The Most Concerning AI Bubble Signals

The central bubble test is whether AI demand and the value it creates will ultimately justify the vast infrastructure spending underway. Americans for Financial Reform called the bubble “primed to pop,” and Maya Jenkins warned that working people could bear heavy costs if it bursts. Ken Griffin said markets were “almost certainly” in a bubble, citing uncertain returns and companies’ differing ability to capture value. John Gray disagreed, arguing that capacity is being built to meet demand rather than in the hope that demand will emerge. Gartner’s John Lovelock said compute capacity currently exceeds demand, partly because freemium services give away tokens and processing power. Gartner estimates data-center owners need $8.5 trillion in revenue through 2030 to cover announced capital spending, financing costs, and returns, against no more than $3.6 trillion in projected enterprise and consumer AI revenue. It also forecasts that 80% of neoclouds could fail by the end of 2030, putting $3.5 trillion in AI infrastructure spending at risk of impairment.

Recent strains include Firmus canceling its Australian IPO after banks could not find enough buyers at its desired price, CoreWeave debt trading at distressed levels, and Microsoft exiting a planned UK data center after power delays, exposing Nscale. Goldman Sachs upgraded CoreWeave to buy, saying debt concerns were overblown and the stock oversold, but the host said it was too soon to dismiss the sector’s jitters. Reports of a possible $45 billion Anthropic contract with Nscale and a potential Nvidia investment of hundreds of millions in Firmus remain rumors. A neocloud failure could have limited direct economic impact but seriously damage confidence in the AI trade. More concerning signals would include a distressed firm failing to find a rescuer, bond distress spreading to projects leased to Meta, Microsoft, or Google, or a hyperscaler or AI lab canceling capacity because it no longer needs it. Microsoft’s UK departure was attributed to power delays, not weaker demand.

Business Insider reported that AWS expected about $5 billion in Nvidia GB300 racks at US facilities before final deployment, followed by about $3 billion more in racks each month, while lacking enough power to switch on some stored hardware. Amazon said staging equipment ahead of construction is routine. The reporting does not establish that Nvidia is forcing deliveries or that AI demand is weakening; insufficient powered data centers is a simpler explanation. Holding scarce, high-priced GPUs in storage could also help Amazon secure future chip access.

The scale and location of AI-related credit exposure matter. The Information reported that insurers bought more than $740 million of CoreWeave debt in its last bond sale, with the named firms’ portfolios managed by Blackstone. The host considered that amount small across large portfolios, though a default would still cause losses. Data-center debt is held as corporate bonds, sometimes investment grade, and has not yet been repackaged into securities rated above the underlying loans. Concentrated exposure of tens of billions of dollars at a few private-equity-owned insurers, or repackaging that obscures the underlying risk, would be more concerning. Americans for Financial Reform said speculative data-center debt is reaching pension and retirement funds, endowments, and other institutions; the Wall Street Journal reported that AI drove $138 billion in gains across 13 public pension funds since 2022. Private-equity firms have also paired with insurers, including Apollo’s 2022 merger with Athene. An AI credit default could impair insurers and expose annuity investors, although the firms discussed are mostly offshore life insurers and annuity sellers, not major auto and home insurers.

Revenue concentration and customer quality remain risks. Nvidia reported that two customers accounted for 39% of its second-quarter revenue, while a reported Anthropic IPO prospectus said its top two customers supplied one-fifth of first-half revenue. Ramp data showed that the top 1% of businesses accounted for 80% of OpenAI and Anthropic spending on its platform, and that 80% of those top AI customers were venture-backed startups. The host cautioned that Ramp’s tech-oriented customers may not represent the wider market and that startup spending could weaken if venture funding dries up. At the same time, Ramp found that businesses across spending cohorts tripled AI spending per employee since summer: the 10% cohort rose from about $200 to about $800 per employee, while the highest-spending startups were around $9,000 per employee per month. Rising spend beyond the top 1% suggests adoption is spreading, but it is unclear whether broader enterprise spending will approach startup levels. Enterprise AI spending is still growing around 50% per quarter, despite earlier predictions that businesses would stop paying.

Uncertain valuations could constrain funding for private AI firms and neoclouds. Reporting put OpenAI’s expected fundraising valuation at $1.25 trillion, down from an earlier report of $1.4 trillion; some Anthropic investors had sought $2 trillion, while more recent reporting described another group seeking $1.5 trillion. West Riggs of Truist Financial said investors were “truly paralyzed,” and a Deloitte and Touche capital-markets leader described a valuation stalemate. If firms and banks cannot agree on IPO prices, neoclouds may struggle to obtain their next funding tranche. The host expects Anthropic’s IPO to provide a clearer test of the AI trade: a strong reception at $1.5 trillion would be less concerning than a weak one, but neither a bubble burst nor the absence of risk is certain.