The AI Bubble in 2026: What the Infrastructure Binge Actually Costs

01 The Biggest Capex Number Ever Assembled

Somewhere between the launch of ChatGPT in November 2022 and the end of 2026, the largest technology companies on Earth decided to become, first and primarily, construction companies. Microsoft, Alphabet, Amazon and Meta — the four American hyperscalers — now spend more on capital expenditure each year than most countries spend on everything. Combined capex, which hovered in the low hundreds of billions of dollars across the group in the early 2020s, has run at roughly six hundred billion dollars annualized across the group in 2025-2026, and the majority of it is data centers, power infrastructure and AI accelerators.

Hank Green's "The Riskiest Moment of the AI Bubble" — viewed about 2.1 million times — is one of the more effective mainstream framings of why this number is different in kind from ordinary tech investment. The core observation: a large fraction of this spending is not funded by revenue from AI products, because the revenue does not yet exist at anything close to this scale. It is funded by expectations of that revenue, priced into the most valuable equity in the world and lent against by everyone from chipmakers to sovereign funds.

This is what a boom looks like from the inside: real money, real buildings, real gigawatts, and a forward-looking revenue story whose present-day proof is thin. The most recent boom happened in the 2020s before seeing increased acceleration and media coverage — a period sometimes referred to as an AI spring, in deliberate contrast to the AI winters that preceded it. Springs and winters are meteorology; capital expenditure is accounting. The accounting is what this analysis tracks.

02 What the Money Actually Buys

Strip out the abstraction and the spend has a concrete shape. Graphics processing units — GPUs — cost roughly thirty to forty thousand dollars each at the flagship end, and a frontier training cluster can hold hundreds of thousands of them. Every accelerator needs a server around it, power delivery, cooling, networking and a building, which multiplies the chip cost several-fold before a single model is trained. One widely used planning figure across the industry is about ten to fifteen megawatts of data-center capacity, and roughly one to two hundred million dollars of capital, per fifty thousand high-end accelerators — and the Stargate consortium announced by OpenAI, SoftBank, Oracle and MGX in January 2025 talks about five hundred billion dollars of investment over four years, at an announced scale of gigawatts per campus.

The scarce input is increasingly electricity, not silicon. A single hyperscale AI campus is planned at hundreds of megawatts to multiple gigawatts, comparable to the load of a mid-sized city. American utilities have re-run their demand forecasts upward specifically because of data-center interconnection queues; the constraint story of 2026 is transformer lead times, grid interconnection waits and gas-turbine backlogs, not ideas.

None of this is speculative froth in the 1999 sense of pets.com banners. It is heavy industrial infrastructure with multi-decade depreciation schedules, financed against cash flows that — outside the hyperscalers' core cloud businesses — remain largely hypothetical. That combination is what makes this moment genuinely risky rather than merely dramatic: the assets are real, and so is the debt.

03 The Capex Curve

The clearest way to see the boom is the trajectory of combined capital expenditure at the four hyperscalers, as widely reported in their quarterly earnings. The figures below are reported annual totals — Microsoft on a June fiscal year, the others on calendar years — assembled from public earnings statements; treat them as reported estimates rounded to the nearest five billion dollars.

The slope after 2024 is the story. It took the group roughly a decade to build capex to about eighty billion dollars a year; the AI era added roughly half a trillion dollars of annual spend in about two. Very few categories of private investment in history have compounded at anything like this pace — which is precisely why the comparison that keeps surfacing is not to a product cycle but to the railway manias and the dot-com build-out.

04 The Circular Financing Problem

The most specific criticism of the boom is not the size of the spending but the shape of the money. In September 2025, Nvidia announced it would invest up to one hundred billion dollars in OpenAI, on the condition that OpenAI use the funds to buy Nvidia systems. OpenAI, in turn, committed to spending about three hundred billion dollars on Nvidia chips over an extended horizon. In the same season, OpenAI agreed to purchase one-point-seven-five gigawatts of capacity from a former Intel site being developed by Nvidia partner Amkor and SoftBank's Stargate venture, while Oracle signed cloud contracts reportedly worth around three hundred billion dollars to run OpenAI's workloads, financed in part by debt raised against the expectation of those same payments. The chipmaker funds the customer; the customer buys from the chipmaker; the cloud vendor borrows against the customer's promise to rent.

The last company to do something structurally similar was Cisco in 2000 — vendor financing to telecom customers who then bought Cisco equipment with it, a fact that Cisco executives of the era have themselves cited as a cautionary tale. Nvidia and OpenAI argue that this is different: the commitments are real capital against a real compute shortage, and demand for frontier training capacity continues to exceed supply. That argument is not absurd. It is also exactly the argument every vendor-financed boom makes, and the difference between a virtuous circle and a circular one is invisible until revenue from end customers — people outside the circle paying for AI — grows to cover the interest.

The honest summary of the circularity debate: it is not yet a Ponzi structure, but it is a closed loop whose stability depends on a variable, external AI revenue, that no participant can currently measure. When the key assumption of a financial structure is unmeasurable, risk does not disappear — it moves to whoever is last in the chain.

