As photonic technologies and optical networking become critical to AI performance and energy efficiency, the US and China are entering a new phase of tech competition centered on controlling the light-based hardware and supply chains that connect next-generation data centers.
In August, news broke that the Trump Administration might ban new Chinese-made optical transceivers, thumb-sized modules that convert electricity into light and sit at the end of nearly every fiber in an AI data center. With TrendForce data showing that China has 56 percent of global transceiver manufacturing capacity, industry was quick to argue that the United States could not build data centers without them. When the Federal Communications Commission published its final rule in September, it left transceivers off the list.
It is understandable that the FCC held back at the moment, as dependence runs in both directions. Any rash move to restrict trade could quickly escalate, backfire, and harm US industry. However, that does not mean strengthening the US semiconductor stack is the wrong objective. Washington is right to look closely at next-generation optical and photonic technologies, a critical area that has received little policy attention.
The twenty-first century is likely to be defined by light. The transceiver is the visible layer. In the future, photonic technologies, the component that encodes data onto light, will probably sit inside every module. The race to harness the photonic technologies could reshape the balance of technological power between the United States and China.
Frontier AI models need tens of thousands of processors acting as one computer. They cannot afford to wait for a slow network between these processors to return results. This is a problem particularly worth worrying about because the cost and the technical standards are getting more demanding.
Power is now the ceiling on new data centers, so every watt saved on data transmission is a watt available for compute. A large cluster of 128,000 GPUs, for example, can require almost half a million pluggable transceivers to keep them all connected, each one drawing between 9 and 15 watts. That works out to roughly 4.5 to 7.5 megawatts for moving data alone, before any computing is done.
A new technique called co-packaged optics has emerged. It places the optical engine right next to the switch chip, removing the pluggable module. Nvidia claims its photonic switches are 3.5 times more power-efficient than traditional networking, and Broadcom says its latest co-packaged switch cuts optical interconnect power by about 70 percent compared with pluggable modules.
Since 2015, the PRC has made photonic technology a national priority. The 14th Five-Year Plan designated it a strategic technology, directing funding toward national labs from agencies such as the Ministry of Science and Technology and the National Natural Science Foundation and establishing dedicated research and development programs. This allowed for long-term capability-building rather than immediate commercialization. The possibility of “leapfrog development” in chip design has driven Beijing’s strategic obsession with photonics over the last decade. That interest has accelerated since the US restrictions on advanced chipmaking machines arrived in 2022, because photonic chips offer a way to ease the bottleneck: they can deliver efficiency gains through design rather than through the most advanced lithography.
Last year, Huawei founder Ren Zhengfei told People’s Daily that China could address its lack of advanced chips by pursuing “group computing”. Optical networking lays the ground for group computing. Huawei’s plans for the Atlas 950 SuperPoD, due in the fourth quarter of 2026, suggest that its customers will be able to do this without worrying about slow networking, by clustering up to 8,192 Ascend processors interconnected by an all-optical fabric called UnifiedBus. But early industry analysis of its predecessor, CloudMatrix 384, found that it beat the raw compute of Nvidia’s GB200 NVL72 rack by using roughly four times the power. In optics, the gap between China and the rest of the world may be narrowing faster than it is in logic chips.
Washington should think carefully about each layer of the optical stack. On the US side, the strengths are more entrenched. The Broadcom–Marvell duopoly makes the all-important signal-processing chips. High-speed lasers come mainly from Lumentum and Coherent in the United States and from Japanese suppliers. Chinese firms have captured little of that market. Nevertheless, China holds a major advantage at the bottom of the chain: indium phosphide, the substrate for those lasers.
Indium phosphide lacks the glamor of more commonly noted semiconductor inputs, but it has become a pressure point. In February 2025, China added it to its export-control list, so exporters now need licenses. The price of a six-inch wafer has since risen about 250 percent, and China refines about 70 percent of the world’s indium. The substrates themselves are made mostly by AXT and Sumitomo, but AXT, though American-listed, makes its indium phosphide in China through its Beijing subsidiary Tongmei, and its shipments have been delayed by Chinese export permits.
In July, the Commerce Department signed a letter of intent for a $300 million CHIPS award to GlobalFoundries with a roughly 1 percent government stake to accelerate US silicon photonic leadership. This is a solid starting point, but it is not enough to address China’s whole-of-society push into photonic and optical equipment.
First, the US government should publish a photonic strategy that identifies critical supply chain layers the United States intends to own, such as foundry capacity, packaging, and lasers. Such a strategy should also recognize that judging capability solely by the specifications of an individual chip can miss important system-level gains. Improvements in optical interconnects, for example, can allow processors to communicate and operate together more efficiently, boosting overall computing performance. Establishing state-backed pilot lines and facilities where companies can test their optical networking technologies and photonic chips in next-generation data center environment will be essential.
Second, the United States must secure the strategic materials that enable the photon. Indium phosphide substrates and high-speed lasers are the weak point. Public money should support non-Chinese substrate capacity. Nvidia’s $4 billion investment in Lumentum and Coherent covers lasers; the substrate beneath them still needs a sponsor.
Third, the US government should help coordinate the industry to establish relevant standards. There needs to be shared standards and independent testing so that chips, lasers, connectors, and networking equipment from different companies can work together. Likewise, the US should also negotiate practical partnerships with Taiwan, Japan, and European partners to ensure that critical components meet these standards.
The chip war was fought over what could be etched. The next phase will be fought over what can be connected, and over who controls the light and photon inside the connection.
