SEMICON Taiwan 2026: The Optical Interconnect Buildout Is Re-Concentrating the AI Supply Chain in Taiwan
The forums opened in Taipei on Monday and the exhibition halls open Wednesday, and by any measure this is the largest edition SEMI has run: more than 1,300 exhibitors, 4,300 booths, 18 national pavilions, over 100,000 people from 65 countries. Numbers like that are easy to skim past. The substance underneath them matters more, because the question dominating almost every session is not which accelerator is fastest. It’s whether the connections between accelerators can keep up at all.
That shift has a strategic consequence that hasn’t been priced into most Western discussions of Taiwan risk. For four years the story has been diversification — Arizona, Kumamoto, Dresden, the slow migration of leading-edge capacity off the island. Meanwhile a new layer of the stack is being built, and it is being built here, almost entirely.
Why the wires became the problem
Frontier training compute is growing four- to fivefold a year. Power, not transistor count, is the binding constraint, and inside a large AI cluster the energy spent shuttling data between chips can exceed the energy spent computing on it. Copper is running into its limits on bandwidth, on reach, and above all on watts per bit. As clusters grow from a single rack to multi-rack systems (Broadcom’s core switching people are talking publicly about interconnecting a million accelerators), the electrical link stops being an implementation detail and becomes the ceiling.
The answer the industry has settled on is optics moved inside the package. Co-packaged optics puts the transmitters and receivers that carry data as light directly alongside the switch chip or processor, instead of leaving them in pluggable modules out on the board. Light is far cheaper than copper in power terms over distance. The engineering problem is that optical components have historically been assembled in a completely different manufacturing culture from silicon, with different yields, different test methods, and different tolerances.
Solving that is a packaging problem. Packaging is Taiwan’s.
What is actually being commercialised
The technical program this week reads as a production schedule rather than a research agenda, which is the real news.
TSMC is pushing its COUPE photonic engine platform. ASE is building out a CPO ecosystem across subsidiaries and its VIPack platform. Powertech, working with Broadcom, plans low-volume optical engine component production this year and integration into CPO switches during 2027. Largan Precision — an optics company that made its name in phone camera lenses — has a fiber-array order in hand and is preparing trial production on its first automated CPO line. Materials Analysis Technology is installing silicon photonics test equipment, which is the least glamorous item on the list and possibly the most telling, since nothing scales until it can be tested at volume.
Around that core sits the substrate work. AUO and Corning are showing glass core substrates that pair semiconductor-grade glass with AUO’s redistribution layer process, aimed at the warpage and signal-loss problems that appear when AI processor dies get very large. ASMPT is presenting thermocompression and hybrid bonding tooling for the same set of applications. Suppliers from outside the island are present in force, but they are present as vendors into a Taiwanese assembly base.
The equipment numbers give the scale. Global semiconductor manufacturing equipment sales hit $135.1 billion, up 15% year on year. SEMI’s 300mm outlook puts fab equipment spending at $133 billion this year and $151 billion in 2027, the first time the industry crosses $150 billion in a single year, with roughly $155 billion projected by 2029. TrendForce expects AI server shipments up more than 28% in 2026, with ASIC-based systems taking 27.8% of shipments, their highest share since 2023.
The concentration argument
Here is the part worth sitting with. Fab diversification addresses one specific dependency: wafer fabrication at the leading edge. It does not address advanced packaging, and it does very little for the photonics-plus-packaging hybrid now being industrialised.
CoWoS capacity has been sold out globally. TSMC has revived Longtan for 1.4nm work and for a CoWoS hub, and that capacity is spoken for too. The optical layer is being stacked on top of a bottleneck that already exists, using the same test houses, the same substrate suppliers, the same OSAT floor space, the same engineers. A CPO switch that fails at Taiwanese assembly cannot be reassembled in Arizona, because the line, the tooling, and the accumulated yield knowledge aren’t there and won’t be for years.
The speaker list makes the dependency explicit rather than hiding it. Google’s Amin Vahdat gave the opening keynote, his first public keynote in Asia, on AI infrastructure and high-speed networks. Nvidia’s CTO Michael Kagan closes. Meta sent its VP of AI and compute infrastructure. Infineon sent its COO. These are the buyers, and they came to Taipei because this is where the physical problem gets solved. SEMI also arranged something it has never done before: a closing fireside chat putting five Taiwanese supply chain leaders on one stage, spanning IC design, wafer manufacturing, packaging and test, substrates, and electronics manufacturing. That lineup is a map of the chokepoint.
What it means for cross-strait risk assessment
Analysts who model a Taiwan contingency tend to reason in terms of fab output and how quickly it could be replaced. That model is already dated. The exposure is shifting toward a layer that is newer, thinner, more specialised, and more concentrated than wafer fabrication ever was — and one that most substitution studies don’t even enumerate.
There is also a timing asymmetry. Optical interconnect capacity is being committed now, on 2027 and 2028 production schedules, at a moment when nobody is going to slow down a hyperscaler roadmap in order to build redundancy into a supplier base outside Taiwan. Buyers with capex plans this aggressive take the capacity where the capacity exists.
Beijing does not need to read SEMI’s press releases to know this. The exhibition floor is the disclosure.
The industry spent four years working to make Taiwan less indispensable. This week in Taipei it is quietly making Taiwan indispensable in a new way.