We're building data centres for a dream

The world is in the middle of one of the largest infrastructure buildouts in modern history, and almost all of it is a bet on a single technology: artificial intelligence. Big Tech is projected to spend roughly $650 billion on AI infrastructure in 2026 alone, up from $383 billion the year before and just $80 billion in 2019. Data centres are being erected at a pace that rivals the fibre optic boom of the late 1990s. Power grids are straining under the weight of facilities that didn't exist five years ago. And the question nobody seems to want to sit with is simple: what happens if the demand curve these buildings were designed for never actually arrives? Jim Rickards, the economist and former CIA advisor, is one of the few people asking that question out loud. His answer should give us pause.

The arms race logic

The AI infrastructure buildout follows a competitive logic that is almost impossible to opt out of. Every major cloud provider, from Microsoft to Meta to Amazon to Google, is racing to secure chips, lock down power capacity, and build the physical layer that will run the next generation of AI models. Morgan Stanley has warned that a transformative leap in AI capabilities is imminent, citing Elon Musk's claim that applying 10x the compute to large language model training effectively doubles a model's intelligence, and that the scaling laws supporting this are holding firm. The implication is clear: if you stop building, you fall behind. If you fall behind, you lose the platform war. So everyone builds. This is classic arms race dynamics. The decision to invest isn't based on a careful cost-benefit analysis of each individual data centre. It's based on the fear of what happens if your competitor builds and you don't. That logic can produce rational individual decisions that add up to a collectively irrational outcome.

The numbers behind the dream

The scale of spending is staggering. According to S&P Global, the four largest hyperscalers, Microsoft, Amazon, Alphabet, and Meta, are on track to spend around $635 billion on data centres, chips, and other AI infrastructure in 2026. Goldman Sachs expects AI-related capital expenditure to exceed $500 billion this year across the broader industry. Morgan Stanley has gone further, warning that AI capex is about to dwarf the dot-com-era telecom boom in both scale and duration. That's a comparison banks don't make lightly. Meanwhile, global AI spending is projected to reach $2 trillion by 2026, according to Gartner. The IEA projects that global data centre electricity consumption will double to around 945 terawatt-hours by 2030, with AI as the primary driver. In the United States alone, data centres consumed 183 terawatt-hours of electricity in 2024, roughly equivalent to the annual electricity demand of Pakistan. That figure is expected to more than double by 2030. These are not speculative startups burning through venture capital. These are the most profitable companies on earth, borrowing against future demand that may or may not materialise at the scale they're pricing in.

What Rickards is actually saying

Jim Rickards' argument isn't that AI is a fad. It's that the financial architecture supporting the buildout may be hiding risks the market isn't pricing in. In a series of presentations released in early 2026, Rickards laid out a case that the rush to build data centres, acquire advanced chips, and expand cloud capacity is creating conditions that deserve far more scrutiny. His core concern is contagion: AI infrastructure investment has become so deeply woven into construction, energy, manufacturing, and finance that problems in one part of the AI market could affect confidence across a much larger part of the economy. As he put it, the same forces driving excitement now could become the source of serious trouble ahead. The risk isn't confined to tech stocks. It's structural. This echoes a point made by Man Group in their February 2026 analysis: "The AI industry has been built for a demand curve that may not arrive and whose economics may deteriorate if it does." They noted that AI-related exposure is increasingly metastasising through the economy, and that this quiet migration of risk is the bubble's distinctive feature.

The dot-com parallel, and where it breaks down

The comparison to the dot-com era is impossible to avoid. In the late 1990s, telecom companies laid more than 80 million miles of fibre optic cable across the United States, driven by wildly inflated projections that internet traffic was doubling every 100 days. The actual growth rate was far more modest. The result was massive overcapacity, a wave of bankruptcies, and a bust that extended well beyond the telecom sector. The parallel is instructive, but it's not a perfect map. As KKR pointed out in their analysis, today's data centre cycle is fundamentally different in at least one important respect: it's underpinned by long-term contracts with the world's most financially robust technology companies, not speculative startups. The Richmond Federal Reserve noted that the 1990s telecom infrastructure eventually got used, just not on the timeline the original investors expected. That distinction matters. The question isn't whether AI infrastructure will ever be useful. It almost certainly will. The question is whether the economics work on the timeline the current wave of investment requires. As one analyst bluntly put it: "If the economics don't work, doing it at massive scale doesn't make the economics work any better, it just takes an industry crisis and makes it into a national economic crisis."

