The artificial intelligence boom has a very physical price tag. Behind every chatbot answer, image generator and coding assistant sits a growing network of data centres filled with expensive processors, cooling systems and power contracts. This week, fresh analysis of the industry’s spending plans has sharpened an uncomfortable question for investors and technology leaders alike: is the industry building infrastructure far faster than the revenue from AI products can realistically pay for it?
Reporting by The Register highlighted forecasts suggesting that planned spending on AI infrastructure could outstrip the revenue the sector expects to earn from selling AI services. A separate analysis argued, in stark terms, that paying for the current wave of data centres at prevailing wages would take the equivalent of hundreds of millions of jobs — more than the entire American workforce. The figures are illustrative rather than literal, but they capture the scale of the bet now being placed.
Why Data Centres Cost So Much
A modern AI data centre is not a warehouse with a few servers in it. Training and running large models requires dense clusters of advanced graphics processors, high-speed networking that lets thousands of chips work as one machine, and enormous amounts of electricity. Power, once a background utility cost, has become a strategic constraint. Operators compete for grid connections, sign long-term energy deals and, in some cases, look at dedicated generation to keep facilities running around the clock.
Then there is depreciation. The chips at the heart of the boom are extraordinarily expensive, and each new generation makes the previous one less competitive. If a facility’s hardware loses economic value in a few years while the building and power contracts run for decades, the maths only works if demand for AI computing stays high and prices hold up. That is a demanding set of assumptions.
The Revenue Side of the Equation
None of this means AI lacks real customers. Businesses are paying for coding tools, customer support automation, search features and productivity software, and consumer subscriptions are growing. Our coverage of how AI agents are quietly taking over everyday office work shows how quickly these tools are moving into everyday workflows. The open question is pace: revenue is growing quickly from a modest base, while capital spending has jumped to levels normally associated with railways, telecom networks and other historic infrastructure build-outs.
History offers both comfort and warning. The fibre-optic networks laid during the dot-com era eventually carried the modern internet, but many of the companies that paid to lay them did not survive to benefit. Excess capacity can be a gift to the next generation of users and a problem for the investors who funded it at peak prices. AI infrastructure could follow a similar pattern — socially useful, commercially painful for some of its earliest backers.
What Happens If the Maths Stays Tight?
If revenue does not catch up as fast as expected, several adjustments are likely. First, efficiency will matter more: smaller models, better software and chips designed for specific tasks can deliver similar results with far less computing power. Second, prices for AI services may rise, or free tiers may shrink, as providers try to align usage with cost. Third, consolidation could follow, with smaller operators renting capacity from giants or leaving the market altogether.
Energy politics will also intensify. Communities asked to host large facilities increasingly want to know what they receive in return — jobs, tax revenue and grid upgrades — and whether household bills will be affected. Regulators, meanwhile, are watching both competition and environmental impact more closely than they did at the start of the boom.
A Boom That Still Needs Its Customers
It would be a mistake to read the scepticism as proof that AI is a mirage. Electricity networks, cloud computing and mobile broadband all went through periods when spending ran ahead of proven demand. The technology is already embedded in software development, design, research and office work, and its capabilities continue to improve.
But the era of treating data centre announcements as self-evidently good news is ending. The next phase of the AI story will be told less in gigawatts and chip counts, and more in ordinary business language: customers, margins, renewal rates and cash flow. Companies that can show paying users for the capacity they build will thrive. Those relying on the assumption that demand always arrives in time will find that, eventually, the AI maths has to math.