05 Inference Costs and the Revenue Gap

The revenue question is sharper than it looks, because the cost structure of frontier AI got worse, not better, as models improved. Chain-of-thought reasoning models — the class that arrived in late 2024 and dominates frontier use in 2026 — generate extensive internal reasoning text before answering, which multiplies the compute per query by an order of magnitude or more relative to a single-pass model. OpenAI has publicly described losing money on some heavy chatbot subscriptions, and industry-wide pricing has repeatedly reset downward even as per-user compute has risen. Adoption, meanwhile, is real but uneven: OpenAI's annualized revenue, reportedly on the order of twenty billion dollars in late 2025, is genuinely enormous for a company founded a decade earlier — and is roughly three percent of the annualized infrastructure run-rate of its four biggest backers.

Enterprise adoption tells a similar double story. Surveys through 2025-2026 repeatedly find that most large organizations have deployed AI somewhere, and only a small minority — often a quarter or less, depending on the survey — report measurable, material returns at the business-unit level. The pattern that recurs: pilots are cheap and impressive; production is expensive, organizationally hard and slow to compound. The capability frontier is running well ahead of the adoption frontier, and the gap between the two is exactly where the capital is sitting.

None of this means the revenue cannot arrive. Electricity, search advertising and mobile data all took a decade or more from infrastructure binge to dependable margin. But it means that the current spend is a wager on a medium-term adoption curve, and the entities making the wager are levering their balance sheets to a degree that assumes the curve's shape rather than tests it.

06 Rails, Fiber and the Shape of a Break

Every infrastructure mania gets compared to the last one, so it is worth doing the comparison with numbers rather than vibes. British railway investment peaked in the 1840s at several percent of annual GDP — an extraordinary figure — before the 1847 panic wiped out marginal investors while leaving a rail network that industrialized the country. The dot-com era drove US telecom fiber build-out; about one hundred billion dollars of long-haul network investment in 1996-2001, much of it debt-financed, ended with WorldCom's bankruptcy, Global Crossing's collapse and roughly ninety percent of the laid fiber sitting dark — unused for years until internet traffic finally caught up. Both episodes left durable assets and destroyed the people who financed them at the wrong point on the curve.

By GDP share, AI build-out sits between the telecom build and the railway mania — larger relative to the economy than fiber, far smaller than the 1840s. The more instructive comparison is structural. Railways and fiber shared a feature that AI data centers partly lack: the asset was the product. A finished railway sold tickets; a finished fiber route sold bandwidth. A finished AI data center sells intelligence whose unit economics still depend on a model, an electricity price and an adoption curve that all remain in motion.

So what would an actual break look like? Not a single dramatic crash, most likely, but the sequence the debt markets fear: AI revenue plateaus for two or three quarters, one leveraged middleman — a neocloud provider, a financed Stargate vehicle, a smaller AI lab on multi-year compute contracts — misses a payment, credit spreads on AI-linked debt widen, and the whole chain from OpenAI's compute commitments down to gas-turbine orders reprices at once. The hyperscalers themselves would not go bankrupt; their core businesses fund most of the spend. What breaks is the marginal operator and the marginal chip order — which is precisely the exposure Nvidia's own guidance, and everything priced off it, carries.

07 The Case That This Time Is Different

The bubble framing has a serious counterargument, and it deserves a fair hearing. Unlike pets.com banners, the assets being built are general-purpose industrial infrastructure with long useful lives: buildings, substations, cooling plants and accelerators that can serve any workload, not only frontier training. Hyperscalers are funding the majority of capex from operating cash flow, not debt — Microsoft, Alphabet, Amazon and Meta each generate a hundred billion dollars or more in annual operating cash — which is a categorically different footing from the debt-financed telecom carriers of 1999. And the underlying demand signal, inference at scale, is growing rather than shrinking; every reasoning model shipped increases the compute consumed per user, which is the opposite of the bandwidth-glut dynamic that sank fiber.

There is also a genuine shortage argument. Frontier training clusters are capacity-constrained through this decade on the planning horizons utilities and chip fabrication actually work on, and every major participant behaves as though compute is the binding constraint on capability. When a resource is scarce and the entities buying it have cash-flow cover, the spend can persist far longer than bubble metaphors predict.

The steelman, in one sentence: this may be less a bubble than an industrial mobilization — a rare case where the market is building ahead of demand on purpose, accepting waste as the cost of not being late. Railway investors in 1845 would have said exactly the same thing, and been both right about the rails and ruined by the timing.

08 What to Watch

Boom-watchers do not need to predict; they need to monitor the right gauges. Four matter most. First, the ratio of external AI revenue — revenue from outside the vendor-investor circle — to annualized AI infrastructure spend; today it is small and the direction of travel is more informative than the level. Second, credit spreads on AI-linked debt: a sustained widening there precedes any equity reckoning by quarters. Third, Nvidia's disclosed customer concentration and receivables growth, the cleanest public window into whether the circle is closing. Fourth, utility interconnection queues and power-price signals in the regions hosting the big campuses, because the electricity constraint binds before the capital constraint does.

The riskiest moment of any boom is not the peak of enthusiasm; it is the plateau that arrives when the money has already been spent, the capacity is already coming online and the revenue is still two years out. The generative AI technologies at the center of this cycle — large language models generating text and code, image models, video models and world models — may well justify every dollar being committed. The historical base rate says the infrastructure usually ends up justified and the financiers usually do not, and the two outcomes have never yet been reconciled by anyone's quarter.