Co-packaged optics explained: why AI data centers are moving from copper to light

Co-packaged optics, usually shortened to CPO, is a way of building network switches for AI data centers so that the parts that turn electricity into light sit on the same package as the switch chip. It matters because the network that links thousands of AI chips has become a major power draw, and the industry's biggest names are betting on light to cut it.

What problem it solves

Every AI cluster is a network. Training a large model means thousands of accelerators exchanging data constantly, and the cables between them have to be fast. Inside a rack, short copper links work. Between racks, over longer distances, light travelling through fiber carries far more data with less loss.

The conventional answer is a pluggable optical transceiver: a small module that slides into the front of a switch, takes an electrical signal from the switch chip and converts it to light. It works, and it is serviceable, because a failed module can be swapped. But the electrical signal has to travel from the chip, across the circuit board, to the module at the panel. That trip burns power.

The scale makes that matter. Optics & Photonics News reported in March 2025 that an AI factory can use up to 2.4 million optical transceivers and up to 24 MW to run them, which it said may be more than 10% of total data center power. Nvidia's chief executive Jensen Huang claimed the new approach could save data centers "tens of megawatts."

What co-packaged optics changes

In CPO, the optical engines move to sit beside the switch chip on the same package. The electrical path shrinks from the length of a circuit board to a much shorter hop. Less distance means less power spent pushing the signal and, in principle, a cleaner signal.

Nvidia describes the technology as the fusion of electronic circuits and optical communications, built on silicon photonics, which uses chip-making methods to manufacture the light-handling parts. According to Optics & Photonics News, its design also uses a micro ring modulator that it says adds further savings, and CPO is used for short-range connections to improve electrical efficiency.

Marvell, a competitor and supplier in this field, makes the same case from the other direction. In its November 2, 2026 release about acquiring Lumentum it says bandwidth demands and "the physical limitations of copper-based connectivity are accelerating the transition toward optical technologies." It also lists a new vocabulary for how far data has to go: scale-across between sites, scale-out between racks, scale-up within a rack and scale-in within a package.

What Nvidia has announced

At its GTC conference on March 18, 2025, Nvidia announced two photonics switch lines.

• Quantum-X Photonics (InfiniBand): 144 ports of 800 Gb/s, liquid cooled, due later in 2025.
• Spectrum-X Photonics (Ethernet): configurations from 100 Tb/s to 400 Tb/s total, for availability in 2026.

The company said, compared with traditional methods, that the switches use 4 times fewer lasers and deliver 3.5 times more power efficiency, 63 times greater signal integrity, 10 times better network resiliency at scale and 1.3 times faster deployment. Its ecosystem partners include TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, Sumitomo Electric and others.

The catch

Every one of those multipliers comes from Nvidia, measured against an unspecified baseline of traditional methods, in a product announcement. The sources we opened offer no independent field measurement. That does not make them wrong. It means the right label is "company claim," and the right time to believe a multiplier is after an operator runs the hardware at scale and reports.

There are also real engineering trade-offs that the marketing does not dwell on. A pluggable module can be swapped at the panel; whether optical engines built into the package are as easy to service is a question operators will weigh, and the sources we opened do not answer it. Nvidia's resilience claim leans on needing 4 times fewer lasers, since fewer parts means fewer things to fail. Whether that holds over years of continuous operation has not been shown publicly.

Why the supply chain is in the news

CPO and the older pluggable modules both depend on a narrow set of laser and optics suppliers. That is why a deal like Marvell's $115 billion all-stock agreement to buy Lumentum, announced November 2, 2026, is a story about AI hardware. Optics & Photonics News names Lumentum as the supplier of lasers for Spectrum-X, and Marvell says it wants to combine its switching and custom silicon with Lumentum's lasers and optical switching.

Marvell's chief executive Matt Murphy described it this way: "AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation." That is an executive's sales pitch. The structural point underneath it is sound: if the network is a large share of cluster power, the companies that make the light matter more than they did.

For the demand side of this story, see our tracker of Big Tech AI capital spending, which shows how much money is chasing exactly this kind of build-out.

Our take

CPO is not hype in direction. Power is the constraint on AI data centers, the network is a meaningful share of it, and moving the optics next to the chip is a plausible way to reduce it. Multiple large suppliers and the biggest chip company are converging on the idea.

What is hype is the precision. A figure like 3.5 times or 63 times, from the seller, against an undefined baseline, tells you what the company hopes to show, not what a customer will see. The practical test is simple: ask for deployed numbers, from an operator that is not the vendor, covering reliability as well as power. Until those exist, plan for the technology and discount the multipliers.