Product value and waste. And here lies the challenge for data centres

Among the reasons why so many voters backed the Democratic Party last Tuesday was a widespread concern about data centres. Land use and high demands for water and electricity – by 2025, electricity demand from US data centres already accounted for around 5 per cent of the total – make them a threat to environmental sustainability. But economic sustainability is by no means a given either. The stakes are enormous. Following the October 2026 quarterly results, capital expenditure by the major technology investors is estimated at over 800 billion dollars for the year. These investments contribute significantly to US growth, although estimates vary.

Why build new data centres? The idea is that improvements in language models occur at relatively constant rates as the number of parameters, the volume of data and the computing power used in training increase. The industry’s strategy therefore seems to be: more computing power means better models and a market advantage. However, there is a logical leap here. The empirical patterns observed so far do not correspond to a law of nature. There will come a time when an additional data centre will cost more than the benefits it generates, even in terms of learning. The issue is complex and involves economic factors, not just the functioning of LLMs. In the race to accumulate computing power, the perception that geopolitical supremacy is at stake in this arena also plays a part. If American hegemony is on the line, the temptation is to write a blank cheque. Some estimates suggest annual revenues in the region of trillions to justify the projected capacity. Revenues are growing rapidly, but they are not easy to quantify. There is a difference between the run-rate – that is, the annualised value of the latest available data, projected over twelve months – and the revenues recorded in the accounts.

Not all data centres are the same. Buildings, electrical connections and cooling systems can last for decades. Accelerators and servers have much shorter technology cycles. Extending their lifespan reduces depreciation, not obsolescence. Let’s imagine a scenario in which AI truly transforms the world of work, but revenue grows more slowly than installed capacity. We can also imagine that efficiency improves more rapidly than demand. It is the infrastructure gamble that would fail: the technology would survive, but the first generation of investors would take a real hit.

It wouldn’t be the first time. Nineteenth-century railways and fibre-optic networks have already produced this paradox: investors ruined and companies benefiting from infrastructure that survived the bubble. With AI, however, there is an additional problem: the rapid obsolescence of investments. DeepSeek has shown just how quickly the costs required to achieve competitive performance can fall. The Chinese have, broadly speaking, taken a different approach. Through the ‘Data East, Computing West’ programme, launched in 2022, they have sought to shift activities that are less sensitive to latency – that is, those for which a slight delay in data transmission does not compromise the service, such as training or storage, unlike real-time applications – to regions with more favourable energy costs. This has resulted in massive infrastructure, perhaps too much so: according to media reports, many facilities have remained underutilised.

Paradoxically, the need to change course has brought the Chinese perspective closer to the consumer’s point of view and further away from ambitions of grandeur. Rather than seeking the most powerful model, the Chinese are aiming for widespread adoption. The official objectives of the latest government initiatives concern the roll-out of artificial intelligence in manufacturing, services and public administration. Open-weight models – that is, models whose trained parameters are publicly available, downloadable and usable by anyone, without necessarily going via the manufacturer’s servers – and contributions to computing power aim to reduce barriers to entry. There is, due to US administrations (both Trump and Biden), a geopolitical constraint: US controls are pushing Beijing to support domestic hardware even when it is less efficient. Simply put, China leverages its relative abundance of energy and demand-side subsidies—with reductions of up to 50 per cent on electricity bills for data centres using Chinese chips—to compensate for the shortage of advanced silicon.

The real question is not who will win the race, but what relationship exists between computing power and the value produced. The US is banking on the possibility that ever-greater computing power will generate benefits profitable enough to recoup substantial investments. China is banking on the value of less expensive computing capacity, distributed across the real economy. If AI were to evolve from an advisory tool into a system capable of autonomously performing complex tasks, the demand for computing power could grow beyond current forecasts. Greater efficiency would not necessarily reduce consumption: it would make applications that are currently impossible viable, ranging from scientific research to medicine, from manufacturing to services. It is unclear, however, what the business model will be. Earning more tokens does not necessarily mean making more money, let alone generating a return on the capital employed. Investors’ reaction to the half-yearly results of high-tech companies in recent weeks may perhaps herald a market correction.

It is normal for there to be waste during the early stages of a technology’s development. The crucial question is: how much, and on what scale? We will only find out in due course. Both the advocates and the critics of the proliferation of data centres may be right. But not necessarily for the reasons they think.