What Is Co-Packaged Optics (CPO)? Why AI Clusters Are Moving Light Onto the Chip Package
Co-packaged optics moves the optical engine inside the switch or accelerator package, replacing the faceplate pluggable module. This guide explains what CPO is, why AI clusters need it, the power math behind it, and what it trades away.
On this page
The short answer
Co-packaged optics (CPO) is a packaging approach that places the optical engine - the component that converts electrical signals to light and back - inside the same package as a switch ASIC or accelerator chip, so data leaves the chip as light instead of traveling across a circuit board to a separate pluggable transceiver.
That one move - shortening the electrical path from the 10 to 25 centimeters between a chip and its faceplate to a few millimeters inside a package - is the whole idea. Everything else about CPO, good and bad, follows from it. A shorter electrical path means the chip does not have to spend power pushing a fragile high-speed signal across a board, through a connector, and into a module that then has to recover that signal with its own power-hungry electronics before converting it to light. Less signal to rescue means less power per bit, less heat at the faceplate, and more bandwidth that fits in one switch.
CPO is arriving now, rather than five years ago or five years from now, because of AI clusters. A single training fabric can contain tens of thousands of optical links, and at 800 gigabits per port the power those links consume has stopped being a rounding error. This guide explains what CPO actually is, does the power arithmetic in plain numbers, compares it honestly with the pluggable modules it would replace, and sets out what it trades away - because it does trade things away, and any source that skips that half is selling something. For the week-by-week market context, see our GPU Market Watch; to check whether your own cluster design is compute-bound or communication-bound, use the AI cluster bandwidth estimator.
Key facts
- Definition: CPO integrates optical engines into the same package (or package substrate) as a switch ASIC or accelerator, replacing the faceplate pluggable transceiver as the point where signals become light.
- Electrical reach: in a pluggable design the high-speed electrical signal typically travels about 10-25 cm of circuit board plus a connector; in a CPO design it travels a few millimeters. (Architecture comparison, as of October 2026.)
- Switch generations it belongs to: 51.2T switch ASICs - 64 ports at 800G - are the generation where CPO designs have moved from conference demos to announced products; 102.4T designs point to 64 ports at 1.6T.
- Power, as arithmetic: at 800 Gb/s, a link budget of 15 picojoules per bit means 12 watts for that one link's signaling; at 5 picojoules per bit it means 4 watts. CPO's claim is to move real links toward the second kind of number (illustrative calculation - see the worked example below).
- Lasers: in most CPO designs the laser itself stays outside the package, in a separate, replaceable external laser source, because lasers are heat-sensitive and are the component most likely to need service.
- What does not change: the fiber, the network protocols, and the GPUs. CPO changes where the electrical-to-optical conversion happens, not what the network does.
- Status as of October 2026: early production in large AI fabrics, with pluggable optics still the volume default almost everywhere else.
How CPO works, part by part
Start with the conventional design CPO replaces. In a normal high-speed switch, the switch chip sits near the middle of a circuit board. Its high-speed electrical outputs - SerDes lanes - run across the board to cages on the front panel. A pluggable transceiver slides into each cage. Inside that module, electronics retime and clean up a signal that has been degraded by its journey across the board, and a laser turns the cleaned signal into light for the fiber.
Every element of that chain exists for a reason, and together they made sense for decades: the module is a standard part, made by many vendors, that a technician can swap in seconds. The problem is purely physical. The faster the signal, the more energy it takes to push it across each centimeter of board without it smearing into noise, and the more electronics the far end needs to reconstruct it. At 100G per lane and 800G per port, that tax is paid 64 times per switch, on thousands of switches.
CPO deletes most of the journey. The optical engine - a small photonic chip carrying modulators and detectors, usually built with silicon photonics - is mounted on the same package substrate as the switch chip, connected by millimeters of package wiring instead of centimeters of board. The SerDes signal barely leaves the package before it becomes light. Fiber connectors on the faceplate then carry light, which does not degrade the way a board trace signal does, out to the rest of the fabric.
Two supporting choices recur in real CPO designs. First, the laser usually lives outside the package, in a pluggable external laser source that feeds light into all the optical engines. Lasers lose efficiency and lifetime when hot, a chip package is a hot place, and a failed laser should not mean scrapping a switch. Second, the heavy signal-processing electronics that a pluggable module carries largely disappear, because there is no long, lossy channel left to compensate for. That disappearance is where most of the power saving comes from - which the next section quantifies.
The power math, in plain numbers
Interconnect power is easiest to grasp per bit. Engineers quote it in picojoules per bit (pJ/bit): how much energy it costs to move one bit across the link, counting the driver on the chip and the electronics at the conversion point.
Worked example (illustrative): an 800 Gb/s port moves 800 billion bits per second. At 15 pJ/bit, that is 800 x 109 x 15 x 10-12 = 12 watts spent just moving bits across that one port's electrical path and conversion electronics. At 5 pJ/bit, the same port costs 4 watts. Difference per port: 8 watts.
Now scale it the way an AI operator has to. A 64-port switch at the higher figure spends roughly 500 more watts on link signaling than the same switch at the lower figure. A fabric of 1,000 such switches spends roughly half a megawatt more - continuously, before a single GPU has computed anything, and before the cooling needed to remove that heat is counted. Nobody needs the figures to be exact to see the shape of the problem: at AI-cluster scale, picojoules per bit is a power-plant question, not an electronics detail.
