For decades, copper wiring has carried data inside the world's data centers. That era is now drawing to a close. As artificial intelligence pushes computing infrastructure to unprecedented scale, the physical limits of copper, most notably its bandwidth ceiling, heat, and power consumption, are forcing the industry to make a fundamental switch from electrons to photons and from electrical signals to light.
The shift is not theoretical. In 2026, several of the world's most powerful technology companies are moving silicon photonics and co-packaged optics from laboratory prototypes into commercial production. At the center of this transition sits TSMC, whose new manufacturing platform is positioning Taiwan as the critical enabler of a technology that could reshape how every major AI data center on the planet is built.
Modern AI clusters connect thousands of GPUs, each exchanging vast amounts of data at extreme speed. Traditional copper interconnects work well at shorter distances, but as per-lane data rates approach 400 gigabits per second, copper struggles. Signal integrity degrades. Power consumption per bit rises sharply. Heat builds up in already-dense racks.
Silicon photonics solves these problems by routing data as light through tiny waveguides etched into silicon chips. Optical signals travel farther with less loss, consume far less power per bit, and generate substantially less heat. A single data center with 1,000 network switches could save over 2.6 megawatts by switching to optical interconnects. Industry participants broadly agree that 400 gigabits per lane is the tipping point where optical becomes cheaper and more efficient than copper. That threshold is now being crossed.
TSMC has begun production of COUPE, with volume shipments ramping through the middle of the year. The first-generation platform delivers 1.6 terabits per second of data transfer via standard connectors, double the current copper Ethernet maximum. A second generation, expected in 2027, will push throughput to 6.4 terabits per second using TSMC's CoWoS advanced packaging with co-packaged optics integration. A third generation targeting 12.8 terabits per second is already on the roadmap.
The power efficiency gains are striking. COUPE reduces per-port power consumption from roughly 30 watts to around 9 watts, a 3.5-fold improvement. TSMC claims the platform delivers five to ten times better overall power efficiency and ten to twenty times lower latency compared to conventional optical modules.
TSMC is also filing silicon photonics patents at roughly double the rate of Intel, signaling its intent to dominate this space the way it dominates advanced chip manufacturing.
The two most significant early adopters of COUPE are Nvidia and Broadcom, whose chips already sit at the heart of most AI data centers.
Nvidia has developed two new switch platforms built on TSMC's COUPE technology: Spectrum-X Photonics for Ethernet networks and Quantum-X Photonics for InfiniBand. The top-end Spectrum-X configuration delivers 400 terabits per second of aggregate throughput across 512 ports of 800 gigabit-per-second connections, while requiring 3.5 times less power than conventional approaches. Nvidia says the switches will enable AI factories to connect millions of GPUs across sites. Its Quantum-X Photonics InfiniBand and Spectrum-X Photonics Ethernet switches are expected to be available this year.
Broadcom, meanwhile, is bringing its Tomahawk 6-Davisson switch to market, a 102.4 terabit-per-second platform that also uses TSMC's COUPE technology to stack photonic engines directly onto switching silicon. Broadcom has already demonstrated third-generation co-packaged optics at 200 gigabits per lane and is developing a fourth-generation solution at 400 gigabits per lane.
Both companies are converging on the same conclusion that optical integration is no longer optional for next-generation AI infrastructure.
The momentum extends well beyond the chip giants. Ayar Labs, a silicon photonics startup backed by Nvidia, AMD, and TSMC, demonstrated the industry's first working COUPE-based optical connectivity solution in late 2025, alongside Taiwan's Alchip Technologies. AMD has invested US$280 million in a Taiwan-based silicon photonics research center. Taiwan's Silicon Photonics Industry Alliance now counts more than 30 member companies.
TSMC is not without challengers. Samsung Foundry plans to launch its own optical engines in 2027, with turnkey co-packaged optics services following in 2029. Samsung's pitch centers on its ability to offer memory (including HBM), foundry services, advanced packaging, and silicon photonics from a single supplier. Intel, which has long invested in silicon photonics through its own research programs, offers advanced packaging platforms that could serve as alternatives, though its foundry business has yet to secure major external photonics customers.
For now, however, TSMC's roughly three-year head start in volume manufacturing, combined with its existing relationships with Nvidia and Broadcom, gives it a commanding position.
The commercialization of silicon photonics in 2026 marks a rare inflection point. Just as fiber optics replaced copper in long-distance telecommunications decades ago, co-packaged optics is now poised to replace copper inside the data center itself, the last major domain where electrical interconnects still dominate.
Every major AI data center built from the late 2020s onward will likely incorporate optical interconnects as standard infrastructure. The companies that control the manufacturing platforms, optical components, and testing technology for this transition stand to capture a significant share of the trillions being invested in AI infrastructure globally.
Taiwan, with TSMC's COUPE platform at its center and an expanding ecosystem of component suppliers, is once again positioned at the critical juncture of a global technology shift. When copper gives way to light, much of that light will be made on the island.