There’s a familiar hum in markets right now: vast sums are being committed to chips, data centers and start-ups on the premise that artificial intelligence will remake entire industries. But beneath the swagger are financing structures—special‑purpose vehicles, GPU‑backed loans, and private‑credit deals—that look worryingly like the scaffolding of past financial manias.
An industry built on borrowed chips
Look at CoreWeave. Once a crypto miner, it became a bellwether for today’s AI financing: a blockbuster IPO, a string of multibillion‑dollar commercial deals with the likes of OpenAI, Meta and Nvidia, and an income statement that reads like a growth playbook on steroids. The company projects roughly $5 billion of revenue this year against about $20 billion of spending, and it has taken on roughly $14 billion of debt to bridge the gap. A single customer—Microsoft—may provide as much as 70% of its revenue. In other words, the business model depends on a tiny number of big counterparties and an assumption that future demand will make the current heavy lifting pay off.
That pattern—build now, hope revenue catches up later—plays out across hyperscalers and cloud operators. The hardware needed to train and run advanced models is insanely expensive; estimates of industry spending this year run into the hundreds of billions. Where the cash comes from matters. When companies lean on equity, losses fall to shareholders. When they pile on debt, lenders and the broader financial system can get pulled into a correction.
Old tricks, new scale
A worrying feature is how familiar the financial engineering looks. Lenders and sponsors are resurrecting tools once associated with previous crises. Special‑purpose vehicles let firms keep debt off corporate balance sheets; asset‑backed securitisations of data‑center leases are being marketed to investors; and vendors are lending against the chips themselves. If older GPU models rapidly lose value as newer ones emerge, that collateral can evaporate fast—prompting margin calls, firesales of hardware and a vicious downdraft in prices.
Those aren’t hypothetical pathways. Private credit has swelled: non‑bank lenders now hold large slices of the loans that fund this build‑out. Unlike regulated banks, many of these funds disclose far less about exposures, and their investors—pension funds, insurers and endowments—are often the ultimate backstop. That makes the web of risk harder to map and potentially more contagious.
Circular deals and the optics of growth
Another awkward trend is circularity. Big chipmakers and cloud providers have struck deals that blur the lines between investment and revenue. Chip suppliers or cloud hosts sometimes invest in or extend favorable financing to customers that then buy their products—an arrangement that mechanically inflates demand. Formal accounting may be correct, but the arrangement invites reasonable skepticism about how much growth is genuine and how much is self‑reinforcing choreography.
Nvidia, the dominant supplier of AI accelerators, illustrates both the scale and the complexity: equity investments, supplier relationships and enormous chip sales are entwined in ways that make it hard to disentangle who is really taking the economic risk.
Two kinds of bubbles
Not all speculative surges are identical. One useful lens distinguishes mean‑reversion manias—financial fads that leave little lasting technical progress—from “inflection” bubbles where oversized investment funds infrastructure that materially changes an economy. Railroads and the internet fit the latter category: both saw speculative excess, but the underlying technology produced durable change.
If AI is an inflection event, the overinvestment will sting many investors while leaving the world richer and more productive. If it is primarily financial sleight‑of‑hand—assets overbought and revenue over‑promised—the fallout could be concentrated and painful. The central uncertainty is timing and distribution of returns: who captures the profits, how soon, and whether debt holders will be repaid if enthusiasm cools.
What would make a crack look like 2008—or 2000?
Three flash points worry investors and regulators:
- A glut of computing capacity that forces rental rates down and pushes marginal providers into insolvency.
- Rapid obsolescence of collateral (older GPUs losing value when new architectures arrive), triggering lender losses and fire sales.
- A private‑credit unwind where opaque lenders suffer losses that ripple into banks, insurers and public markets.
None of these outcomes is inevitable. But the scale is new: some forecasts suggest debt tied to AI infrastructure could reach into the trillions within a few years if build‑out continues at pace. That amplifies stakes.
How investors and companies are responding
Prudent actors are raising long‑dated financing, choosing geographies with less competition for power, and seeking diversified customers rather than single big tenants. Others are doubling down: bigger seed rounds, aggressive valuations and vendor financing that keeps the arms race humming.
For those allocating capital, the dilemma is classic: participate and risk loss if the mania pops, or sit out and risk missing the winners. The pragmatic answer many asset managers land on is selectivity—exposure to high‑quality operators with durable cash flows and careful scrutiny of the debt being underwritten.
Why this is different (and why that doesn’t make it safe)
AI isn’t just another gadget: it can change labor markets, customer interactions and productivity in ways that are hard to reverse. The technology’s pace also compresses investment horizons—what looks like overbuild today might be a necessary step to reach next‑generation capabilities tomorrow. Still, the combination of unparalleled technical uncertainty and large scale financing is a novel cocktail. It deserves both excitement and caution.
If you want a quick snapshot of how companies are expanding capacity and experimenting with new deployment models, look at efforts from cloud and hyperscale players—some are exploring radical placement ideas such as orbital data centers to relieve terrestrial constraints, a sign of how inventive the infrastructure race has become (Google’s Project Suncatcher moves the idea of data centers into space). Meanwhile, the race to ship new models and capabilities continues apace—Microsoft, for example, is rolling out new in‑house image models to complement its AI stack (Microsoft Unveils MAI‑Image‑1).
AI’s technical frontier remains contested. Experts still debate whether current systems are approaching human‑level generality, even as adoption accelerates across business functions and consumer apps (AI’s tipping point fuels both optimism and skepticism).
The story unfolding now will be written by technologists, dealmakers, regulators and the market’s appetite for risk. Some infrastructure will be redundant, some start‑ups will flame out, and a handful of winners may justify every dollar spent. Expect turbulence. Expect breakthroughs. And expect that the eventual accounting—who won, who lost and why—will be complicated and instructive in equal measure.