The Trillion-Dollar Treadmill, Updated (August 2026) What the arithmetic still says — and why the markets are starting to look nervous
There is a number at the centre of the artificial-intelligence boom that almost no one likes to say out loud. Once you write it down, the whole thing starts to look strange. So let us write it down again, with the latest figures.
The spending
In 2026 the four largest American technology companies — Microsoft, Alphabet (Google), Amazon and Meta — are still guiding toward roughly $725–760 billion of combined capital expenditure in a single year. The overwhelming majority of it is artificial-intelligence infrastructure: graphics chips, custom silicon, data-centre buildings and the electricity to run them. That remains up roughly 75–80 percent from the previous year.
Looking further out, Goldman Sachs continues to expect those same four companies to spend a combined $5.3 trillion between 2025 and 2030. Broader global AI-related investment is now estimated at around $1 trillion this year alone.
What has changed since July is the scale of the commitments. Recent analysis of company filings shows that nine major tech firms (including the big four plus Oracle, Nvidia and others) now carry roughly $2.4–3 trillion in AI-related off-balance-sheet purchase commitments, leases and contractual obligations. That is several times their reported annual capital spending. These are long-term promises that must be paid whether or not the revenue appears on schedule.
Some of the cash-rich giants have already seen free cash flow turn negative or sharply lower in recent quarters as the spending outruns operating cash generation. They are also borrowing more: hyperscalers issued well over $150 billion in debt in the first part of the year, with full-year projections still rising.
The earning
Microsoft’s AI business was running at an annual rate of about $37 billion earlier in 2026 (growing more than 100 percent year-on-year at the time). OpenAI has been variously reported in the $25–41 billion range. Anthropic has grown dramatically and is now cited in the $47–65+ billion run-rate territory by mid-to-late summer.
Adding contributions from Google, Amazon, Meta, other model providers and the broader ecosystem, the total identifiable revenue the world currently pays for AI itself is still only a few hundred billion dollars a year at most — and much of that revenue is concentrated in just two private companies. Even optimistic tallies remain a fraction of the capital being spent.
Spent (or committed) this year: hundreds of billions, with multi-trillion obligations locked in. Earned from AI per year: still far less.
The industry is still spending, in a single year, several times what the entire world pays it for AI annually.
The asset that dies
A factory lasts decades. A railway lasts a century. AI hardware does not.
Secondary-market data and operator economics continue to show rapid value decay. Top-of-the-line accelerators lose a large share of their economic value within roughly three years as newer generations arrive. Accounting lives of five or six years remain common, which — as Michael Burry and others have noted — can understate the true economic cost.
The math, made simple
Cloud computing is a high-margin business, but not infinitely so. After electricity, cooling, networking, staff and the cost of capital, a strong operator might keep around 30 cents of net profit on every dollar of revenue. That is a generous assumption.
At a 30 percent net margin, recovering $1 of investment requires about $3.30 of revenue.
So the hundreds of billions spent in 2026 alone would still need well over $2 trillion of cumulative AI revenue to pay itself back. At current run rates that is many years of every dollar the world currently pays for AI — just to recover a single year of building.
And the hardware still dies in about three years.
That means the spending is not a one-time payment you recover at leisure. To stay competitive you must largely buy it all again every few years — and the bill is not merely repeating, it is growing, and it is now locked in by multi-year contracts.
Simply standing still — replacing what wears out — still costs on the order of hundreds of billions every year, forever, before any growth or return. Covering that replacement bill alone would require AI revenue many times higher than today, just to keep the lights on and break even on the treadmill.
This is the heart of it. The investment never fully converts into an owned, paid-off asset. It behaves less like a building and more like a permanent, rising cost of merely continuing to exist in the business.
So is it madness?
Not necessarily. The bull case remains real.
The correct comparison is still a semiconductor fab. Chipmakers have always lived on this kind of treadmill. It is survivable at vast scale with compounding demand and genuine pricing power. It bankrupts everyone else.
Two forces can still save the AI buildout:
- Each new chip generation does far more useful work per dollar.
- Old chips do not go straight to zero — they cascade into cheaper everyday workloads.
If demand and efficiency both compound fast enough, for long enough, the silicon becomes a consumable while the real durable assets (software, subscriptions, user lock-in, data) become valuable. That is still the bet: that performance-per-dollar and demand will keep compounding faster than the hardware decays.
Who pays if the bet is wrong — and what the markets are saying right now
The cash-engine giants can fund the treadmill out of existing profits more or less indefinitely, though with permanently thinner margins. The more fragile links are pure-play builders and anyone carrying heavy leverage or large fixed commitments. Borrowed money and long-term contracts do not care that the collateral is obsolete in three years.
In August 2026 the markets are starting to show the strain. Technical strategists at BTIG have pointed to an extreme “factor musical chairs” pattern: 57 trading days already this year in which price action and market breadth moved in opposite directions — matching the highest level in nearly three decades. Capital is rotating erratically across factors with little rhythm. Semiconductors have been capped below key moving averages. There has been no classic capitulation day so far this year, which itself is a warning that pressure is building rather than being released. The risk, they argue, is that when the music finally stops, investors are forced into cash and the first broad, high-correlation selloff in many months arrives.
This is exactly the kind of fragility the arithmetic predicts when a market is dominated by a handful of mega-cap AI names while the underlying spending treadmill continues to steepen.
The bottom line
Strip away the excitement and the math still says something simple and uncomfortable:
The world’s most valuable companies are spending (and committing) trillions on assets that expire in a few years, to chase a revenue stream that remains a fraction of the cost, on the faith that efficiency and demand will compound fast enough to outrun the decay — while using longer accounting lives and off-balance-sheet structures that make the present look calmer than it is.
It might still work. For two or three of them it probably will. But the treadmill is steeper than it looked in July, the obligations are stickier, and the market is beginning to dance a more nervous dance around it.
The one thing more dangerous than the cost of running remains being the first to step off.
A note on the numbers Hard figures (company guidance, Goldman estimates, Anthropic/OpenAI run-rate reports, the $2.4–3 trillion commitment analyses, BTIG’s 57-day divergence statistic) come from recent earnings releases, Wall Street Journal/Bloomberg-style reporting of filings, and market strategist notes as of mid-August 2026. The simple arithmetic (margin assumptions, payback multiples, replacement costs) is deliberately back-of-the-envelope and uses the same conservative 30 percent net-margin and three-year economic-life assumptions as before. Change the assumptions and the exact numbers move; the shape of the problem does not. Companies revise guidance frequently and do not always break out “AI revenue” cleanly.