Testing the AI Infrastructure Bubble Thesis

The primary filings now support a narrower answer than either “normal investment” or “bubble.” AI infrastructure spending and long-dated commitments are exceptionally large, and several suppliers now finance, guarantee, invest in, or purchase capacity from companies that also buy their products. Demand is also expanding quickly, with major cloud businesses reporting rapid growth and Microsoft still describing Azure demand as above available capacity.

As of September 3, 2026, the evidence points to a fragile capital expansion, not a demonstrated economy-wide glut or credit crisis. The financing links are real. Whether they become self-reinforcing losses still depends on independent end demand, utilization, contract recourse, and the resale value of the assets.

What the latest disclosures establish

Question	Primary-source observation	What it means
Is demand already collapsing?	Azure grew 43% in Microsoft's FY26 Q4, AWS grew 37% in 2026 Q2, and Google Cloud grew 82% in 2026 Q2. Microsoft said customer demand still exceeded available capacity.	Current revenue and capacity signals argue against declaring an immediate infrastructure glut. They do not validate every project now under construction.
Is investment absorbing cash?	Alphabet's first-half cash capex was $80.6 billion against $84.9 billion of operating cash flow. Meta reported $31.1 billion of Q2 capex and $0.8 billion of free cash flow. Amazon reported trailing-twelve-month free cash outflow of $7.6 billion, primarily because property-and-equipment purchases rose.	Cash cushions have narrowed at several firms even while operating businesses remain large and profitable.
Are financing and product demand coupled?	Amazon disclosed investment, financing, and long-term cloud arrangements with Anthropic and OpenAI. NVIDIA disclosed equity commitments, cloud-capacity purchases, customer guarantees, and long-dated supply commitments.	The coupling is no longer only a media hypothesis. The filings still do not show that the related purchases are uneconomic or unable to reach independent users.

The observations come from Alphabet's 2026 Q2 10-Q, Microsoft's FY26 results, Meta's 2026 Q2 release, Amazon's 2026 Q2 10-Q, and NVIDIA's fiscal 2027 Q2 10-Q. Their periods and accounting definitions differ, so the figures are signals rather than a league table.

Four hypotheses to keep separate

Hypothesis	What would support it	What would weaken it
infrastructure overbuild	falling utilization, delayed projects, falling rental prices, persistent capex ahead of external demand	high utilization, durable contracted demand, improving unit economics
equity bubble	valuations require implausible revenue or margin growth	cash flow catches up while valuation multiples normalize
credit bubble	weak borrowers fund short-lived assets with expensive or mismatched debt	long maturities, strong coverage, limited recourse, diversified customers
circular financing	suppliers or investors fund customers whose purchases support the supplier's own revenue	ordinary arm's-length sales, transparent terms, limited reciprocal exposure

These mechanisms can coexist, but they are not interchangeable. The current evidence supports unusually heavy investment and material financing links. Equity overvaluation still requires a valuation test; a credit bubble still requires evidence about borrowers, maturities, coverage, and refinancing; circularity requires tracing how financed purchases become revenue and who absorbs a shortfall.

Research path

1. Financial metrics defines what to extract from filings before assigning a Minsky-style stage.
2. Financing structure distinguishes normal project finance from genuinely circular exposure.
3. Monitoring turns the thesis into observations that could support or falsify it.

Evidence still needed before a market call

For each company or transaction, retain the exact period, currency, accounting definition, source document, and excerpt. At minimum:

• cash-flow statements and capex reconciliation;
• debt maturities, coupons, covenants, guarantees, and recourse;
• lease and purchase commitments;
• customer or supplier concentration where disclosed;
• utilization, pricing, backlog cancellation terms, and contracted duration;
• revenue and margin that can reasonably be tied to the infrastructure.

The next review should use 2026 Q3 filings to test whether cloud growth, utilization, and cash generation are catching up with commitments. A broad bubble call would become more credible if independent demand slows while cancellations, refinancing pressure, guarantee exposure, or unsold capacity rise together. It would weaken if external workloads absorb new capacity and cash coverage recovers without further financing support.

My reading · Opened notes are remembered in this browser.