Co-packaged optics (CPO)

Co-packaged optics (CPO) integrates the optical engine that converts electrical signals to light into the same package as a switch or accelerator chip, replacing pluggable transceivers. By shortening the electrical path, it cuts interconnect power (NVIDIA and Broadcom cite roughly 3.5x and 70% gains) and raises bandwidth density for AI data centers.

What co-packaged optics is, in plain terms

Inside a data center, chips talk to each other over either copper wires or beams of light in optical fiber. Light carries far more data over distance with less loss, but you need an **optical engine** to convert electrical signals into light and back. Traditionally that engine sits in a **pluggable transceiver** plugged into the front panel of a switch, several centimeters of copper trace away from the switch chip. **Co-packaged optics (CPO)** moves that optical engine off the front panel and right next to the main chip (a network switch ASIC or an AI accelerator), on the same package substrate. The electrical signal now travels millimeters instead of centimeters before becoming light. CPO is sometimes described as putting the fiber connection almost directly onto the chip.

How it works

A CPO module combines a **silicon photonics** engine (waveguides, modulators, and photodetectors etched on silicon) with the laser light source and driver electronics, all assembled close to the switch or accelerator die. Fibers attach to the package, and external laser modules often supply the light. Shortening the electrical path is the whole point. Over long copper traces, high-speed signals (now 200 Gbit/s per lane and rising) degrade, so chips spend power on **equalization and retiming (SerDes)** to clean them up. With optics millimeters away, that signal-conditioning overhead shrinks dramatically, which is where most of CPO's power and signal-integrity gains come from.

Why it matters for AI and data centers

AI training clusters connect tens of thousands of GPUs, and the network between them has become a bottleneck for both bandwidth and electrical power. Copper struggles past roughly a meter at these speeds, and pluggable optics burn significant energy and panel space. The vendor numbers are striking. **Broadcom's** 51.2-Tbps "Bailly" CPO Ethernet switch (announced at OFC 2024) claims about **70% lower optical-interconnect power** and 8x better silicon-area efficiency versus pluggables. **NVIDIA**, at GTC 2025, said its Quantum-X and Spectrum-X Photonics switches use CPO to deliver roughly **3.5x better power efficiency, 63x better signal integrity, and 10x better network resiliency** — saving an estimated tens of megawatts across a large AI "factory."

Where CPO sits in the photonics supply chain

CPO is a system-integration play that pulls together several layers: - **Switch / accelerator ASIC:** Broadcom (Tomahawk), NVIDIA (Quantum/Spectrum), Marvell. - **Silicon photonics + advanced packaging:** TSMC's **COUPE** platform is a leading foundry option for stacking the photonic engine with electronics; Intel, GlobalFoundries and others also play here. - **Lasers and light sources:** EML/DFB laser makers such as **Lumentum, Coherent, Sumitomo Electric (TSE: 6451)** and Mitsubishi feed the silicon-photonics engines. - **Optical engines / interposers:** specialists like **POET Technologies** (its Optical Interposer) and **Ayar Labs** (optical I/O chiplets) provide integration building blocks. No single company owns the whole stack, so partnerships across these layers are the norm.

Who the key players are

**Broadcom** was first to market with a 51.2T CPO switch and emphasizes power and bandwidth-density leadership. **NVIDIA (NVDA)** unveiled the most prominent CPO products in 2025 — silicon-photonics switches built with an ecosystem that includes TSMC and laser suppliers such as Sumitomo. **Marvell (MRVL)** is pushing CPO into custom AI accelerators with its 3D Silicon Photonics Engine, and its Celestial AI unit has touted multi-terabit optical scale-up links. Among smaller pure-plays, **POET Technologies (POET)** is building its Optical Interposer as a manufacturing platform for both pluggables and CPO, winning an ECOC 2025 innovation award and signing partners like LITEON and Sivers. **Sumitomo Electric (6451)** supplies lasers and assembly into NVIDIA's photonics ecosystem.

What's changing now

Through 2025 and into 2026, CPO is moving from demos toward early deployment. Broadcom's Bailly began shipping and the design has reached volume production at system integrators; NVIDIA's Quantum-X Photonics InfiniBand switches were slated for late 2025 with Spectrum-X Ethernet following in 2026. TSMC's COUPE silicon-photonics packaging is ramping. Real hurdles remain: getting a working laser permanently *inside* a hot, hard-to-repair package, fiber attach and alignment at scale, thermal management, testing, and serviceability (a failed pluggable is easy to swap; a failed co-packaged engine is not). Many analysts therefore expect a **hybrid** future where pluggables, linear-drive optics (LPO), and CPO coexist, with CPO winning first where power and density pressure are highest.

Frequently asked

What is the difference between co-packaged optics and a pluggable transceiver?

A pluggable transceiver is a removable module that holds the optical engine at the switch's front panel, several centimeters of copper away from the switch chip. CPO moves that optical engine onto the same package as the chip, just millimeters away. The shorter electrical path cuts power and improves signal integrity, but CPO is harder to repair because you cannot simply unplug a failed module.

Why does AI need co-packaged optics?

Large AI clusters link tens of thousands of GPUs, and the interconnect's bandwidth and power consumption have become limiting factors. Copper can't carry today's per-lane speeds much beyond a meter, and pluggable optics use a lot of energy on signal conditioning. CPO raises bandwidth density and cuts interconnect power, helping operators scale clusters without the network dominating the power budget.

Is co-packaged optics in production yet?

It is in early deployment. Broadcom's 51.2T "Bailly" CPO switch (announced 2024) reached customers and volume production at system integrators. NVIDIA announced its Quantum-X and Spectrum-X Photonics CPO switches at GTC 2025, with InfiniBand versions targeted for late 2025 and Ethernet in 2026. Broad adoption is still ramping.

What companies make co-packaged optics?

Key players include Broadcom and NVIDIA (switch ASICs plus photonics), Marvell (CPO for custom AI accelerators), TSMC (its COUPE silicon-photonics packaging), laser suppliers like Lumentum, Coherent and Sumitomo Electric, and integration specialists such as POET Technologies and Ayar Labs.

Will CPO replace pluggable optics?

Not entirely, at least not soon. Pluggables are serviceable and flexible, and intermediate options like linear-drive optics (LPO) also save power. Most analysts expect a hybrid market where CPO is adopted first in the highest-bandwidth, most power-constrained AI switching and scale-up links, while pluggables persist elsewhere.

Related companies