The Most Concerning AI Bubble Signals

The host opens by saying that fears of an AI bubble are growing as Anthropic’s IPO approaches. Americans for Financial Reform released a report calling the bubble “primed to pop,” with senior policy analyst Maya Jenkins warning that working people would pay a high price if an AI crash occurred and urging policymakers to act. The host says the important question is not whether every bearish headline is correct, but which warning signs matter and what future developments would make them more serious. Market conditions have become less comfortable. Technology stocks have been in roughly a three-week downtrend, the Nasdaq fell 3% on Thursday, and Goldman Sachs reported that hedge funds sold technology stocks at their fastest pace in at least a decade. The 30-year Treasury yield also fell to a one-month low of 5.38%, which the host interprets as an early sign of money moving from risky AI-related assets into safer government bonds. Gartner added to the pressure by predicting that 80% of neocloud companies could go bankrupt by 2030 and that $3.5 trillion of AI infrastructure spending could face impairment. Its forecast says data-center owners would need $8.5 trillion in revenue through 2030 to cover announced spending, financing costs, and returns, while total enterprise and consumer AI revenue might reach only $3.6 trillion. The host contrasts Citadel CEO Ken Griffin’s claim that markets are “almost certainly” in a bubble with Blackstone executives’ more optimistic view that hyperscalers are highly creditworthy tenants and that AI capacity is being built to keep up with demand. The host’s central conclusion is that demand—and whether its value exceeds infrastructure spending—is the issue behind nearly every other argument.

Before examining individual risks, the host distinguishes a normal bear market from a systemic credit crisis. A 20% or 30% stock-market decline can be painful, particularly for people dependent on retirement accounts, but markets can recover if banks and the wider financial system remain sound. A credit crisis is more dangerous because defaults and contagion can impair institutions throughout the economy and potentially require central-bank intervention. The host says discussions using the broad term “AI bubble” often lose this distinction; the dot-com crash, which erased more than 75% of the Nasdaq over two years, represents a much more severe outcome than an ordinary market sell-off. The neocloud sector is currently the market’s main area of concern. Firmus canceled an Australian IPO after banks could not find enough buyers at the desired price. Some CoreWeave debt traded at distressed levels, Microsoft withdrew from a planned U.K. data center because of power delays, and the withdrawal left Nscale exposed. At the same time, there are signs that larger AI companies are trying to stabilize the sector: Anthropic is rumored to be negotiating a $45 billion contract with Nscale, Goldman Sachs upgraded CoreWeave to a buy, and Nvidia is reportedly considering a sizable investment in Firmus after its failed IPO. The host says these neocloud problems should not be dismissed, but their direct economic scale is limited. Nscale is targeting a $30 billion IPO, while the largest AI-related firms together are worth around $10 trillion. The entire neocloud sector could theoretically fail without causing major direct economic damage, although the narrative impact would be substantial. Nvidia has reportedly provided $170 billion of support through credit guarantees and data-center rental backstops, while frontier labs and hyperscalers have incentives to announce multibillion-dollar contracts to keep the infrastructure pipeline intact.

The host identifies specific developments that would make neocloud distress much more alarming. One would be a troubled company failing to find a rescuer, rather than being supported by Nvidia, a hyperscaler, or an AI lab. Another would be distress spreading from CoreWeave’s own bonds into project debt tied to data centers leased to Microsoft or Google. That would suggest investors were losing confidence not merely in a neocloud intermediary, but in the creditworthiness or demand of its hyperscaler tenants. The most consequential signal would be a hyperscaler or frontier lab abandoning capacity because it no longer needed it. The host stresses that Microsoft’s U.K. exit was specifically attributed to power delays. A cancellation attributed to weak demand would be fundamentally different: it would indicate that the infrastructure was being built ahead of customers rather than being delayed by the inability to connect it to the grid. This framework leads the host to a broader conclusion. Financial stress among small infrastructure providers is not itself proof that AI demand is collapsing, especially while larger companies have both the resources and the incentive to prevent failures. The key question is whether distress remains confined to financing structures and construction bottlenecks or begins to reveal that the end customers do not want the capacity. The host therefore treats the neocloud sector as worth watching, but not yet as evidence of an imminent systemic crash.

Business Insider reported that AWS expects to receive about $5 billion of Nvidia GB300 server racks at U.S. facilities before final deployment, followed by roughly $3 billion of GB300 racks per month. Sources said Amazon has room for the racks but not enough power to turn them on, and at least one AWS team reportedly lacked a clear forecast for when installation could occur. Amazon responded that staging hardware before construction is routine and that preparing to plug in GPUs once power becomes available is sound business practice. The report also said Amazon is trying to secure scarce GPUs ahead of competitors. Critics interpreted the story as evidence that Nvidia is forcing hyperscalers to accept chips they cannot use, or “channel stuffing” inventory to recognize revenue. Under that theory, Nvidia’s reported demand would be partly artificial, and unused GPUs would resemble the overbuilt telecommunications infrastructure of the dot-com era. The host accepts the reporting but rejects that conclusion. His “Occam’s razor” explanation is that data centers are being built too slowly and connected to power too slowly, while Amazon is protecting its future access to scarce chips. Old GPUs still sell above their original prices and remain in active use, so storing new hardware for several months does not necessarily imply waste. The host says this story is being stretched too far as a sign that the AI buildout is collapsing. Dark-fiber problems involved too much already-installed infrastructure and too little demand; the AWS situation appears to involve too little powered infrastructure to serve demand. Even if Nvidia has extraordinary negotiating power, the warehouse story would demonstrate pricing power and supply scarcity more readily than fake demand.