This is also why the industry did not jump straight from copper to CPO. The intermediate steps - better pluggable modules, and linear pluggable optics (LPO), which remove the module's signal-processing chip and let the switch SerDes drive the optics directly - each cut part of the same tax while keeping the module replaceable. CPO is the end of that progression: the shortest possible electrical path, and therefore the lowest signaling power, at the price of giving up the module itself. Our existing guide to multi-GPU setups and NVLink scaling covers the same communication-overhead problem inside a single server; CPO is the corresponding answer between switches and between racks.
CPO vs. pluggable optics: the honest comparison
- Power per bit: CPO wins, structurally. Millimeters of package wiring need far less drive energy and no retiming electronics compared with 10-25 cm of board plus a connector plus a module.
- Bandwidth density: CPO wins. A faceplate full of module cages has a physical port limit; a faceplate that only has to pass fibers can carry far more bandwidth per switch, which is what 102.4T switch designs need.
- Field service: pluggables win, decisively. A failed pluggable is a 30-second swap by one technician. A failed CPO optical engine is not a field repair at all - the design answer is redundancy, replaceable external lasers, and in the limit, replacing the switch.
- Sourcing: pluggables win. Modules are standardized and multi-vendor; an operator can dual-source them and upgrade optics without touching switches. CPO ties the optics to the switch vendor's package design.
- Upgrade path: pluggables win. New module generations drop into old cages within limits. CPO optics change when the switch generation changes.
- Reach inside the fabric: both end in fiber, so both cover the same distances once the signal is light. The difference is only where conversion happens.
Read that list carefully and the deployment pattern predicts itself: CPO goes first where one operator builds a huge, uniform fabric, runs it for a full hardware generation, and values power and density over swap-ability - which is a precise description of a hyperscale AI training cluster, and of almost nothing else. Enterprise and smaller cloud networks, where a handful of switches must serve for years and parts must be replaceable at 2 AM, will rationally stay with pluggables much longer.
What CPO trades away
- Serviceability. The central trade. Optics move from the most replaceable part of the switch to the least replaceable, so reliability engineering - redundant engines, graceful degradation, laser sparing - has to absorb failures that used to be a module swap.
- Yield and cost coupling. A package that combines a large switch die with several optical engines is only good if every part of it works. Packaging yield problems therefore tax the most expensive component in the system, not a cheap module.
- Vendor lock-in. The optics are part of the switch package, so choosing CPO is choosing one vendor's silicon, optics and roadmap together.
- A young ecosystem. Test equipment, sparing practice, failure statistics and second sources for CPO are all immature compared with 30 years of pluggable practice. Early adopters are buying the learning curve along with the hardware.
- Thermal design pressure. Packing conversion electronics against a hot ASIC concentrates heat in the package and is one reason lasers are kept external.
None of these are arguments that CPO fails. They are the reasons its rollout is measured in switch generations rather than quarters, and why "CPO replaces pluggables" headlines should be read as "CPO takes the top of the AI fabric market first, over several years."
Who should care, and when
- You design or buy switches for a large AI fabric: CPO is a roadmap question now. Ask vendors for power per bit at the port speeds you will actually deploy, for the external laser sparing model, and for what a failed optical engine means operationally.
- You rent GPU clusters: you will not choose CPO directly - but it is part of why newer fabrics quote better network power and higher all-to-all bandwidth. The practical step is unchanged: check whether your workload is communication-bound before paying for more GPUs, with the AI cluster bandwidth estimator.
- You run a small or on-prem setup: pluggable optics and plain copper inside the rack remain the right answer for the foreseeable future. File CPO under "know what it is" rather than "plan a purchase."
- You follow the supply chain: CPO shifts value from module assembly toward advanced packaging, silicon photonics and external laser sources. Our GPU Market Watch tracks the pricing and capacity side of that shift as it shows up in the GPU market.
This is the first article in our optical interconnect series. The next articles take the pieces in turn: copper versus optical interconnects inside the AI rack, how much bandwidth an AI cluster actually needs, NVLink domains versus scale-out networks, 800G and 1.6T optical modules, and the power bill of the interconnect itself. Each will link back here, because the definition above is the one the whole series uses.
Frequently asked questions
What is co-packaged optics in simple terms?
Co-packaged optics (CPO) is a way of building a switch or accelerator where the optical engine - the part that turns electrical signals into light - is placed inside the same chip package as the main processor, instead of in a separate pluggable module on the front panel. The data travels as light almost from the moment it leaves the chip, instead of crossing a circuit board as an electrical signal first.
How is CPO different from a pluggable transceiver?
A pluggable transceiver sits in a cage on the faceplate, 10 to 25 centimeters of circuit board away from the chip, and can be pulled out and replaced in seconds. A CPO optical engine sits millimeters from the chip inside its package, so the link uses far less power - but the engine itself is not field-replaceable. If it fails, you replace a laser source or the whole switch, not a module.
Why do AI clusters need CPO?
Because the electrical link between the chip and the faceplate has become a power and density problem. At 800G per port and 51.2T per switch, driving high-speed electrical signals across a board, and recovering them in a pluggable module, consumes a large and growing share of the networking power budget. Shortening that electrical path to millimeters is the main lever CPO pulls.
When will CPO replace pluggable optics?
Not completely, and not soon. The practical expectation as of October 2026 is a long hybrid period: CPO appears first in large, single-vendor AI fabrics where one operator controls the whole switch, while pluggables remain the default everywhere flexibility, multi-vendor sourcing and field service matter. CPO grows generation by generation - it does not flip a switch.
Does CPO make GPU clusters cheaper to run?
It reduces the power spent moving data between chips, which is a real and growing line in a large cluster's power bill, and it can allow denser switches. It does not reduce the cost of the GPUs themselves, and early CPO systems carry packaging and yield costs of their own. The honest claim is narrower: CPO attacks the interconnect power bill, which pluggable optics cannot shrink much further.