TSMC is betting on photonics
TSMC is accelerating work to move from copper wires to silicon photonics — essentially using light instead of metal to connect chips — and it’s filing patents at roughly twice the rate of Intel as it prepares for 2026 mass production, which could cut latency and power for big AI systems. (x.com) This matters because interconnect limits are becoming a core bottleneck as GPUs and accelerators pack more compute close together, so a successful photonics ramp would shift where performance gains come from and who controls key supply lines. (x.com)
Inside an artificial intelligence server, the slow part is often not the math chip. It is the traffic jam between chips, where electrical signals run through copper traces that burn power and lose speed as distances and data rates climb. (tsmc.com) Silicon photonics is the workaround. It puts tiny optical parts on silicon so data can move as pulses of light through fiber, the way a laser pointer carries information farther than a metal wire can. (tsmc.com) One trick makes light links especially useful in data centers: wavelength division multiplexing. That means one fiber can carry multiple colors of light at once, so a single strand acts like a multilane highway instead of a one-lane road. (trendforce.com) The hard part is not making one optical component. The hard part is stacking electronic circuits, which do the computing, directly on photonic circuits, which move the light, without wasting power or creating alignment problems. (tsmc.com) Taiwan Semiconductor Manufacturing Company has been building exactly that stack in a platform it calls Compact Universal Photonic Engine. The design uses System on Integrated Chips stacking so an electrical die sits on top of a photonic die with a very short die-to-die path. (tsmc.com) The company’s public high-performance computing page now says its 65 nanometer silicon photonics technology is already in volume production. The same page says TSMC has demonstrated optical modulation at 200 gigabits per second and more than 99 percent three-dimensional stacking yield on engineering samples. (tsmc.com) What is new is the timetable for moving from pieces to full systems. Multiple industry reports in 2025 and 2026 say TSMC’s Compact Universal Photonic Engine is slated for volume production in 2026 as the company pushes toward co-packaged optics, which puts the optical link right beside the processor package instead of at the edge of the server. (trendforce.com) That packaging shift matters because the biggest artificial intelligence systems now stitch together thousands of graphics processors, and every extra watt spent on moving bits is a watt not spent on training or inference. Ansys said in its April 24, 2024 announcement with TSMC that the Compact Universal Photonic Engine is aimed at speeding chip-to-chip and machine-to-machine communication for cloud, data center, and high-performance computing systems. (ansys.com) Intel used to be the name most people associated with silicon photonics. TrendForce reported on September 3, 2025 that TSMC filed about 50 United States patents in the area in 2024 versus Intel’s 26, after the two companies were roughly even in 2023 at 46 and 43 filings. (trendforce.com) TSMC’s edge is not just the optical device. It is that the company already controls the advanced packaging steps that artificial intelligence chip customers use to combine logic, memory, and interconnect in one product, so photonics can be added as another layer in the same manufacturing flow. (tsmc.com 1) (tsmc.com 2) If that 2026 ramp lands on schedule, the next jump in artificial intelligence hardware may come less from shrinking transistors and more from shortening the distance data has to travel. In that version of the market, the company that owns the optical stack, the packaging stack, and the foundry line gets to decide who can scale first. (trendforce.com)
