"You Know Light Travels Faster Than Electricity?" AI Semiconductor Race Shifts Focus... Samsung and SK Join the CPO Wave [Chip Talk]

Co-Packaged Optics (CPO)
A Technology That Utilizes Light to Boost Data Transmission Speed
Efficiency Becomes Critical with the Rise of AI Data Centers
Challenges Remain in Securing Yield and Maintenance

The focus of competition in artificial intelligence (AI) semiconductors is shifting from the computational performance of chips to data transmission speeds. No matter how fast a chip can process calculations, it is difficult to fully utilize its performance unless data can be exchanged in a timely manner. Against this backdrop, Co-Packaged Optics (CPO), which transmits data at higher speeds and lower power consumption by using light instead of electrical signals, is emerging as a next-generation AI data center technology.

Optical Engines Around the Chip: Faster with Light

CPO is a technology that attaches the optical engine, which exchanges data with the semiconductor chip, close to the chip or integrates it into a single package. Traditionally, electrical signals from the chip travel across the substrate and circuitry to deliver data. However, the longer the distance, the greater the power consumption and the weaker the signal becomes. The core of CPO is to convert electrical signals to optical signals near the chip for data transmission, thereby reducing the distance electrical signals must travel and lowering both power consumption and signal loss.

Silicon photonics is regarded as a key technology for realizing CPO. It combines optical technology with silicon semiconductor processes to transmit light and convert signals. In addition, CPO leverages advanced packaging technology that integrates the optical engine and semiconductor chip into one package.

CPO technology has become increasingly important, as the advent of AI data centers has shifted the competitive standard in the AI market. In the past, the performance and computational speed of a single chip were most important, but now it is more crucial how quickly and efficiently a chip communicates with high bandwidth memory (HBM). In AI data centers, numerous graphics processing units (GPUs), accelerators, memory, and network devices are closely interconnected. Even if computing chips perform at high levels, if data cannot flow smoothly between chips, the efficiency of the entire system can be limited.

According to the UK market research firm IDTechEX, the market for CPO technology is forecast to reach USD 95 million (approximately KRW 13.1 billion) this year, growing at an average annual rate of 30% through 2035. In ten years, the market is expected to surpass USD 1.2 billion (around KRW 1.65 trillion).

Global Competition Begins, Samsung and SK Join the Race

Comparison diagram of the existing method and CPO method. SK Hynix.

The competition surrounding CPO is expanding beyond semiconductor design to encompass optical engine manufacturing, packaging, and system integration capabilities. Taiwan’s TSMC has already been developing silicon photonics and advanced packaging technologies. Through its “COUPE” technology, which integrates electronic chips that process electrical signals and photonic chips that manipulate light, TSMC is focused on raising the integration density of optical engines, with mass production scheduled for the second half of this year. This approach particularly integrates silicon photonics, HBM, and 3DFabric (3D stacking and packaging technology) within high-performance computing (HPC) platforms for AI applications.

Nvidia and Broadcom have also mass produced and shipped network platforms and switch products incorporating CPO technology. In particular, CPO requires a wide range of fields including switch application-specific integrated circuits (ASICs), optical engines, lasers, optical components, optical fibers and connectors, foundry processes, advanced packaging, system assembly, and testing. As a result, optical communication component manufacturers, packaging firms, and system manufacturers are also participating in the supply chain.

Recently, Korean memory companies Samsung Electronics and SK Hynix have also entered the CPO technology race. Samsung Electronics is working to secure silicon photonics technology and a platform, targeting commercialization of CPO technology in 2027, and mass production of AI semiconductor-silicon photonics integrated products in 2028. The company is pursuing a 'turnkey' strategy encompassing silicon photonics, HBM, system semiconductor foundry, and packaging.

SK Hynix is focusing on optical interconnects between memory and processors. The company’s researchers even published a paper on their CPO technology roadmap in the international journal Nature Electronics. Through an architecture that connects memory and processors via optical interconnects, they outlined ways to alleviate data movement bottlenecks in large-scale AI systems and suggested future commercialization potential.

Yield, Maintenance, and Mass Production: Commercialization Still a Hurdle

CPO technology is still in the early stages of commercialization. As a highly complex process, there are challenges that must be addressed before establishing mass production systems. These include ensuring precise alignment of optical components and semiconductor chips, as well as managing heat generated inside the package. Reducing defects during manufacture and securing high yield rates are also crucial.

Maintenance approaches also pose challenges. With conventional plug-in optical modules, problematic modules can simply be replaced, whereas with CPO—where the optical engine is integrated into the package—inspection and repair systems must be established to suit the overall system architecture. The degree to which CPO can lower manufacturing and operational costs compared with existing methods may also determine the pace of commercialization.

Some experts predict that CPO technology adoption will accelerate globally. At CES 2026, Nvidia identified the third quarter of this year as a critical inflection point for commercialization. An industry insider commented, “In the past, the focus was on improving the performance and competitiveness of memory semiconductors themselves, but now there is growing recognition that simply making good memory isn’t enough. It's now essential not only to secure competitiveness in individual technologies but also to consider how to connect and optimize different technologies together.”

This content was produced with the assistance of AI translation services.