Co-Packaged Optics Explained: Why AI Networking Is Moving Optics Next to the Switch Chip
Co-packaged optics (CPO) moves the part that turns electrical signals into light off the front panel of a network switch and onto the same package as the switch chip. Today, almost every fiber link in a data center runs through a pluggable transceiver: a small module, about the size of a pack of gum, that slots into the front of a switch or a server network card. CPO gets rid of most of those modules.
The reason is power. AI clusters connect tens or hundreds of thousands of GPUs, and every link needs optics at both ends. As link speeds climb to 1.6 and 3.2 terabits per second, the electrical path between the switch chip and a pluggable module burns more and more energy just pushing bits across a few inches of circuit board. Shorten that path to millimeters and most of that waste goes away.
For background on why AI data centers run on light at all, see what is optical connectivity.
Three ways to build an optical link
| Approach | Where the optics sit | Signal processing | Main advantage | Main drawback |
|---|---|---|---|---|
| Pluggable transceiver | Removable module on the front panel | DSP chip in each module cleans up the signal | Easy to swap, mix vendors, mature supply chain | Highest power; DSP adds heat and cost |
| Linear pluggable optics (LPO) | Removable module on the front panel | No DSP; the switch chip’s own SerDes drives the optics | Lower power and latency than standard pluggables | Tight signal budgets; works best over short reaches |
| Co-packaged optics (CPO) | Optical engines on the switch package | Very short electrical path; no module DSP | Lowest power per bit, highest bandwidth density | Harder to service; new supply chain; vendor lock-in |
A DSP (digital signal processor) in a pluggable cleans up the electrical signal after its trip across the circuit board. It works well but uses a lot of energy. LPO drops the DSP and relies on a cleaner signal from the switch chip, a middle step that kept pluggables alive for another generation. CPO goes further and puts the optical engine right next to the chip.
Why power per bit matters
Network power is usually counted in picojoules per bit (pJ/bit). A rough rule of thumb in the industry is that a DSP-based pluggable link costs somewhere in the mid-teens of pJ/bit, while CPO designs aim for single digits. Nvidia has said its silicon photonics switches use about 3.5 times less power for optics than pluggables.
That sounds small until you multiply it. A large AI cluster can need millions of optical transceivers. At hundreds of thousands of GPUs, the optics alone can eat tens of megawatts, power that could otherwise run more GPUs. With grid access the hardest constraint on new data centers (see AI data center power explained), every watt saved in networking is a watt that can go to compute.
There is a second benefit. Fewer separate parts means fewer things to fail. Nvidia claims its photonics switches are far more resilient than a wall of pluggables, because a lot of link flapping in big clusters comes from individual transceivers.
Who is shipping what
Broadcom was first to market with CPO switches. Its Bailly design paired a 51.2 terabit Tomahawk 5 switch chip with optical engines on the package, and its Davisson generation moved to the 102.4 terabit Tomahawk 6.
Nvidia announced CPO switches at its GTC conference in March 2025. Quantum-X Photonics (InfiniBand) came first, with Spectrum-X Photonics (Ethernet) following. Nvidia’s design uses micro-ring modulators and relies on TSMC’s COUPE silicon photonics packaging. That is the same TSMC advanced packaging ecosystem described in CoWoS explained.
Marvell agreed in December 2025 to buy Celestial AI, a startup building optical links for connecting chips and memory, for about $3.25 billion. That is a bet on optics inside the rack, as well as between switches. Our piece on Marvell’s connectivity moat covers why connectivity is central to Marvell’s case.
Other players include startups such as Ayar Labs and Lightmatter working on optical chip-to-chip links, and the big optical component makers (Lumentum, Coherent and others) that supply lasers.
Scale-out first, scale-up next
AI clusters have two networks.
- Scale-out links racks to each other through switches. This is where CPO is arriving first, because switch boxes already concentrate a huge number of optical ports.
- Scale-up links GPUs inside a rack or pod so they can act like one giant processor (Nvidia’s NVLink, for example). Today this runs mostly on copper cables, which are cheaper and very reliable over a meter or two.
Copper runs out of reach as racks get denser and speeds rise. Most roadmaps expect optics to take over scale-up links in later GPU generations, which is why companies like Celestial AI are worth billions before they have large revenue.
The objections
CPO has been “two years away” for about a decade. The reasons it took so long are real.
- Serviceability. If a pluggable fails, a technician swaps it in a minute. If an optical engine on a CPO package fails, you may need to replace the whole switch. Designs now put the laser, the part most likely to fail, in a separate replaceable module (an external laser source) for that reason.
- Supply chain. Pluggables come from many vendors that compete on price. CPO ties the optics to the switch maker, which data center operators worry will raise prices.
- Manufacturing yield. Joining fiber to a chip package with micron accuracy at high volume is hard, and a bad optical engine can scrap an expensive package.
- Testing. Pluggables can be tested before they ship. Optical engines on a package are harder to test until the whole assembly is done.
Who wins and who loses
The shift changes where the money goes.
- Gains: switch chip makers who control the package (Broadcom, Nvidia), TSMC and packaging specialists, laser suppliers, and companies that make fiber attach and connectors.
- Pressure: pluggable transceiver makers and DSP vendors, at least in the switch ports that move to CPO first. Pluggables will not disappear soon. Servers, older switches and long reaches will use them for years, and LPO buys them more time.
What to watch
The adoption curve is the main question. Watch how many hyperscalers deploy CPO switches at scale rather than in pilots, whether external laser modules deliver the reliability that operators need, and when scale-up links inside the rack go optical. For where networking sits in the full AI spending stack, see the AI capex value chain, and for the memory side of the same packaging story, HBM explained.