces back to one root cause — the short electrical path — which is what removes the DSP, the loss, and the per-module laser all at once.

7. CPO challenges and tradeoffs

CPO isn't free, and an honest evaluation weighs each gain against what it costs to live with. The bullets below are the headline; the magnitude is in the paragraphs that follow, because the magnitude is what actually decides the call.

Serviceability is the real change, and it's bigger than it looks. A failed pluggable is a two-minute swap by any technician, with a spare transceiver from any vendor. When the optic is fused into the package, the unit of replacement becomes the line card or the whole switch sled — so the spare you stock is no longer a $-hundreds module but a sled holding hundreds of fibers, and shipping one across regions is slow and costly. This is why hyperscalers plan to overbuild ports by 5–10% and keep hot spare switches: a dead lane gets routed around, not rushed to repair.

The failure domain, not the failure rate, is the worry. Counterintuitively, the early field data cuts in CPO's favor — Meta reported fewer failures with co-packaged optics than with pluggables, because removing the pluggable connector removes a common failure point. The concern is blast radius: when something does fail, it can take several ports with it instead of one, and in a synchronized training job a single stalled link can idle a large slice of the cluster. That's why vendors are chasing a failure rate below 0.1 FIT with repair measured in hours before operators will accept the larger domain.

Lock-in and cost are the remaining two. 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 goes away at that tier. And while the power math is compelling, first-generation CPO still carries a higher per-port cost than high-volume pluggables — the equation improves as volume grows, but it isn't there yet. None of this argues against CPO. It explains why, in 2026, co-packaged optics is roughly 0.5% of optical modules in AI data centers (projected toward 35% by 2030): it starts where the power wall is hardest and the operator is most sophisticated, and pluggables and LPO stay the default everywhere else.

8. CPO applications and use cases

CPO is arriving where density and power pressure are greatest, then spreading outward. The pace below reflects deployment reality in 2026, not just technical readiness.

Where CPO fits, and how fast each sector is likely to adopt it.

AI and ML fabrics are the primary driver, but the distinction that matters is where in the fabric. CPO lands first on the densest, shortest, most homogeneous links — the scale-up tier knitting GPUs together within and across racks, where bandwidth-per-rack is everything and the links are uniform enough to commit to one vendor. The broader scale-out leaf/spine network adopts more slowly, because that's where multi-vendor sourcing and field serviceability are hardest to give up.

Beyond AI, the picture varies by sector — what each runs, and whether CPO's tradeoffs make sense, differ widely:

Read top to bottom, the table tracks CPO's spread: it leads in AI training where the power wall bites hardest, and thins out toward enterprise/edge where pluggable and LPO remain the right call.

The throughline: in 2026 CPO is a top-of-fabric, highest-pressure-link technology — about 0.5% of AI-data-center optical modules today, concentrated exactly where the power wall is hardest, and spreading outward only as the ecosystem matures.

9. CPO in AI data centers

Almost all CPO deployment in 2026 is in AI data centers, so it's worth being specific about where in an AI fabric it actually lands — because the answer decides which optics you buy and which you don't. AI fabrics have two distinct networks, and CPO treats them very differently:

• Scale-up — the GPU-to-GPU fabric (where CPO lands first). This is the dense, high-bandwidth mesh tying GPUs together within and across a few racks — NVLink-class today. The links are short, uniform, and bandwidth-per-rack is everything, so the power and density win is largest and committing to one vendor is acceptable. This is the leading edge of CPO.
• Scale-out — the leaf/spine cluster network (where pluggables hold on). This is the broader Ethernet or InfiniBand fabric stitching thousands of nodes together. It's where multi-vendor sourcing and field serviceability matter most, so it adopts CPO more slowly — and where your pluggable FR4/DR4 optics keep doing the job.

What this means for the optics you specify in an AI build:

• In-rack scale-up stays copper — DAC, ACC, and AEC for the shortest GPU-to-GPU hops, the lowest-power option until reach runs out.
• Rack-to-rack and leaf links are where 800G–1.6T optics live — DR4 for ~500 m, FR4 for ~2 km — whether the switch is pluggable or CPO.
• On a CPO switch, those optics move to the NIC/host side in OSFP or QSFP-DD, while the switch faceplate is just fiber.

For the platform-by-platform picture of which optic each NVIDIA system takes, see our 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

To make the optics question concrete, consider one leaf switch with 128 ports of 800G connecting to GPU servers, built two ways.

The same 800G leaf, pluggable vs CPO — the switch side changes, the host side doesn't.

In the pluggable build, every switch port carries an 800G transceiver and every server port carries its match, so the link has a module at each end — fully serviceable and multi-vendor, but a hot, dense faceplate. In the CPO build, the switch faceplate has no transceivers — only fiber leaving the package, plus replaceable laser modules. But the host side is unchanged: each server still needs its pluggable optic — a 400G/800G FR4 or DR4 — and in-rack GPU-to-GPU links still run on copper. The lesson for a buyer is simple: a move to CPO changes where you buy optics, not whether. The host-side transceiver becomes the part you qualify most carefully.

11. Operating a CPO network

CPO doesn't only change the hardware you buy — it changes how the network is run day to day. Three shifts matter most, and they're worth planning for before the first switch arrives.

• From field tuning to factory commissioning. Much of a CPO system is assembled and tuned at the factory, so deployment leans more on vendor coordination and acceptance testing than on bench work in your own lab.
• From swap-and-go to monitor-and-predict. A degraded lane can't be swapped on its own, so per-lane health telemetry — laser current, photodiode temperature, engine status — has to feed your monitoring, and troubleshooting shifts toward catching problems early.
• From break-fix to spare capacity. Because a failed engine takes down more than one port, operators provision extra ports — commonly around 5–10% — so a dead lane can be routed around rather than rushed to repair. The replaceable laser module and a clear board-level RMA workflow round out the plan.

The host-side optics, by contrast, are operated exactly as they are today — hot-swappable, monitored through the same tooling, sourced from whoever you choose. That continuity is part of why the host side is the comfortable place to keep your flexibility.

12. Selecting optics for CPO

Whether your next switch is pluggable, LPO, or co-packaged, the decision comes down to matching each link to the right part. The flow below is the quick version: start with reach, then distance, then what matters most on the switch tier.

A simple way to decide what belongs on each link.

Once the part is chosen, a few checks apply on every link — and they matter more in a CPO fabric, where the switch-side DSP is gone and there's 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 and inspect every face; enable the correct host FEC; and qualify any LPO against the specific switch and NIC, since its behavior depends on the host. The checklist below is the version to take into a deployment.

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