The host then examines whether data-center debt could transmit an AI downturn into pensions, insurers, or the broader financial system. The Information reported that large insurers bought more than $740 million of CoreWeave debt in its most recent bond sale, naming American General Life Insurance, Fidelity and Guaranty Life, and Security Life of Denver, whose portfolios are managed by Blackstone. The host says $740 million sounds large in isolation but is a rounding error when spread across large portfolios. It shows that insurers hold this kind of debt, but not yet that they face the sort of mass impairment seen during the 2008 financial crisis. A related risk comes from private-equity ownership of insurers. Apollo merged with Athene in 2022, using a structure associated with Warren Buffett: the insurer receives favorable tax treatment and the investment firm receives a steady stream of capital. The host clarifies that these are generally offshore life insurers and annuity providers, not large household insurers such as State Farm. Their failure could hurt wealthy annuity holders, but it would not automatically create the same public exposure as a collapse of the largest property-and-casualty insurers. Pension funds do have meaningful AI exposure because they invest in stocks and bonds, and the Wall Street Journal estimated that AI generated $138 billion in gains across 13 public pension funds since 2022. The host contrasts this with the subprime crisis: mortgage debt was chopped into securities deliberately marketed as investment-grade and nearly as safe as Treasuries. Data-center debt is generally treated as corporate debt, sometimes investment grade and sometimes not, and its risks are more visible. He would become much more concerned if disclosed exposure rose from hundreds of millions to tens of billions concentrated in two or three private-equity-owned insurers, or if data-center loans were repackaged into securities with higher ratings than the underlying loans.

The discussion turns to Anthropic and Elon Musk’s proposals to give the public equity through different mechanisms. The host sets aside the broader policy questions—whether public ownership is an appropriate way to redistribute AI gains or whether billionaires and large companies should instead pay more taxes—and focuses on the argument that public stakes could make AI firms politically “too big to fail.” The host defines that phrase narrowly as an institution the government feels obligated to protect regardless of what the free market would otherwise do. He says that, in his view, the ship has already sailed. A trillion-dollar company collapsing could create a huge economic shock, and the government intervened to prevent GM’s failure in 2008. Given how much the government-market relationship has changed, he considers it unlikely that officials would simply allow a company of that scale to create a giant hole in the economy, whether or not ordinary citizens own shares. He also rejects the idea that Anthropic and Musk are giving away equity mainly to pre-engineer a bailout. Anthropic may genuinely want to create a large public-benefit fund consistent with its mission. Musk may believe that public giving is effective, may be trying to get ahead of an inevitable redistribution debate, or—more jokingly—may enjoy being loved, with an $11 billion gift to the American public being an efficient route. The host says the issue deserves more attention only if AI companies begin requesting explicit federal guarantees, loan programs, or financing backstops before any crisis occurs.

The host considers revenue quality the most interesting and consequential current risk. Nvidia reported that two customers generated 39% of its second-quarter revenue. The Wall Street Journal’s review of Anthropic’s IPO prospectus found that its top two customers supplied one-fifth of first-half revenue. Ramp data showed that the top 1% of businesses represented 80% of spending volume for OpenAI and Anthropic through Ramp’s platform. The host says concentration becomes dangerous when a major customer can plausibly pull the plug, but Nvidia’s customer concentration is partly structural because only a few companies can build multibillion-dollar data centers. Anthropic’s concentration may also reflect a wholesaler-distributor relationship with hyperscalers that resell access to its models. The more troubling question is whether most enterprise AI demand is coming from a small group of heavily funded power users. Ramp found that 80% of its top 1% of AI customers are venture-backed startups. Some may be wrapping frontier-model tokens into products sold at VC-subsidized prices; others may be using tokens internally with venture capital paying the bills. If venture funding dries up and these companies reach the end of their runways, their AI spending could disappear. There are signs adoption is broadening. Ramp found that businesses across cohorts tripled spending per employee since the summer, while companies outside the top 1%—those in the top 10%—rose from about $200 to $800 per employee. Still, the host asks whether the present power-law distribution will persist or whether it merely reflects different stages of adoption. The risk becomes easier to measure once Anthropic and OpenAI are public companies with audited quarterly financials and executives answering questions from analysts.

The host says Anthropic’s upcoming IPO is a likely catalyst for the entire bubble debate because public markets will have to assign a concrete value to a major AI lab. A reported $80 billion first-half loss went viral, but the host explains that this was an accounting loss; the company’s actual cash loss was reported as $8.7 billion. Even so, valuation expectations have softened. OpenAI’s expected fundraising valuation has reportedly fallen from an earlier $1.4 trillion figure to $1.25 trillion, although the host cautions that the difference could reflect earlier misreporting rather than a genuine downgrade. Anthropic investors had floated $2 trillion, while more recent reporting put the target around $1.5 trillion. The wider IPO market is effectively frozen: fewer than $600 million had been raised during the month. Truist’s head of equity capital markets said people were “truly paralyzed,” while Deloitte’s U.S. capital-markets leader described a stalemate between companies’ valuation demands and what investment banks would offer. That matters to neoclouds because an inability to go public could prevent them from accessing their next funding tranche and leave private-equity investors holding the risk. The host’s key distinction is between pricing and reception. An Anthropic IPO at $1.5 trillion after months of $2 trillion speculation would be disappointing but probably manageable. If it prices at $1.5 trillion and then receives a weak market response, that would validate deeper fears about demand, revenue quality, and the state of the AI trade. The episode closes with caution: previous bubble warnings—in 2024 about scaling limits and last year about poor enterprise returns—were followed by reasoning models, improved chips, and enterprise spending that continued growing by about 50% per quarter. Those precedents do not guarantee future optimism, but they justify taking dramatic bubble claims with “bags of salt” until Anthropic’s public-market test provides better evidence.