Silicon Photonics: From Quiet Plumbing to Strategic Weapon
Silicon photonics is the use of optical components built on silicon chips to move data as light instead of electricity, allowing AI data centers to connect processors at far higher bandwidth and lower power than traditional copper links while shrinking link distance limits and energy costs across massive computing clusters.
The core takeaway is blunt: optical interconnects for AI are now as strategic as GPUs, and the fight over who controls them will shape the next era of AI infrastructure. The race to solve AI's copper bottleneck has pushed optical transceivers and co-packaged optics from niche technology into the center of AI infrastructure geopolitics. Photonics, once treated as background plumbing, is now the bottleneck and the bargaining chip. As clusters scale into hundreds of thousands of processors, the limiting factor is no longer how fast chips can compute, but how fast silicon photonics chips and AI data center transceivers can move data between them.
Why Optical Interconnects Now Matter as Much as GPUs
AI models are distributed across thousands of accelerators, which must exchange vast amounts of data many times per second. Copper links cannot keep up: as per‑lane speeds approach 200 Gbps, attenuation and crosstalk make passive copper unusable beyond short distances. Optical transceivers solve this by turning electrical signals into photons and back again, carrying more data across longer distances with lower power. In other words, scale‑out AI dies without light. The industry response has been a scramble from 800G to 1.6T modules and co‑packaged optics, which move optical engines closer to switch ASICs and, eventually, accelerators themselves. The race to solve this photonics bottleneck is becoming as important as securing GPUs and high‑bandwidth memory. If you care about AI infrastructure, you now have to care about the wiring behind the rack as much as the processor inside it.
Tower Semiconductor and the New Industrial Ground Game
While headlines obsess over accelerator chips, the quiet industrial story is who will fabricate the silicon photonics chips that make AI data center transceivers work. One foundry stands out for leaning into this niche: it produces specialty silicon photonics and silicon germanium chips used in high‑speed optical modules. Its recent plan to repurpose an existing facility for 300mm silicon photonics and advanced packaging, alongside boosting output at another fab, is a concrete bet that demand for optical interconnects AI hardware is only beginning. This first expansion track is slated to reach full production readiness in the fourth quarter of 2027. A second track, still subject to agreements, would add another 300mm facility next to its existing site. In a market distracted by a sharp pullback in semiconductor stocks, this kind of long‑dated factory plan signals that the smart money expects the optical layer to remain a growth engine even after the current AI hype cools.
China’s Lead and the Tangle in the Photonics Supply Chain
On the ground, one side currently dominates the photonics supply chain. Commercially, Chinese vendors account for nearly two‑thirds of global optical transceiver unit shipments and roughly 60% of datacom transceiver revenue. Innolight alone holds about 27% revenue share, ahead of Western rivals whose shares are in the teens or single digits. Yet this dominance hides deep mutual dependence. Chinese module makers often combine digital signal processor chips from Western firms with laser components from other non‑Chinese suppliers. Western companies, in turn, rely on Chinese indium phosphide — now under export controls — for their lasers. Photonics is “a market defined by deep mutual entanglement,” with supply‑chain ties that cut across the same borders politicians want to harden. That entanglement is exactly what makes photonics such a powerful geopolitical lever — but also what makes any clean decoupling fantasy.
The Emerging Ban and What Comes Next for AI Infrastructure Geopolitics
Policymakers are now turning this entangled photonics supply chain into a pressure point. Regulators are drafting a rule to deny new models of Chinese optical transceivers the authorization needed to be imported, marketed, or sold domestically. The aim is clear: keep Chinese hardware out of the optical layer of future AI data centers, even if that means short‑term disruption. This plays out against a broader backdrop in which AI's latest bottleneck has become a tech‑war battleground and the AI supply chain is being weaponized for summit‑level negotiations, turning silicon photonics into a bargaining chip. Yet the rule is only a draft, open to comments, carve‑outs, and court fights. Analysts warn that Western suppliers lack the packaging capacity and manufacturing scale to replace Chinese modules within a few years, and that the AI infrastructure ecosystem remains heavily dependent on Chinese optical‑module vendors. In trying to wall off Chinese transceivers, policymakers risk slowing their own AI ambitions.
