The AI Infrastructure Arms Race: Tomorrow's Economy, or Yesterday's Bubble?

Everyone is asking the wrong question about AI.

Since ChatGPT launched in late 2022, the debate has fixated on the software: the chatbots, the assistants, the endless demos. But the real story isn't on your screen. It's in the ground, in the steel, in the substations. It's the infrastructure race happening beneath the headlines, and it may be the most consequential capital allocation decision of our generation.

Three Booms, Not One

Strip away the hype and the "AI boom" splits into three distinct waves.

There's the application boom: businesses bolting AI onto everything from customer service to spreadsheets. There's the model boom: OpenAI, Anthropic, Google DeepMind and xAI racing to out-build each other's frontier models. And then there's the one nobody's talking about enough, the infrastructure boom.

This one is different. It's physical. GPU clusters. Data centres. Cooling systems. Substations. Transmission lines. These aren't lines of code you can rewrite overnight. They're concrete and copper, poured and wired over years, at staggering cost. Once built, they don't come back. They're sunk.

And that changes everything.

Every Firm, Building at Once

The numbers make the scale hard to grasp. Microsoft, Alphabet, Amazon and Meta, the four largest US cloud providers, are together tracking toward roughly $760 billion of capital expenditure in 2026, against $413 billion in 2025.¹ In February the same tally was put at $660 to $690 billion across five companies including Oracle.² The estimate has risen every quarter this year, which is itself the point. Break it down by company and it gets more staggering still. Amazon raised its 2026 guidance to $220 billion on 30 July, up from $200 billion. Alphabet lifted the top of its range to $205 billion, and Meta took its own ceiling to $145 billion.³ Microsoft expects about $175 billion, revised down from $190 billion on 29 July when an accounting change shifted future data centre leases off the capital expenditure line.⁴

One detail in those revisions should give anyone pause. Amazon attributed its entire $20 billion increase to the rising cost of memory. Microsoft says roughly $25 billion of its own outlay buys no additional capacity at all, covering higher memory and storage prices.⁴ Between two companies, some $45 billion of this year's spending purchases nothing that was not already planned.

The rebuttal comes from the biggest spender. Andy Jassy told investors that even at $220 billion Amazon will not have enough capacity to meet its 2026 demand, that the same will hold in 2027, and that the demand already booked for 2028 is striking.⁵ If he is right, the arms race is not overbuilding at all. It is running behind. Add Oracle's $50 billion and OpenAI's $500 billion Stargate project with SoftBank, and the picture is unmistakable.²

Guided capital expenditure, 2026

Every transformative technology has demanded physical infrastructure. Railways needed track. Electrification needed grids. The internet needed fibre. AI is no different, except for one thing: the entire industry is building simultaneously, at a pace and scale with no real precedent.

Each decision, taken alone, looks rational. If AI defines the next fifty years, falling short on compute means falling out of the race entirely. In a market where first-mover advantage can compound for decades, underinvestment is the scarier bet.

But here's the problem. When every firm reaches that same conclusion, at the same time, for the same reason, it stops being competition. It becomes an arms race.

Rational Firms, Irrational Industry

Arms races don't happen because anyone wants inefficiency. They happen because no one can afford to stand still while rivals keep building.

That's exactly the dynamic playing out in AI infrastructure. Every dollar one firm spends ratchets up the pressure on its rivals to spend more. Companies aren't just investing to meet today's demand. They're investing out of fear of being the one firm left behind.

And there are real questions about whether the revenue is there to justify it. Sequoia's David Cahn has calculated that AI infrastructure spending from just 2023 and 2024 would need something like $800 billion in AI product revenue over the life of the hardware to earn an acceptable return, with most of that hardware useful for only three to five years before it's obsolete.⁶ Bain & Company puts the bar even higher, estimating the industry will need $2 trillion in annual AI revenue by 2030 to justify the spending trajectory now underway, more than the combined 2024 revenue of Amazon, Apple, Alphabet, Microsoft, Meta and Nvidia put together.⁶ Today's AI revenues remain a small fraction of that.

The result is rational actors producing an irrational outcome. If five competitors each decide they need a million GPUs to secure the future, society ends up financing five million, when, in the medium term, a fraction of that may actually be needed. Every individual bet is defensible. Collectively, they may add up to enormous overcapacity in waiting.

