The bubble callers are back — Gartner's neocloud doomsday, GPUs gathering dust in Amazon warehouses, data center debt landing on insurer balance sheets, and revenue concentrated in a sliver of VC-backed power users. Today's episode sorts the signals that actually matter from the proxy fights, and for each one, names exactly what would have to happen to turn caution into alarm.
Every bubble signal is a proxy fight for one question: demand.
Warehoused GPUs, distressed neocloud bonds, pension fund exposure, frozen IPOs — each has its own logic, but they all collapse into whether AI demand ultimately exceeds what's being spent on the infrastructure. Most of today's warning signs aren't yet what the doomsayers claim: the GPU pile is a power problem, not a demand problem; data center debt isn't being disguised the way subprime was; the neoclouds are minnows the industry keeps backstopping. But each has a clear escalation path — and the Anthropic IPO is about to replace speculation with audited quarterly answers.
Americans for Financial Reform published a report Wednesday warning the AI bubble is primed to pop, with analyst Maya Jenkins arguing working people will pay a high price for an AI crash and that policymakers should be taking urgent action. The report's real target: arguing against an AI bailout, which many are certain the administration would indulge at the earliest opportunity.
The Nasdaq dropped 3% Thursday amid a three-week tech downtrend, with Goldman reporting hedge funds sold tech at their fastest pace in at least a decade and Meta signaling its next AI build-out will be partly funded by equity dilution. Meanwhile the 30-year Treasury hit a one-month low of 5.38% — early, but a classic flight-to-safety pattern of money leaving the AI trade for safe havens.
Gartner claims $3.5 trillion in AI infrastructure spending faces impairment due to lack of demand: data center owners need $8.5 trillion in revenue through 2030 to cover capex, financing, and return on capital, against a total AI revenue forecast of no more than $3.6 trillion. Chief forecaster John Lovelock told CNBC that compute capacity currently exceeds demand and nearly every AI market they track is likely to collapse to a duopoly.
— Ken Griffin, Citadel CEO, at a Wall Street conference. Griffin's argument hinges on uncertain ROI forcing a spending rethink: "all of corporate America is looking at their bills from Anthropic and OpenAI and going, 'Okay, we're going to need to think about how we're sure we're getting productivity out of this.'" The grain of salt: Citadel profits most when markets are volatile.
— Jon Gray, Blackstone President, on Bloomberg. Blackstone's president answered Griffin directly on Bloomberg, while colleague Jas Kaira dismissed data center debt fears because hyperscaler tenants are "the most creditworthy companies in the world." The caveat: Blackstone holds massive equity and debt positions across the neocloud sector, including early stakes in CoreWeave and Firmus.
Every talking head in this debate has a dog in the fight, but there's one through line that ultimately matters: how much AI demand will there be, and is the value of that demand higher than what's being spent on the infrastructure? Every other bubble signal is a proxy for that question.
Apocalyptic 'AI bubble' talk flattens a crucial distinction: a 20-30% bear market is painful but self-healing, while a credit crisis — mass defaults and contagion impairing banks and the broader financial system — can require central bank intervention. The dot-com crash wiped over 75% off the Nasdaq across two years; you can have one kind of crash without the other.
Firmus's Australian IPO was canceled when banks couldn't find buyers at the desired price, a portion of CoreWeave debt traded down to distressed levels implying a reasonable chance of default, and Microsoft pulled out of a UK data center over power delays, leaving NScale high and dry. It's the one segment of the AI trade genuinely worth watching closely.
Anthropic is rumored to be in talks for a $45B NScale contract, Goldman upgraded CoreWeave to a buy, and NVIDIA may invest hundreds of millions in Firmus after its failed IPO — on top of the $170B in credit guarantees and rental backstops NVIDIA has already provided neoclouds, per The Information. Perspective matters: against an industry whose key firms are worth ~$10 trillion collectively, a $30B company is a minnow — the economic risk is small, but the narrative effect of a failure would be gigantic.
One: a neocloud gets into trouble and nobody rides to the rescue. Two: distress spreads from CoreWeave's own bonds to project debt for data centers leased to Meta, Microsoft, or Google — markets doubting the tenant, not just the landlord. Three: a hyperscaler or lab walks away from capacity because of demand, not power — Microsoft's UK exit was blamed specifically on power delays, and a demand-blamed cancellation would be a totally different story.
Business Insider reports AWS expects roughly $5 billion in NVIDIA GB300 racks staged at US facilities before final deployment, with ~$3 billion more each month for the foreseeable future — and at least one team with no 'clear forecasted path' to installing them because the power isn't there. Amazon says staging hardware ahead of construction is best practice, and that it's securing scarce GPUs ahead of competitors.
