The clash between China and the US spans data centers and chips, energy and model training, all the way to talent, networks, and military uses
AI Brief
Your highlights
Between the US midterm elections and the resignation, after a scandal halfway between espionage and sex, of Nvidia founder Jen-Hsun Huang, there is one piece of news that has gone unnoticed in recent days. And it demonstrates, once again, the existence of a cold war over artificial intelligence between the United States and China. Beyond the statements and conciliatory face-to-face meetings.
For a few weeks now, 600,000 Nvidia chips have been sitting idle in Hong Kong, waiting to find out whether they will be allowed into China. They are the H200s, among the most powerful processors for training AI models. But in this case Donald Trump and the US have nothing to do with it; this time it is Beijing that is holding back imports: it does not want its technological future to depend on a foreign company.
US-China, the technological gap between models has narrowed
It is a story that captures the situation well: AI is now a matter of national sovereignty, security and industrial power. Because it concerns the ability to develop and govern the machines that produce (and will increasingly produce) knowledge, wealth and power. The game is complicated and there is no single playing field. There are data centers (and therefore chips), but also the energy to power them, the data to train the models, the communication networks, the talent and the military applications.
A report by the Stanford Institute for Human-Centered AI offers a picture of this competition. The technological gap between the best US models and the Chinese ones has practically closed: as of March, the American advantage was just 2.7%. Yet, on the private investment front, the distance remains enormous. In 2025 the United States spent 285.9 billion dollars on AI, compared to China's 12.4 billion.
But economic superiority does not coincide with technological superiority. Chinese AI DeepSeek has proven it: developing competitive models does not require the computational resources of US Big Tech. Which is why they are going through a complicated period. The costs of using their models are decidedly higher than the Chinese ones. So, even in the West, Beijing's algorithms are gaining ground. One case above all is that of Revolut, which recently moved its roughly 2,000 software engineers from Anthropic and OpenAI services to GLM, a model from China's Z.ai. According to Revolut, the company now spends about a tenth.
"China has all the conditions to become the great protagonist of AI. It has a highly developed power grid, enormous production capacity, batteries, water availability and access to the materials and rare earths needed to build data centers and chips. To this is added the expertise in developing code and technology, with costs that are rapidly approaching zero. It is a competitive advantage that is difficult to close," explains Francesco de Mojana, former Permira partner and now a private investor and board member of several companies, including Wix, an Israeli tech company listed on the Nasdaq. "Chinese models, moreover, are 'open weight', unlike US ones which are proprietary. The former can be downloaded, run locally and customized, offering greater possibilities of control over data and privacy," says Laura Sanità, professor at the Department of Computing Sciences of Bocconi University.
To the competition between models is added another dimension, a geographical one. The US hosts 5,427 data centers, more than ten times as many as any other country. Numbers that, however, do not make them autonomous. The production of the most advanced chips, in fact, depends on the world's largest semiconductor manufacturer, Taiwan's Tsmc. American power, therefore, also rests on an industrial dependency located a few kilometers from China and one of the areas of greatest military tension on the planet.
Speaking of the military sphere, a recent report by the US Government Accountability Office underlines how America's adversaries are investing in AI capabilities that can alter the balances of international security. The problem is not only who owns the most powerful algorithms, but who manages to integrate them most quickly into their military structures. There is, in fact, a misconception to overcome: thinking that the AI competition concerns only who will succeed in building the smartest model. What is decisive is who owns the chips, who powers the data centers, who governs the platforms and who sets the conditions of access.
And what about us? As is well known, the risk is that Europe becomes the territory where models developed elsewhere are used. In June, the European Commission presented a package to strengthen technological sovereignty, intervening on semiconductors, cloud, AI and open source. It is a necessary direction, but will it be enough? Investments, industrial capacity, research, energy, infrastructure and above all a common policy are needed.
"Europe had an advantageous position in research, but internal divisions have squandered many opportunities. If we only use AI developed outside Europe, we are opening our companies and our data to technologies we do not control. And we cannot rule out that, through AI agents, new forms of vulnerability in access to our systems may be created," De Mojana continues.
The problem for Europe, moreover, is that AI is so pervasive that falling behind in this sector means condemning ourselves to a secondary or tertiary role in many markets. And in research as well. "There is an accessibility problem: the licenses of proprietary models can be so expensive as to make their use unsustainable for part of the scientific community. We cannot allow a series A and a series B science to emerge, in which only those with sufficient economic resources can access the best tools," Sanità concludes.