U.S. Big Tech companies are pouring astronomical sums into artificial intelligence infrastructure, but they have not widened the gap in computing power over China as much as expected, according to a new analysis. China trails the United States in absolute scale, yet cheap energy and government support are helping it close the distance faster.
Hong Kong's South China Morning Post reported on the 5th, citing a recent report by credit rating agency Moody's, that U.S. hyperscalers — operators of large-scale data centers — are spending 5.6 times more than their Chinese rivals, while the actual gap in physical computing capacity is only twofold.
Moody's said the five largest U.S. hyperscalers — Microsoft, Amazon Web Services, Alphabet, Meta and Oracle — along with AI cloud provider CoreWeave, are set to spend $785 billion in capital expenditure this year alone, a figure projected to reach $1 trillion in 2027. By comparison, capital spending by major Chinese technology companies is expected to total about $140 billion this year and $165 billion in 2027, far behind the U.S. total. Based on projected spending for this year, U.S. investment is about 5.6 times China's.
The difference in installed data center capacity, however, was about twofold. That is still a wide margin, but not as wide as the spending gap. Figures from the International Energy Agency cited by Moody's show U.S. installed data center capacity at 52 gigawatts as of the end of 2025, against 28 gigawatts in China. By 2030, U.S. capacity is projected to reach 100 gigawatts and China's 67 gigawatts, narrowing the gap further. China's data center capacity is expected to grow at an average annual rate of 19%, outpacing the U.S. growth rate of 14%. "While power output does not directly correspond to AI processing capability, it is clear that the gap in capital spending is far larger than the gap in physical capacity," the SCMP said.
Model performance also reflects China's catch-up. Leading Chinese models — Moonshot AI's Kimi K3, Z.ai's GLM-5.3 and Alibaba Group Holding's Qwen3.8 — now stand alongside the world's best open-weight competitors on benchmark platforms such as Artificial Analysis and Chatbot Arena. Open-weight models release the weights, or parameters, of a trained AI so anyone can download, run and fine-tune it, but training data and code are not disclosed, making them more restricted than fully open-source models.
Chinese Chips Still Draw Four Times the Power of Nvidia's
Moody's said policy incentives from the Chinese government have helped the country secure more computing resources per dollar invested. In particular, it credited China with sharply lowering data center construction costs by placing power-hungry, heat-generating computing work in inland regions with ample land and cooler climates. The SCMP added that subsidies, fast-track approvals and tax breaks are also being extended to AI companies.
There are limits, of course. The biggest constraint cited is restricted access to Nvidia's most advanced chips under U.S. sanctions. China's homegrown chip technology has improved considerably, but power consumption remains high. Huawei Technologies' CloudMatrix 384, for example, is reported to deliver twice the computing performance of Nvidia's GB200 NVL72 while consuming about 4.1 times more power.
Another hurdle is that monetization of China's low-cost AI models remains at an early stage. In recently announced quarterly results, Microsoft posted $59 billion in cloud revenue, Amazon $42 billion and Alphabet $25 billion. Alibaba, by contrast, reported $7 billion and Baidu about $1 billion.
Still, there are signs the gap is beginning to narrow. Alibaba's cloud business posted 45% revenue growth in the June quarter, its fastest pace in 22 quarters, as demand for AI-related products accelerated, the SCMP said.