Critics cast the AWS story as NVIDIA channel-stuffing and proof that demand is soft — but Amazon wouldn't order tens of billions in chips to warehouse them indefinitely, GPUs still trade at a premium with old hardware selling above original price, and the Occam's razor read is fear of losing future allocation. The dark fiber comparison fails: that was too little demand for installed infrastructure; this is too little infrastructure to service demand.
The Information reports giant insurers including American General Life and Fidelity & Guarantee Life — portfolios managed by Blackstone — bought more than $740 million of CoreWeave's last bond sale. Meaningful, and a default would leave a hole, but spread across portfolios that size it's nothing like the mass-impairment setup of the GFC.
The Apollo-Athene model — private equity firms merging with insurers for favorable tax treatment and a steady capital flow — is a genuine pocket of AI-debt exposure. But these are mostly offshore life-and-annuity insurers, not State Farm: an AI credit default could impair them, but the risk lands mainly on wealthy annuity investors, not the broad public.
The Wall Street Journal reports AI drove $138 billion in gains across 13 public pension funds since 2022 — exposure is unavoidable when AI has produced the bulk of market returns for four years. Americans for Financial Reform warns data center debt is being repackaged and sold to pensions and endowments 'in a manner eerily reminiscent of the subprime bubble.'
The problem in 2008 wasn't chopping mortgages into CLOs — it was rating them investment grade and treating them as safe as Treasuries, with the structuring explicitly used to obfuscate risk. Data center debt today sits on balance sheets as ordinary corporate bonds, sometimes investment grade and sometimes not, and no one is pretending it's safe as houses. What would change the picture: disclosed exposure jumping by orders of magnitude into a handful of PE-owned insurers, or debt repackaged into securities rated higher than the loans underneath.
Anthropic's and Musk's public equity gifts revived fears that the labs are engineering bailout protection — but the government didn't stand aside for GM in 2008, and it's massively unlikely to let a trillion-dollar company blow a hole in the economy regardless of who holds the shares. The 'engineering a bailout ahead of a crash' theory doesn't hold up either; the escalation signal to watch is AI companies asking for explicit federal guarantees, loan programs, or financing backstops before any crisis.
Two customers made up 39% of NVIDIA's Q2 revenue; a fifth of Anthropic's first-half revenue came from its top two customers per the leaked IPO prospectus; and Ramp data shows the top 1% of businesses account for 80% of OpenAI and Anthropic spending on its platform. But the dynamics differ: NVIDIA's concentration reflects how few companies can build multi-billion-dollar data centers, and much of Anthropic's is wholesaler-distributor flow through hyperscalers — revenue quality only becomes a real issue when a big customer could plausibly pull the plug.
Ramp found 80% of its top 1% of AI customers are VC-backed startups — some wrapping frontier tokens in VC-subsidized services, much of it internal use funded by venture dollars. That's fine while AI venture is a raging bull market; the problem arrives if VC money dries up and those startups hit the end of their runway.
If the startups spending ~$9,000 per employee per month are the top of the chart now, will all enterprises get anywhere near that in a decade — or is AI spend permanently subject to power laws? Encouragingly, Ramp sees businesses across all cohorts tripling spend per employee since the summer, with the top 10% going from roughly $200 to $800 per employee and growing faster than the median. But there's no guarantee.
Today the revenue-quality debate runs on Ramp data, the often-misunderstood Silicon Data Token Index, and other unofficial slivers of the picture. Post-IPO, Anthropic and OpenAI will report fully audited financials quarterly and face equity researchers on earnings calls — if revenue quality isn't improving, everyone will know.
Anthropic's viral $80 billion first-half loss was an accounting figure; the true cash loss was $8.7 billion. Still, the air is coming out of the numbers: The Information reports OpenAI's round will close at $1.25 trillion versus earlier reporting of $1.4 trillion — possibly misreporting rather than a downgrade, but investors are reading it as one — while Anthropic backers who once floated $2 trillion are now reportedly pushing for $1.5 trillion.
Less than $600 million has been raised this month, with Truist's head of equity capital markets saying 'people are truly paralyzed' and Deloitte citing a stalemate on valuation. If neoclouds can't go public, they can't access their next tranche of funding — leaving private equity holding the bag.
Frothy markets often top on a catalyst that breaks the suspension of disbelief, and for many the Anthropic IPO fits the bill. Pricing at $1.5 trillion after months of $2 trillion talk would be not great but not alarming; pricing at $1.5 trillion and then getting a weak market reception would confirm a lot of people's worst fears about the AI trade.
The bubble-burst narrative recurs around this time every year: 2024's scaling-wall panic preceded the first reasoning models, and last year's 'enterprises won't keep paying' fears — embodied in the MIT 95%-of-pilots-fail report — preceded enterprise spend booming at around 50% a quarter. Previous calls being wrong doesn't make future ones wrong too; it just means every bombastic headline deserves a bag of salt while we wait for the Anthropic IPO to deliver a more definitive answer.