AI Tech NewsYour portal to artificial intelligence
Home / What Is CPO? How Co-Packaged Optics Brings Light to AI Servers and Data Centers
AI IndustryInfrastructure

What Is CPO? How Co-Packaged Optics Brings Light to AI Servers and Data Centers

5 min read

Co-packaged optics (CPO) places optical components right next to the switch chip so AI servers can exchange data with light. Here's what it is, why AI needs it, and where it stands.

In this article
  1. The Short Version
  2. What Problem Does It Solve?
  3. Key Features
  4. Which Products Exist Today?
  5. How Is It Different from Pluggable Transceivers?
  6. What Should Non-Engineers Know?
  7. Who Should Pay Attention?
  8. Who Can Probably Skip It for Now?
  9. Current Hurdles and Risks
  10. Our Take
  11. Sources

Training a large AI model often takes tens of thousands of GPUs working together. Those GPUs are spread across many servers, and they constantly trade data. No matter how fast each GPU is, the whole system slows down if data can't move between them quickly enough.

Co-packaged optics (CPO) is a technology built to tackle that bottleneck. It packages the optical components that convert electrical signals into light right next to the switch chip. Data leaves as light almost immediately, instead of first traveling a long stretch of copper across the circuit board.

You'll often see headlines about "CPO servers." In practice, that usually means AI server racks and networks built around CPO switches. Today, the CPO products in real production are mostly network switches, not the servers themselves.

The traditional approach is like trucking goods from a factory to a train station a few miles away. CPO builds the train station right at the factory door, so goods board the train the moment they're made.

The Short Version

CPO moves optics next to the chip to shorten power-hungry copper paths. The goal is to let AI data centers connect more GPUs while using less power and suffering fewer failures. Switch products began entering production in 2026, but cost, testing, and serviceability remain hurdles before wide adoption.

What Problem Does It Solve?

Inside a data center, servers and switches are usually linked with fiber. The standard approach today uses pluggable transceivers: small modules, roughly the size of a USB stick, that plug into the front panel of a switch and convert electrical signals to light.

Pluggables are convenient. If one fails, you pull it out and swap in a new one. But AI workloads have exposed their limits:

  1. Rising power use: Signals leave the chip and travel across the board before reaching the transceiver. The faster the signal, the more it degrades along that copper, and the more power it takes to compensate.
  2. Huge component counts: A large switch can hold over a hundred transceivers, each with its own lasers. Across an entire AI cluster, that adds up fast.
  3. More points of failure: More components mean more chances for something to break. When a link drops during AI training, an entire job may have to pause or restart.

CPO's logic is simple: if the trouble lies in the path from chip to transceiver, put the optics right next to the chip.

Key Features

1. Optical Engines Sit Beside the Chip

CPO packages optical engines, the parts that convert between electrical and optical signals, together with the switch chip. Electrical signals travel only a very short distance before becoming light, which cuts both signal loss and power draw.

This relies on silicon photonics, which uses standard semiconductor manufacturing to build tiny structures on silicon that can guide and process light.

2. Fewer Lasers, Lower Power

NVIDIA says its CPO switches use about 4x fewer lasers than designs built on traditional transceivers, and it claims several-fold gains in power efficiency. These are vendor figures, and real-world benefits will depend on how each data center deploys them.

3. Connecting More GPUs

Bigger AI models need more GPUs linked together. CPO switches offer more bandwidth per box and use less power per port, which helps operators build larger GPU clusters within the same power budget.

Which Products Exist Today?

Company Product Type Public details
NVIDIA Quantum-X Photonics InfiniBand switch 144 ports at 800Gb/s, liquid cooled
NVIDIA Spectrum-X Photonics Ethernet switch Up to 400Tb/s total bandwidth; reported in full production as of August 2026
Broadcom Tomahawk 6 Davisson Ethernet switch chip 102.4Tbps, third-generation CPO, shipping announced October 2025
TSMC COUPE Silicon photonics packaging platform Optical-electrical integration for NVIDIA, Broadcom, and others; volume production targeted for 2H 2026

NVIDIA designed its CPO switches to network its new Vera Rubin AI platform. According to StorageReview, cloud providers including CoreWeave, Oracle Cloud, and Microsoft Azure are already integrating them.

How Is It Different from Pluggable Transceivers?

Comparison Pluggable transceivers Co-packaged optics
Where the optics live Switch front panel Next to the chip, in the same package
Copper path Longer Very short
Power use Higher Lower (per vendor figures)
Repairs Pull and replace More complex; serviceability must be designed in
Supplier choice Many interchangeable vendors More tightly tied to the chipmaker
Maturity Very mature Early production

In short, pluggables win on flexibility and easy repairs, while CPO wins on power and density. Expect the two to coexist for years.

What Should Non-Engineers Know?

  • AI's bottleneck isn't just GPUs: Power and data movement are becoming key limits on how fast AI data centers can grow.
  • Optics are replacing copper over time: It starts with switches. The industry is also exploring bringing optics into GPUs and servers, but no firm timeline exists yet.
  • Taiwan's supply chain plays a big role: TSMC fabricates the silicon photonics, SPIL handles packaging and testing, and Foxconn assembles the switch systems. That's why CPO gets so much attention in Taiwanese media.

For a closer look at how fiber connects to chips, see our earlier article on Corning's Glass Bridge and CPO.

Who Should Pay Attention?

  • Readers following AI data center infrastructure trends.
  • Content creators and founders covering semiconductors, networking, and optical communications.
  • Technical leaders planning large GPU clusters or cloud data centers.

Who Can Probably Skip It for Now?

  • Teams using ordinary cloud instances or small servers: CPO targets the largest AI clusters and won't show up in typical server rooms anytime soon.
  • Anyone treating CPO headlines as investment advice: an industry trend doesn't guarantee any company's profits, so check financial reports and professional analysis.

Current Hurdles and Risks

  1. Cost is still high: Market research suggests CPO currently costs at least 10% more than comparable transceivers.
  2. Testing is hard: Once optics and chips share a package, wafer-level testing and fiber alignment get much harder. TSMC has publicly named these as production challenges.
  3. Repairs work differently: A failed pluggable can be swapped on its own. CPO requires new approaches to maintenance and redundancy.
  4. Volume is still unproven: Research firms such as TrendForce see 2027 and 2028 as the years that will decide whether CPO reaches true high-volume production.

Our Take

CPO isn't about replacing pluggable transceivers overnight. Its first job is to ease the power crunch in the very largest AI clusters. When a data center's power budget is fixed, every watt saved on networking is a watt that can go to GPUs instead.

Over the next two years, we think the specs matter less than three questions: how often deployed links actually fail, whether the repair process works in practice, and whether costs fall enough for more cloud providers to adopt it. Those answers will only become clear once the first CPO clusters have run in production for a while.

Sources

Continue exploring

AI Industry

What is LPU? The Secret Behind the 'Language Processing Unit' Designed for AI Inference, 10x Faster than GPU

LPU (Language Processing Unit) is a novel AI chip architecture developed by Groq, designed specifically for large language model inference. By discarding the dynamic scheduling and external memory of GPUs in favor of deterministic static scheduling and massive on-chip SRAM, it generates tokens several to over ten times faster than traditional GPUs, making it a crucial hardware accelerator for real-time AI conversation and voice applications.

Read more →