Is AI a bubble?

Why — the first-principles explanation

Separate four questions that are often collapsed into “the AI bubble.” First, is the technology useful? A useful technology can still be sold at an irrational price. Second, are public and private valuations consistent with plausible future cash flows? That is an estimate, not a present fact. Third, how is the infrastructure financed? Cash, equity and debt create different failure paths if returns disappoint. Fourth, how much of the current macro economy depends on continued investment?

The BIS Bulletin “Financing the AI boom: from cash flows to debt” provides measurable evidence for the financing and macro questions. By mid-2025, US expenditures on IT manufacturing facilities and data centres were equivalent to about 1% of GDP, while total IT-related investment including other equipment and software reached about 5% of GDP. Data-centre and semiconductor-facility investment contributed an average 0.4 percentage points to US GDP growth over the subsequent three years; total IT investment accounted for almost half of GDP growth in recent quarters. Those are investment and contribution measures, not proof of future profits.

The financing signal is more cautionary. The BIS reports that outstanding private-credit loans to AI-related sectors grew from near zero to over $200 billion, rising from less than 1% to almost 8% of outstanding private-credit loan volumes; it estimates around $300–600 billion by 2030 under projected investment growth. The BIS says macroeconomic and financial stability risks appear moderate, but warns that the boom’s sustainability hinges on high earnings expectations and notes a disconnect between debt pricing and equity valuations. Debt does not prove a bubble; it increases the cost of being wrong.

A useful conclusion can therefore be conditional: real infrastructure, genuine scientific progress, excessive valuations and future distress can coexist. The 1990s telecom buildout left useful fibre after many investors lost money. Do not ask for a yes/no prophecy. Ask what revenue, margins, utilization, cash flow and refinancing assumptions must be true—and what happens if each one misses.

An example that makes it click

How to do it

1. Define the claim: are you asking about model capability, a company’s valuation, private-market funding, infrastructure spending or the wider economy?
2. Separate observed facts from forecasts. Record the date, source, geography and definition for every number before comparing it with a price or valuation.
3. Read the business evidence: recurring revenue, customer retention, gross margin, cash flow, capital expenditure, depreciation, backlog and concentration.
4. Map the infrastructure economics: data-centre utilization, power and cooling costs, chip replacement, network constraints, lease terms and who bears stranded-asset risk.
5. Trace financing: operating cash flow, equity issuance, corporate bonds, leases, private credit, guarantees and off-balance-sheet structures.
6. Stress-test the thesis with slower adoption, lower prices, model substitution, higher energy costs, delayed construction, regulation and a refinancing shock.
7. Compare equity and debt signals. Different risk pricing can reveal disagreement, but it is evidence to investigate—not proof that one market is correct.
8. Check the adoption denominator. “Users,” pilots, revenue, production deployments and accepted outcomes measure different things and should not be added together.
9. Set a valuation and concentration rule before buying or committing capital. A good technology thesis does not remove portfolio, liquidity or tax risk.
10. Revisit the evidence as earnings and filings arrive. If the required assumptions stop holding, reduce exposure or change the thesis; do not average down solely because the technology is important.

Key facts

• The BIS January 2026 Bulletin says AI investment is surging and accounts for a substantial share of economic growth, while the boom’s sustainability depends on high earnings expectations.
• By mid-2025, BIS estimates put US investment in IT manufacturing facilities and data centres at about 1% of GDP and total IT-related investment at about 5% of GDP.
• BIS estimates data-centre and semiconductor-facility investment contributed an average 0.4 percentage points to US GDP growth over the three years after 2022; total IT investment accounted for almost half of GDP growth in recent quarters.
• Outstanding private-credit loans to AI-related sectors grew from near zero to over $200 billion, and their share rose from less than 1% to almost 8% of outstanding private-credit loan volumes.
• The BIS estimates private credit to AI-related firms could reach around $300–600 billion by 2030 under projected AI-investment growth; this is a scenario estimate, not a guarantee.
• The BIS says current macroeconomic and financial-stability risks from the AI boom appear moderate, while warning that a failure to meet expected returns could produce sharp corrections.
• The BIS identifies a disconnect between debt pricing and equity valuations: loan spreads on AI-related private-credit loans are close to other loans while equity valuations imply outsized future returns.
• Stanford’s 2026 AI Index says technical capabilities, investment and adoption are rising while data transparency and governance preparedness lag, making independent measurement more important.
• A useful technology and a bad investment can coexist. Cash flows, valuation, leverage, utilization and refinancing—not the word “AI”—determine the investment outcome.
• A thematic forecast is not personal financial advice. Country, account, tax, liquidity and suitability rules differ; use current filings and qualified local advice for an actual decision.

What authoritative sources say

People also ask

The same question, asked other ways

This page answers one intent expressed in 2 phrasings. How the index is organized →

• Is there an AI bubble?

Related questions