WORLD

Taiwan's TSMC, 30 firms push silicon photonics alliance to tackle AI power crisis

by
Seo Jiyeon
Published : Sept. 7, 2026 - 14:24:33
    • Copy Completed!

View Korean Original

Alliance aims to lead silicon photonics supply chain

Light-based data transfer cuts power use by 30–50%

Nvidia invests $4 billion as technology race accelerates

TSMC targets mass production of next-gen CoPoS packaging by 2028

TSMC's headquarters at the Hsinchu Science Park in Taiwan [Getty Images]
TSMC's headquarters at the Hsinchu Science Park in Taiwan [Getty Images]

TSMC and about 30 other leading Taiwanese companies are moving to establish a silicon photonics alliance, as power consumption and data transmission bottlenecks in AI data centers emerge as critical challenges. Taiwan aims to leverage its existing semiconductor manufacturing strength to secure dominance over the next-generation optical semiconductor supply chain.

Taiwan's Ministry of Economic Affairs announced Sunday on Facebook that TSMC and roughly 30 other companies are pursuing the formation of the alliance, a development that drew significant attention at Semicon Taiwan 2026, a semiconductor exhibition held in Taipei from Wednesday through Friday.

The ministry said the alliance would work collectively to resolve key technical challenges in co-packaged optics, or CPO, with the goal of keeping Taiwan at the forefront of the silicon photonics supply chain.

Silicon photonics fabricates optical integrated circuits on silicon substrates to transmit data using light rather than electrical signals. It can process large volumes of data faster than conventional electrical-signal-based connections, making it a leading candidate for next-generation AI semiconductors and data centers.

The ministry said silicon photonics can cut power consumption by 30 to 50 percent and is a key technology for overcoming data transmission bottlenecks that arise as AI computing performance scales up. As AI models grow larger and the volume of data exchanged between chips surges, data transfer speed and energy efficiency are becoming as decisive as raw processing power.

CPO integrates silicon photonics with AI processors to boost data transfer speeds and energy efficiency. Because optical components must be placed in close proximity to the semiconductor, the technology demands a tight combination of design, manufacturing and packaging expertise — making it one of the central challenges in commercializing silicon photonics.

Global technology companies are also accelerating related investment. Nvidia, the top player in the AI semiconductor market, invested $4 billion in silicon photonics technology in March, moving to secure next-generation optical interconnect capabilities.

The ministry highlighted Taiwan's well-developed semiconductor ecosystem as a competitive advantage in this technology race. Taiwan holds more than 90 percent of the market for advanced processes at 7 nanometers and below, and about 50 percent of the global packaging and testing market. The strategy is to extend that manufacturing and packaging edge into silicon photonics, cementing a position in the global supply chain that would be difficult to replace.

Taiwanese President Lai Ching-te last year unveiled a plan called the "AI New Ten Constructions," which includes silicon photonics, targeting production output of 15 trillion Taiwan dollars and the creation of 500,000 AI jobs by 2040.

Meanwhile, TSMC is reportedly planning to begin mass production of its next-generation advanced packaging technology, chip on panel on substrate, or CoPoS, in the second half of 2028. Taiwanese media, citing sources, said TSMC is pursuing CoPoS as a successor to its existing chip on wafer on substrate, or CoWoS, technology.

CoPoS uses square panels measuring 310 by 310 millimeters instead of conventional 12-inch circular wafers, and TSMC plans to raise material utilization rates from below 70 percent to more than 90 percent. Sources said the company will begin equipment and materials verification this year, conduct small-scale trial production in 2027, and move to full mass production in the second half of 2028.


sjy@heraldcorp.com
This content was produced with the assistance of AI translation services.

MOST READ