The energy problem nobody can hand-wave away

Every data centre needs power, and the AI buildout is consuming it at a rate that is already straining grids. Morgan Stanley's "Intelligence Factory" model projects a net U.S. power shortfall of 9 to 18 gigawatts through 2028, a 12% to 25% deficit in the power needed to run it all. Developers aren't waiting for the grid to catch up. They're converting Bitcoin mining operations, firing up natural gas turbines, and deploying fuel cells. Reuters reported that U.S. power consumption hit a second consecutive record high in 2025, and electricity prices nationwide have risen by an average of 7%. S&P Global forecasts that data centre grid-power demand will nearly triple by 2030, reaching 134.4 gigawatts. The IEA expects data centres to account for roughly 50% of U.S. power demand growth for the remainder of the decade. If AI demand meets projections, this is manageable. If it doesn't, we will have built some of the most expensive infrastructure in human history to solve a power problem that didn't need solving at this scale.

Jevons paradox, or its evil twin

There's a compelling counterargument to the overbuilding thesis, and it comes from a 160-year-old economic observation. Jevons paradox, named after the Victorian-era economist William Stanley Jevons, holds that when a resource becomes more efficient to use, total consumption of that resource tends to increase rather than decrease. Cheaper coal didn't mean less coal consumption. It meant factories, trains, and steamships found new uses for it. Applied to AI: as models become cheaper and more efficient to run, the number of applications, users, and workloads expands dramatically. DeepSeek demonstrated this in early 2025 when it showed that competitive AI models could be trained and run far more cheaply. Nvidia's stock briefly tanked on the assumption that less compute would be needed, but Microsoft CEO Satya Nadella responded by invoking Jevons paradox directly: "As AI gets more efficient and accessible, we will see its use skyrocket." Gartner projects data centre electricity demand will grow 16% in 2025 and double by 2030. Bloomberg NEF expects U.S. data centre power demand could reach 106 gigawatts by 2035. If Jevons holds, the buildout isn't just justified, it's barely enough. But Jevons paradox has an evil twin. If cheaper AI means margins collapse, then the infrastructure gets used but nobody can afford to maintain it. The demand exists, but the revenue doesn't follow. This is the scenario that should worry investors more than a simple lack of demand: a world where AI is everywhere, indispensable, and unprofitable.

The risk asymmetry

The interesting question here isn't whether the AI buildout is right or wrong. It's about the shape of the risk. If the optimists are right, the infrastructure gets used, the economics work, and the companies that built aggressively are rewarded. The returns are good, perhaps great. If the pessimists are right, hundreds of billions of dollars in infrastructure sits underutilised, the contagion Rickards warns about ripples through construction, energy, and finance, and the early investors absorb enormous losses. The infrastructure probably still gets repurposed over time, just as fibre optic cables eventually found their use, but the financial damage is done. The asymmetry is striking. The upside is strong returns. The downside is a potential economic crisis that extends well beyond the tech sector. And because the buildout is happening so fast, the feedback loop is compressed. We won't have decades to course-correct. Morgan Stanley's own timeline suggests the critical period is months, not years. Harvard economist research has highlighted another uncomfortable detail: the gap between the long-term vision of AI and whether growth will materialise fast enough to pay for the buildout. These companies can end up underwater even if AI grows fast, just less rapidly than they hope for. Add to that the unusual "circular financing" arrangements between customers and suppliers, where it appears that some companies are effectively paying their customers to buy their products, and the picture gets murkier.

What to watch

None of this means the buildout is doomed. It means the bet is larger, more concentrated, and more consequential than most people realise. A few things are worth tracking: Utilisation rates. Are the data centres being built actually getting filled? If utilisation remains high and growing, the buildout is justified. If vacancy rates climb, that's an early warning sign. Revenue versus capex. The gap between what AI companies are spending and what they're earning needs to close. One analysis estimated that to break even on 2025-2026 spending alone, the industry would need approximately $1 trillion in cumulative AI revenue, a figure that remains distant. Energy costs and availability. Power is the binding constraint. If energy costs continue rising or grid capacity can't keep pace, it will slow the buildout regardless of demand. The contagion question. Watch how deeply AI-related investment is embedded in sectors that have nothing to do with technology. The more entangled it gets, the more Rickards' contagion thesis matters. We're building data centres for a dream. The dream might be real. But the cost of being wrong isn't just money, it's the kind of structural economic disruption that touches industries and communities with no direct connection to artificial intelligence at all. That asymmetry, more than any single forecast or projection, is what makes this moment worth paying very close attention to.