The AI Bubble's Physical Bill: Why Data Centre Construction Is the Real Economic Risk
AI valuations are being set by the quality of the story, not the cash flow, and that would be tolerable if the story had stayed in the stock market. It hasn't: it's been poured into concrete, and concrete doesn't reprice on a Friday afternoon.
Most of what you need to know about AI capital markets sits in two numbers. In October 2025, an employee share sale valued OpenAI at $500 billion, the richest mark ever put on a private company. Three months earlier, a tender offer reportedly pushed SpaceX towards a $400 billion valuation. Neither number matters for what it says about rockets or chatbots. Both matter for what they reveal about the story now financing AI data centre construction, the economic risk being poured into the ground one substation at a time.
Treat these figures less as measurements than as bids in a narrative auction: prices set by the quality of a story rather than by any cash flow an outsider can audit. Call the gap between story and cash flow the narrative premium. In a share price a narrative premium is tolerable, because equity can reprice on a Friday afternoon.
The trouble with this cycle is where the premium has been parked: concrete, steel, substations and fibre, assets that cannot reprice, only default.
The pour is no longer a rounding error. McKinsey estimates the world will need to spend $6.7 trillion on data centre capacity by 2030, $5.2 trillion of it for AI workloads. Microsoft said in January it was on track to invest approximately $80 billion on AI-enabled data centres in a single fiscal year. And the debt has arrived at single-project scale: Meta's Hyperion campus in Louisiana is being financed with a $29 billion package of debt and equity arranged by PIMCO and Blue Owl, structured so the bulk of the borrowing sits off Meta's own balance sheet. When the financing has to be engineered off the balance sheet, somebody has already run the scenario where the tenants don't show.
The fair objection is that the market has rarely looked tighter. CBRE put vacancy across North America's primary data centre markets at a record-low 1.9 per cent at the end of 2024, with the bulk of a record construction pipeline already pre-let. That is the bulls' best evidence, and it deserves a harder look at who the pre-lets are: the same handful of platforms and AI labs whose capital spending is itself financed by the narrative premium. 'Pre-let' increasingly means let to one of five balance sheets making the same assumption at the same time, and vacancy measures the market that exists, not the 2027–28 cohort now being financed. The fibre builders of 2000 had reassuring order books too, right up to the quarter they didn't.
Is the AI bubble real, or is the demand real too?
Both, and the loudest voices on each side keep missing it. The bears say it is all story; the bulls say the demand is genuine. Genuine demand and a valuation bubble coexist comfortably, and history is crowded with the pairing. Britain's railway mania of the 1840s financed track that passengers used for a century and still bankrupted most of the people who paid for it. The fibre dug into the ground in 1999–2001 carried the internet boom that followed, though the firms that paid for the trenches mostly didn't survive to bill for it.
Usefulness is no defence against overbuilding; the underwriting question is whether revenue arrives before the debt does.
On the revenue side, AI demand is real but moves at the speed of organisations, which is to say slowly. Enterprise adoption is a procurement cycle, a change-management programme and a governance argument, the slow craft that practitioners like Abyshire are hired to make routine rather than an overnight transformation. Usage grows with deployment, deployment grows with budget cycles, and meanwhile the build-out runs at the speed of capital markets. The gap between those two clocks is where bubbles live, and right now the capex clock is winning by years.
The bulls' strongest structural point is elasticity: every collapse in the unit cost of compute so far has produced more consumption, not less. Elasticity rescues the renter first, though. The owner who financed the building at 2025 rents only wins if revenue lands inside the debt's window; being early and being wrong have identical cash flows.
Why data centre construction falls harder than code
When a software stock de-rates, the loss is real but the asset is weightless: no creditors' committee, no maintenance bill, no planning consent quietly expiring. A data centre is the opposite. It is a 15-to-20-year financing secured against a building tuned for a narrow class of tenant, in a fixed grid-queue position, with power contracts attached. If the tenants fail to appear at the assumed rents, the asset doesn't drift down gently; it strands, and the loss lands on whoever holds the debt. Increasingly that means infrastructure funds and regional lenders, not venture capitalists who priced in the chance of zero.
A campus is also two assets wearing one valuation. The durable half is the grid position: consented land with contracted power in a market where new connections queue for years. The wasting half is everything above the slab, fitted out for a chip generation that replaces itself every couple of years at ever-rising rack densities. In a downturn those halves will be priced separately: the owner holding power has something to sell, the owner holding yesterday's fit-out has a warehouse with remarkable cabling. Underwriting that prices the site as one number will get the split wrong.
