IN Brief:
- TSMC expects silicon photonics to exceed 50% of the optical-transceiver market during 2027.
- Optical-transceiver sales rose 25% in 2025 and are forecast to increase another 50% during 2026.
- Lasers, fibre connectivity, testing, and manufacturing yield are becoming larger constraints as co-packaged optics approaches production scale.
TSMC expects silicon photonics to account for more than half of the optical-transceiver market during 2027 as AI infrastructure drives faster adoption of optical links between processors, switches, and data-centre systems.
K.C. Hsu, vice-president of advanced packaging technology development at TSMC, set out the forecast during the Silicon Photonics Global Summit in Taipei. The event brought semiconductor, packaging, optical-component, and systems companies together around the problem of moving data efficiently as accelerator clusters grow in scale.
TSMC’s market view is supported by rapid growth in optical networking hardware. Hsu said optical-transceiver sales increased by around 25% during 2025 and are expected to rise another 50% in 2026, while the 100G-and-above segment doubled during 2024 before expanding by a further 60% last year.
The underlying pressure is no longer confined to processor performance. AI accelerators remain useful only if processors, memory, switches, storage, and neighbouring compute nodes can exchange enough data to keep the silicon occupied. Higher electrical signalling rates increase channel loss and power consumption, while long copper traces become progressively harder to operate without more equalisation, retiming, and thermal overhead.
Optical links reduce those penalties over longer distances, and silicon photonics extends the approach by integrating functions such as modulators, waveguides, photodetectors, and coupling structures using semiconductor manufacturing techniques. The technology has been available for years, but the present shift is towards substantially higher deployment volumes and tighter integration with advanced semiconductor packages.
Conventional pluggable optical modules already provide high-speed links between network equipment. Near-packaged optics and co-packaged optics move the conversion between electrical and optical signals closer to switch or compute silicon, shortening the highest-speed electrical section and reducing part of the power consumed simply moving data across a circuit board.
The shorter electrical path creates a more complicated package. Optical engines, switch ASICs, fibres, drivers, thermal structures, connectors, and external or integrated laser sources all have to work within an architecture that can be manufactured and tested economically rather than assembled as a laboratory demonstration.
TSMC expects some suppliers to move co-packaged-optics technology into mass production during the second half of 2026. Its own COUPE platform is being developed alongside SoIC and CoWoS packaging, giving the foundry a route to combine photonic integrated circuits with logic and memory inside an established advanced-packaging ecosystem.
Open co-packaged-optics architectures are simultaneously defining interfaces between switches, optical engines, external lasers, fibres, and package structures. Common interfaces become more important once manufacturers have to source those elements from several suppliers and maintain them across large installed fleets.
The remaining bottlenecks are increasingly industrial rather than conceptual. Hsu identified lasers, optical fibres, fibre connectors, and product testing among the areas requiring further development as silicon photonics moves towards larger-scale deployment.
Lasers remain a particular constraint because silicon is an effective material for many optical functions but an inefficient light source. Practical systems therefore depend on III-V semiconductor lasers or separate external laser modules, which must maintain optical output, wavelength stability, and lifetime while operating near high-power electronics.
Fibre attachment creates another production challenge. Semiconductor packaging has decades of experience around automated electrical interconnects, but coupling light efficiently into a photonic waveguide still depends on mechanical alignment precise enough to affect insertion loss and yield. A process that works with careful laboratory alignment can become expensive when repeated across thousands of optical channels.
Testing becomes more difficult for the same reason. A co-packaged-optics assembly can accumulate considerable value before final system test, combining expensive switch silicon with photonic dies, advanced substrates, fibres, and laser hardware. Discovering a defective optical path only after those elements have been assembled carries a much higher cost than rejecting a die or optical component earlier in production.
Wafer-level and intermediate optical testing therefore become increasingly important as volumes rise. Manufacturers need ways to screen photonic devices, coupling structures, electronic interfaces, and assembled optical paths before each additional packaging stage makes a failure more expensive.
The manufacturing problem also extends into thermal management. High-speed switch ASICs and AI processors dissipate substantial power, while optical components and lasers can be sensitive to temperature changes. Bringing the two technologies into the same package reduces electrical distance but increases the need to manage heat without moving optical characteristics outside their operating range.
Those constraints explain why silicon photonics reaching majority market share would represent more than a change in transceiver technology. The industry would have moved from proving that integrated photonics can deliver the required bandwidth into proving that optical engines can be assembled, tested, cooled, connected, and serviced at semiconductor manufacturing volumes.
That also separates TSMC’s market forecast from its parallel COUPE engineering roadmap. COUPE addresses how individual photonic engines can scale through higher channel rates, more channels, wavelength multiplexing, and tighter packaging. The wider 2027 forecast concerns what happens when those technologies have to be supplied in quantities large enough to become routine data-centre hardware.
Silicon photonics may therefore cross its next threshold away from the optical waveguide itself. Lasers, connectors, fibre attachment, test coverage, package yield, and thermal stability are becoming the measurements that decide how quickly a promising integration technology turns into an industrial supply chain.
