Beijing bet on leading the artificial intelligence revolution, but over-investment in data centers and chips has led to a dead end
What started as a frantic race towards a technological future may have turned into a new episode of an economic bubble. China, following the global surge in artificial intelligence (AI) driven by models like ChatGPT and Claude, allocated billions to build data centers across the country. Today, many of those buildings stand empty, equipped with state-of-the-art servers collecting dust. The term “AI bubble” is already being discussed in the Asian giant.
After years of unchecked growth in the real estate sector, many local governments saw AI as a way to revitalize the economy. The construction of data centers became the new engine, fueled by public funds, advantageous loans, and inflated expectations. The phenomenon even spread to remote regions, far from the main tech hubs, where the availability of cheap electricity and vacant land seemed to justify any investment.
“The enthusiasm for AI was an adrenaline shot,” said an executive in the sector. “They went from building homes to erecting data centers without being clear on what they were going to be used for.”
From training models to inferring them: the market reality
In recent months, the sector has radically changed course. While data centers were conceived to train large-scale language models—a resource-intensive task—the industry has shifted towards inference: efficiently executing pre-trained models quickly and with low latency. This shift has sidelined much of the infrastructure set up in locations far from users and telecommunications networks.
Chips like the NVIDIA H100, which once sold for over $28,000 on the black market, are now rented for a fraction of that price… and still, there isn’t enough demand to make them profitable. The supply of computing power far exceeds the actual needs of the market.
Uncontrolled investments and decisions without technical criteria
The problem has not only been technical. Executives lacking experience in artificial intelligence, companies without a track record in technology, and local officials more concerned with pleasing Beijing than with economic viability have fueled a runaway machine. In some cases, data centers were used as an excuse to access electric subsidies or government loans, without any real intention of putting them into operation.
There have even been identified cases of structural corruption, where infrastructures were built that never connected to customers or networks, serving only as a facade to justify public aid.
The NVIDIA H20 case: $16 billion in unused hardware
Meanwhile, the Chinese government and major tech companies have invested nearly €16 billion in NVIDIA H20 chips, specifically adapted for the Chinese market due to export restrictions imposed by the United States. These GPUs are designed for inference tasks, but their massive rollout has not been accompanied by a clear strategy for utilizing them.
The situation worsens with the emergence of models like DeepSeek R1, which can match ChatGPT’s performance with a fraction of the resources. This has changed the game: it’s no longer about who has more computing power but rather about who knows how to use it more efficiently.
From brick to silicon, and starting over
For many analysts, what is happening in China with AI resembles the “brick bubble” experienced by Spain. The difference is that now the product is data centers and hardware, not homes. The pattern repeats: over-investment, poor planning, speculation, and a market unable to absorb what has been built.
Some voices within the country are already calling for state intervention. The central government is expected to drive the redistribution of assets towards companies that are managing their technological resources properly, in an attempt to salvage what can be saved.
And now what?
With U.S. sanctions partially circumvented—after an apparent informal “agreement” that even included million-dollar dinners as a goodwill gesture—China now faces a phase of adjustment. The country will likely shift from a focus on hardware to a focus on software and services. The goal: to do more with less and regain the strategic sense of its investment in AI.
What seemed like a sprint towards global leadership in artificial intelligence has ended, at least for now, as a harsh reminder that technology requires planning, talent, and clear objectives. Building alone is not enough to lead.
Reference: El chapuzas informático