AI is no longer just software: the real battle is over energy and infrastructure
From the showdown between ChatGPT and DeepSeek to the race to build data centres, secure electricity supplies and control the systems that link artificial intelligence to businesses. The United States, China and Europe are investing in different parts of the same supply chain: the real competitive advantage will lie in having computing power available, secure data and platforms that are difficult to replace
Key points
Artificial intelligence is undergoing a transformation. Until recently, competition seemed to focus mainly on models: who had the most powerful algorithm, who achieved the best results in tests, and who could produce a response at the lowest cost. But now the tide is turning and the challenge is shifting to a much more concrete and tangible level, towards what is needed to make those very same models work on a day-to-day basis and integrate them into businesses and public administrations.
A few figures may help to illustrate the scale of the change we are talking about. As of 31 August this year, China had registered 1,112 generative AI services, compared with 302 at the end of 2024 and 748 at the end of 2025. The same system also listed 731 applications linked to already registered models. This means that DeepSeek – namely China’s largest AI company, founded by Liang Wenfeng, which develops large-scale open-source language models – whilst undoubtedly remaining a key player, is no longer sufficient on its own to describe the market in Beijing. This is because there are complementary companies that build the entire ecosystem required to use the model: Alibaba develops Qwen, Tencent links its models to platforms and developer tools, Baidu integrates Ernie with the cloud, and Huawei links Pangu to its Ascend accelerators.
The price of the token matters less than the cost of the completed work
By September 2026, the Chinese market was already showing very aggressive price segmentation for inference. DeepSeek V4.1-Flash quoted $0.30 per million input tokens not present in the cache and $1.20 per million output tokens at peak rates; MiniMax M2.7 used the same basic pricing structure; Tencent offered GLM-5.3-Flash via TokenHub at 0.8 renminbi per million input tokens and 2.8 renminbi per million output tokens.
However, the nominal cost of the token is not the sole factor that will determine industry-wide adoption. A model that costs less but requires more checks, greater human supervision or a longer validation process may ultimately prove more expensive. That is why, until a model reaches the minimum level of reliability required by the process, quality takes precedence over price. Once several systems exceed that threshold, other variables become decisive, such as integration times, latency, security, data residency, the cost of computational capacity, the possibility of private deployment and compatibility with existing software.
The US race to catch up with the Dragon
Even the United States has now taken the competition beyond the laboratories. Microsoft, AWS, Google, OpenAI, Anthropic, Meta and Nvidia dominate different segments of the chain, such as models, cloud services, accelerators and management software. A few examples: OpenAI, in April 2026, stated that Stargate had already exceeded 10 GW of secured infrastructure capacity in the United States. In the same month, Anthropic announced, in partnership with AWS, up to an additional 5 GW and over $100 billion worth of technology over a ten-year period. Microsoft indicated around $175 billion in capital expenditure for 2026, whilst Alphabet forecast between 175 and 185 billion.
