Chapter 1 · Does the AI Bubble Exist?
The Setup: A $700B/Year Bet on an Imaginary Future Customer
Before judging whether the AI trade is a bubble, understand what is actually being built, who is paying for it, and where the money quietly leaks off the balance sheets.
Big Tech will spend more than $700 billion building AI infrastructure in 2026 alone — roughly two and a half cents of every dollar the United States economy produces, guided to $720–745 billion across Microsoft, Amazon, Alphabet and Meta before Oracle's fiscal-year spend pushes the total toward $835 billion. Nvidia's CEO calls it 'the largest infrastructure buildout in human history.' The number has become so large it no longer registers emotionally — which is precisely why the mechanics underneath deserve a hard look.
The cast of characters matters. At the top sit four hyperscalers with historically strong balance sheets — genuinely profitable, genuinely cash-generative companies. Beneath them sit the AI labs: OpenAI (~$40B annualized revenue) and Anthropic ($65B run-rate as of July 2026) — growing faster than any software companies in history, yet collectively still losing money against their obligations. Beneath THEM sits the financing layer nobody had heard of in 2023: neoclouds (CoreWeave et al.) renting GPUs with debt; SPVs like Meta's $27.3B 'Beignet' vehicle that keep liabilities off the books; and vendor financing from Nvidia, which has committed up to $100B into OpenAI tied to chip purchases and helped assemble a $500B financing vehicle with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.
This structure — supplier finances customer, customer's spending books as supplier revenue, debt hides in special-purpose vehicles between them — is not new. It is Cisco 1999 and it is the shadow-banking system of 2007, reassembled with better logos. That does not make it fraudulent or doomed. It makes it cyclical: this is how capital-heavy booms are always financed at the top, and how they always end.
The Gap: AI Infrastructure Spend vs AI Software Revenue
Big-four hyperscaler capex guidance vs combined annualized revenue of the AI labs + model APIs. The gap is the bet: spend assumes revenue arrives later, at utility scale.
How the money actually moves: hyperscalers sign multi-year lease and purchase commitments (these land in footnotes, not balance sheets); data centers get built by joint ventures funded by private credit (Apollo, Blackstone, Blue Owl) that insurers and pension funds then buy as 'infrastructure bonds'; Nvidia sells the chips — increasingly financed by Nvidia itself; the labs sign $100B+ compute commitments backed by expected future revenue from customers who don't exist yet at scale. Nikkei's investigation tallied $1.65 trillion of such off-balance-sheet obligations across five companies — more than their combined reported debt — with roughly $900 billion signed in a single quarter in mid-2026.
None of this is hidden from regulators — the BIS and the FSB have both flagged circular AI financing as a systemic watch item in 2026. It is hidden from most investors reading headline P/E ratios, because the risk lives in the credit layer: bond covenants, residual-value guarantees, insurer portfolios. When analysts say 'this time the bubble is in private credit,' this is the plumbing they mean.