The physical inputs carry their own momentum. The IEA calculates that data centres consumed about 1.5 per cent of the world's electricity in 2024 and projects that more than doubling to roughly 945 TWh by 2030, slightly more than Japan consumes in total today. Goldman Sachs expects AI alone to drive a 160 per cent increase in data centre power demand over the same period. Turbines, transformers and grid connections all run on multi-year lead times, so the capacity being financed this year is a bet on revenue in 2029, not 2026.
The second-order effects are the part nobody is pricing. Construction firms that re-tooled for multi-year campus pipelines. Specialist electrical and logistics contractors booked out for seasons ahead. Energy suppliers that signed long power-purchase commitments against campuses still on the drawing board. Regional economies, some of them far from wealthy, that rezoned land, upgraded substations and pre-spent the business rates.
An equity correction trims portfolios and ruins a quarter. A pipeline correction removes payrolls, and a payroll takes the better part of a decade to rebuild in the postcode that loses it.
The late-cycle tells are there if you read for pattern rather than press release: companies with no adjacency to compute announcing AI infrastructure strategies, model access priced to flatter adoption with the serving costs left unexamined, capacity announced before tenants are signed. When the marginal buyer of a capital-heavy asset is motivated by fear of missing a theme rather than by contracted revenue, the base rate for that asset class over five years is poor. That is a forecast about capital discipline, or its absence, and it holds even if AI turns out to be everything its believers claim.
What would change my mind
I hold this view at six in ten, not nine, and the evidence that would flip it is specific. If, by late 2027, inference revenue on new capacity covers depreciation plus a normal cost of capital at the major platforms, the campuses fill and the narrative premium turns out to have been a down payment. Renewal behaviour would deliver the same verdict earlier: if firms that piloted AI tools in 2024–25 expand seats and budgets at renewal rather than quietly trimming, the revenue clock is running faster than I've assumed. I'm watching both. Anyone with exposure to the build should be too.
The practical response depends on which side of the trade you sit. If you buy AI rather than sell it, none of this argues for waiting; it argues for discipline about readiness before you build, and for letting someone else's shareholders fund overcapacity you will later rent at a discount, which is exactly what the smart money did with dark fibre in 2002. If you invest in, lend to or supply the build-out, the questions are blunter: who is the contracted tenant, for how long, and what does the asset earn if workloads grow at half the deck's assumption? A sound technical strategy stress-tests the downside case before the concrete does it for you.
The asymmetry is the whole story. If the bulls are right and demand arrives on schedule, tech investors make money and the buildings fill: a good outcome, widely shared. If they're wrong, the equity loss is the small part of the bill, and the large part lands on order books, regional lenders and towns that bet their rezoning on a story written in a valuation deck. Markets can forgive a narrative premium in a share price. Concrete is less sentimental.
Questions people ask
Will the AI bubble burst in 2026?
Timing is the one thing base rates can't give you, so the honest answer is conditional. A broad crash needs a revenue disappointment at the cash-rich platforms, while a correction in debt-financed, speculatively built data centre capacity only needs tenants to arrive late. I'd put better-than-even odds on the second within three years and treat the first as a lower-probability, higher-impact event. Watch enterprise renewal rates and inference revenue per unit of new capacity; those turn before the headlines do.
What happens to data centres if AI demand falls short?
They don't disappear; they strand. Capacity that can't be filled at the assumed rents gets re-let cheaper, repurposed for lower-value workloads or sold at distressed prices, and the loss concentrates in whoever holds the debt rather than in the share prices of the big platforms. The fibre build-out of 1999–2001 is the template: the assets were eventually used, but the original owners and their lenders took the write-downs first.
How is the AI build-out different from the dot-com bubble?
The equity story is similar but the capital structure isn't. Dot-com losses sat mostly in shares held by people who knew they could go to zero. This cycle has layered heavy debt on top of physical assets, plus grid connections and long energy contracts, so a correction travels into construction, utilities and regional credit rather than stopping at portfolios. That plumbing is why the physical bill matters more than the ticker.
Related
- How AI Makes Bad Bosses Worse: The Agreeability Machine in the Corner Office
- The AI Efficiency Paradox: Why Cheaper AI Breaks the Business Model
- Digital Business
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.