What is the actual math behind the bubble worry?
A revenue gap. Fortune’s survey of the debate cites Gartner’s estimate that about 8.5 trillion US dollars of cumulative AI revenue is needed through 2030 to justify current data center spending, and lists the year’s top investor fears: debt-financed buildout and circular funding arrangements between AI companies (Fortune, 17 October 2026, read 19 October 2026). Gartner separately estimates over 3.5 trillion US dollars of AI infrastructure spending exposed to impairment by 2030. The disagreement is not about the spending, which is public, but about the demand curve that must catch it.
What is circular financing and why does it spook investors?
Money that counts twice. A chipmaker invests in an AI company that spends the investment on the chipmaker’s hardware; a cloud provider funds a model lab that commits the funds to cloud capacity. Each deal is defensible alone; together they make revenue look more independent than it is, so the same dollar can appear as demand in several balance sheets. It echoes the vendor financing of the late 1990s telecom buildout, which is why the comparison keeps being drawn.
Who is exposed if the correction comes?
Unevenly, down the stack. The hyperscalers spend from profitable core businesses and would absorb a write-down; the pure-play neocloud layer is debt-financed against GPU collateral that depreciates fast, which is where analysts expect consolidation first. Model labs sit between, priced for growth. For everyone else the exposure is indirect: your risk is not owning the stocks, it is depending on the casualties, the tools, platforms and AI infrastructure that change hands or disappear in a shakeout.
How should a company buy AI during a bubble debate?
As if volatility were certain and the direction unknown. Buy against measured returns, the discipline of AI ROI measurement, so your spending survives scrutiny whatever the market does. Prefer portable architectures over deep single-vendor coupling, and check the balance sheet behind any long commitment. And note the asymmetry: a correction makes compute cheaper for buyers, so a company whose AI pays its own way enters any downturn as the customer everyone wants.