AI Boom Tightens Optical Supply Chains and Fuels Huawei’s Chip Resurgence

Business news

The AI explosion that began in 2022 is now reshaping the entire hardware stack. Nvidia’s CEO Jensen Huang told investors that the company’s annual spend on Taiwan‑based AI supply chain partners could reach $150 billion, a figure that dwarfs its historic GPU‑only spend. AMD’s Lisa Su, Intel’s Pat Gelsinger, Qualcomm’s Cristiano Amon and Arm’s Simon Segars are all scheduled to meet TSMC executives in the coming weeks, not for the Computex showcase but to lock down chips, advanced packaging, substrates, PCBs, cooling and power modules.

At the heart of the new supply‑chain scramble is a shift from traditional copper‑heavy interconnects to co‑packaged optics (CPO) and compact universal photonic engines (COUPE). TSMC’s COUPE merges photonic and electronic dies in a single package, while CPO places the optical engine directly beside the processor, cutting latency and power use. Early adopters—Nvidia, Google and Amazon—are already qualifying CPO for their largest AI clusters.

Market context

Optical component shortages

• Fiber: Nvidia signed a long‑term deal with Corning to multiply U.S. fiber capacity tenfold. The agreement, valued at roughly $3 billion over five years, underscores fiber’s emergence as a strategic AI asset.
• Laser sources: Indium phosphide (InP) substrates, essential for high‑speed lasers, are tightening. China’s recent export curbs on indium have cut global InP supply by an estimated 12 %, pushing prices up 35 % since Q1 2025.
• Pluggable transceivers: Chinese firms Zhongji Innolight and Eoptolink have seen revenues jump 800 % and 900 % respectively since 2022, driven by orders from Google, Nvidia and Amazon. Their rapid growth highlights how a once‑niche market has become a critical AI bottleneck.

Private‑equity retreat from China

Geopolitical pressure is prompting a wave of exits. Princeton Digital Group’s potential $1 billion sale of data‑center assets across six Chinese cities signals the end of a decade‑long PE influx that poured $12 billion into Chinese cloud infrastructure. Domestic buyers are stepping in, but the overall foreign capital exposure to Chinese AI hardware is falling sharply.

Huawei’s modest chip comeback

Six years after the U.S. black‑list forced Huawei out of advanced fabs, the company unveiled a new Kirin line aimed at mid‑range smartphones. Chief semiconductor officer He Tingbo emphasized a design philosophy that works around the lack of EUV lithography, relying on 28 nm and 40 nm process nodes available at domestic foundries. While the chips will not match the performance of flagship Snapdragon or Apple silicon, Huawei projects annual shipments of 30 million units by 2028, enough to generate $2.5 billion in revenue and re‑establish a foothold in the Chinese market.

Memory‑centric AI

SanDisk CTO Shannon Miller warned that AI workloads are becoming memory‑centric. Micron and SK Hynix have each crossed the $1 trillion market‑cap threshold, driven by demand for high‑bandwidth DDR5 and emerging HBM3 stacks. Customers are now signing five‑year supply contracts for NAND and DRAM, a practice unheard of before the AI surge.

What it means

1. Investors should watch optical‑component suppliers – Companies like Corning, Lumentum and the Chinese transceiver makers are likely to see earnings volatility as AI demand outpaces supply. Tight InP markets could also benefit niche players that secure alternative substrate sources.
2. CPO may become a decisive differentiator – Early adopters that qualify co‑packaged optics will achieve lower total‑cost‑of‑ownership for AI clusters, potentially widening the gap between cloud giants and smaller regional players.
3. Huawei’s roadmap signals a partial re‑entry – While the new Kirin chips will not challenge high‑end global players, they provide Huawei with a domestic revenue stream and a platform to iterate toward more advanced nodes as China’s own EUV capabilities mature.
4. Memory‑centric strategies will dominate AI system design – The shift toward larger, faster memory pools suggests that companies investing in HBM3, DDR5‑E and emerging NVDIMM technologies could capture a larger share of AI spend than pure compute‑focused firms.
5. Geopolitical risk remains a key factor – The retreat of foreign private‑equity from Chinese data‑center assets and ongoing export controls mean that supply‑chain diversification will be a priority for multinational chipmakers.

Overall, the AI boom is not only accelerating demand for silicon but also exposing hidden constraints in the optical layer that connects chips to each other. Companies that can navigate these bottlenecks—whether through CPO adoption, strategic fiber contracts, or diversified memory sourcing—are positioned to capture the next wave of AI‑driven growth in Asia and beyond.