Sunk Costs, Locked Decisions

Infrastructure makes this worse, not better. Data centres can't easily be repurposed. AI chips depreciate fast as new generations arrive. Power grids need maintenance whether or not the demand shows up.

Microsoft's own disclosures show how skewed the spending is. Roughly two thirds of its quarterly capital expenditure goes into short-lived assets, principally GPUs and CPUs, with the remainder in buildings and land it expects to monetise over fifteen years and beyond. In the same quarter it extended the assumed useful life of those buildings from fifteen years to twenty five, an accounting choice that spreads the cost further into the future without changing anything physical.⁴

So once the money is spent, it can't be unspent, and that warps the incentives further. Pulling back once excess capacity becomes obvious would mean admitting the earlier spending might never pay off. It's easier to keep building. Yesterday's investment becomes the justification for tomorrow's.

Who Wins, Who's Exposed

Some winners are obvious: chipmakers, electrical equipment suppliers, utilities, construction firms, data centre developers. Entire supply chains are riding this wave of capital expenditure.

The risk sits elsewhere: with the investors financing assets that may never earn their keep, the operators running data centres below capacity while costs keep piling up, and the capital that could have gone anywhere else in the economy but instead got funnelled into one story.

The Technology Can Succeed. The Bet Can Still Fail.

Here's the part that gets lost. AI doesn't need to fail for this to go wrong. History offers two sobering precedents.

The Railway Mania of the 1840s. At the peak of the mania in 1846, Parliament authorised £132.6 million in new railway capital in a single year, covering 9,500 miles of new track. By 1847, railway investment reached roughly 7% of Britain's GDP.⁷ Railway share prices subsequently collapsed by roughly two-thirds between 1845 and 1850, ruining thousands of middle-class households who had invested their savings.⁷ The railways themselves were a triumph, transforming the British economy for a century to come. The people who financed them, largely, were not.

The dot-com fibre glut. In the five years after the Telecommunications Act of 1996, US telecoms poured over $500 billion, mostly debt-financed, into fibre-optic cable, switches and wireless networks, betting that internet bandwidth would triple every few months. It didn't: by various estimates, 85 to 95% of the fibre laid in the '90s remained unused, or "dark," years after the bubble burst.⁸ Companies like Global Crossing and WorldCom went bankrupt owing billions.

The pattern in both cases is identical. The technology won, the infrastructure eventually found its use, but the earliest financiers of that infrastructure were often ruined getting there.

AI could easily follow the same script. If adoption accelerates fast enough, today's infrastructure spend looks visionary in hindsight. If it doesn't, we may find we've financed tomorrow's economy decades before tomorrow actually shows up.

That's not technological failure. That's misallocation, and it's exactly what a bubble looks like. Not a worthless technology, but expectations about speed and payoff drifting away from reality.

The Real Risk

The danger isn't that AI disappoints us. It's that competition itself is pushing the industry to build faster than demand can justify.

An arms race rewards whoever moves first. Markets reward whoever allocates capital best. Those two goals don't always point the same way, and right now, with the industry pushing toward three-quarters of a trillion dollars in spending this year alone, they're pulling further apart than at almost any point in economic history.

AI may well reshape the global economy over the coming decades. But if every firm keeps building simply because every other firm is building, we may look back and realise the defining bubble of this era was never the software.

It was the race to build what sits beneath it.

Footnotes

1. Statista, Big Tech's AI Spending to Reach $760 Billion in 2026 (opens in a new tab), July 31, 2026.

2. Futurum Group, AI Capex 2026: The $690B Infrastructure Sprint (opens in a new tab), February 12, 2026. ²

3. CNBC, Amazon, Meta and Microsoft face skeptical investors this week after Google report sparked sell-off (opens in a new tab), July 28, 2026.

4. Microsoft, Fiscal Year 2026 Fourth Quarter Earnings Conference Call (opens in a new tab), July 29, 2026. ^{2 3}

5. CNBC, Amazon (AMZN) Q2 earnings report 2026 (opens in a new tab), July 30, 2026.

6. IEEE ComSoc Technology Blog, Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom and bust (opens in a new tab), September 27, 2025. ²

7. Market Histories, The Railway Mania: Britain's Victorian Tech Bubble (1840s) (opens in a new tab), February 14, 2026. ²

8. Forbes, Is The AI Boom Headed For Its "Dark Fiber" Moment? (opens in a new tab), March 24, 2025.