Which layer of the AI stack is the bubble in?

The middle layer means cloud, orchestration, data and compute leasing. It carries the most debt: circular financing, credit pushed down the chain, and GPU depreciation mismatched against bond maturities all sit here. FutureX's reading: the bonds run 7 to 10 years, the accounting life of the assets 5 to 6, their economic life perhaps 2 to 3, and FutureX stays away from debt-funded compute leasing. Since March 2026 FutureX has scored seven signals daily; in August 2026 six were lit and liquidity had triggered. To spot the top it watches one thing: when capex growth slows.

On August 17, 2026 the 30-year US Treasury yield touched 5.31%, and the next day 5.336%, the highest since 2007 (public market data). On August 26 Nvidia's CFO put the top five cloud providers' capex at about $800 billion for 2026 and $1.3 trillion for 2027 (confirmed).

The bubble is in neither the model layer nor the application layer. The model layer cannot hold value: no network effects, no switching costs, a gap of about 2.7% between the top US and Chinese models (LM Arena, March 2026), and wholesale tokens already in a gross-margin war. The application layer can hold it: revenue released by falling costs lands there, and the ARR is real. Consensus takes three forms, as in 2000: concentration, circular financing and sameness of narrative. FutureX's view is that the bubble verdict matters less than finding the layer where value is highest.

Securitized data-centre debt grew from $1.3 billion in 2022 to about $26.5 billion in 2025, and JPMorgan expects $30 to 40 billion in 2026 (public statistics); the ten largest US stocks make up about 37% of S&P 500 market value, above the 27% peak of 1999 to 2000 (Goldman Sachs, public statistics).