CPO Still Awaiting! IET Chairman Kao: Mass Production Growth Expected by 2028, Indium Phosphide Remains Tight in Short Term

AI data centers continue to drive demand for high-speed optical interconnects, but co-packaged optics (CPO)—a technology highly anticipated by the market—still requires time before large-scale commercialization. Kao Yung-chung, Chairman of compound semiconductor epitaxy manufacturer IET Technology (IET-KY, 4971), forecasts that testing and small-scale adoption may begin between 2026 and 2027, with significant market volume growth potentially not occurring until 2028. Meanwhile, indium phosphide (InP) substrate supply, pressured by AI demand and geopolitical factors, is expected to remain tight through 2026–2027, with supply-demand balance unlikely to stabilize until after 2028.

Although CPO volume growth is still pending, this does not mean AI optical communication demand is cooling down. Kao emphasizes that as GPU clusters continue to scale, data centers simultaneously demand higher bandwidth density and energy efficiency. Traditional copper wiring and pluggable optical module architectures will eventually face power consumption and transmission loss bottlenecks. The shift from copper to optical—'optical replacing copper'—remains an irreversible industry trend. However, what needs recalibration is the market’s expectations regarding technological maturity and mass production speed.

The market underestimates integration complexity, with heat dissipation, light sources, and repair reliability posing significant barriers. Kao notes that silicon photonics and CPO concepts have been under development for years, yet external observers often underestimate the complexity and difficulty of integrating optics, electronics, packaging, and thermal management. The core of CPO is moving optical components closer to the switch chip, shortening high-speed electrical signal transmission distances to reduce power consumption and signal loss. However, challenges such as laser placement, heat dissipation, and the ability to quickly repair systems after failure must be resolved before transitioning from lab environments to data centers.

Lasers, in particular, are components that generate significant heat and are prone to failure. Integrating the light source directly into the package risks affecting the entire system if the laser fails. As a result, the industry is gradually developing external laser sources (ELS), placing lasers outside the package. This not only simplifies replacement but also enables a single high-power laser to serve multiple transmission channels.

'It may still take three to four more years before widespread adoption becomes evident. At the most optimistic, 2028 will be when growth becomes clearly visible,' Kao believes. While 2026 and 2027 will see more verification and trial deployments, large-scale commercial penetration will depend on system reliability, customer validation, and overall cost—factors that require time to navigate the learning curve.

1.6T drives single-channel 200G, accelerating demand for high-speed detectors. Although CPO volume growth remains distant, demand for high-speed optical communication components in AI data centers is already impacting the supply chain. Kao points out that as switch specifications evolve toward 1.6T, common architectures will push single-channel data rates to 200G. This requires not only high-speed lasers but also optical detectors capable of handling 200G signals.

On the emission side, silicon is not an efficient light source. On the reception side, while technologies like silicon-germanium (SiGe) can serve certain applications, indium phosphide maintains advantages in sensitivity and high-frequency performance under ultra-high-speed, low-light conditions. As such, it continues to play a critical role in high-speed photodetectors, modulators, and related components.

IET has already mass-produced indium phosphide high-speed detector epitaxial products and is developing epitaxial structures for PIN, avalanche photodiodes (APD), electro-absorption modulators (EAM), and heterojunction bipolar transistors (HBT). Kao believes that even as alternative materials continue to develop, indium phosphide’s position in ultra-high-speed optical communications and high-frequency components will remain difficult to fully replace in the short term.

AI demand and export controls create a 'perfect storm' for indium phosphide. Compared to CPO adoption timelines, InP substrate shortages represent a more immediate supply chain bottleneck. Kao describes the current market situation as a 'perfect storm': on one hand, AI data centers are rapidly increasing demand for photodetectors, modulators, and high-speed lasers; on the other, export controls and regional geopolitical risks are increasing supply uncertainty. Even without restrictions, existing supply growth may not keep pace with demand.

Global InP substrate supply has long been concentrated among a few players, including Japan’s Sumitomo Electric Industries and U.S.-based AXT. AXT’s primary production capacity is located in China, making supply more vulnerable to policy changes. Kao states that over 90% of IET’s InP substrates currently come from Sumitomo Electric. While long-term customers can maintain basic supply, existing capacity struggles to meet rapid demand increases.

The supply chain’s inability to respond quickly is also linked to the previous economic downturn. The InP market was weak in 2023 and 2024, but demand reversed sharply from late 2024 into 2025. Suppliers lacked both the incentive and time to expand capacity early. Now, some cloud service providers and AI companies are pushing for supply chain expansion through prepayments, long-term procurement agreements, or funding new capacity. However, building and validating new production lines typically takes two to three years, making it impossible to immediately fill the gap.

Industry seeks solutions through substrate recycling and quantum dot lasers, with balance expected post-2028. Facing supply shortages, the industry is exploring solutions through capacity expansion, substrate recycling, and alternative materials. Kao explains that epitaxial layers are typically only about one micron thick, compared to hundreds of microns for the substrate—akin to a single page in a book. If used epitaxial layers can be removed and the substrate reprocessed, expensive substrates could be reused, reducing reliance on new materials.

Another path involves developing quantum dot lasers on gallium arsenide (GaAs) substrates to partially replace existing InP light sources, thereby reducing InP consumption. However, whether through substrate recycling or new material alternatives, process development, component validation, and customer adoption must occur, making widespread implementation difficult in the short term.

Kao predicts that InP supply will remain tight in 2026 and 2027. Only after new suppliers’ capacity comes online, substrate recycling matures, and more alternatives pass validation—potentially not until after 2028—can the market gradually return to a more balanced state.