The Most Concerning AI Bubble Signals

The AI Bubble Signals Worth Taking Seriously

Nathaniel Whittemore spent an episode walking through every loud AI bubble warning and concluding most don't add up to a system-wide crash. He's probably right on the macro. The useful part is the plumbing he exposes: who's actually paying the frontier labs, who's funding the data centers, and what breaks first if the money gets nervous.

The revenue base is the soft spot. Ramp's data shows the top 1% of businesses drive 80% of OpenAI and Anthropic spend, and 80% of those are venture-backed startups burning other people's money. That's not a healthy revenue base. It's a cohort that stops spending the quarter their next round doesn't close. "Enterprise spend is growing 50% a quarter" is true and also exactly what you'd see right before a venture-funded demand cliff. Whittemore treats the growth as proof of health. The problem is who's funding it.

The compute crunch is real, but the bottleneck isn't chips. Amazon is taking delivery of $5 billion in NVIDIA GB300 racks and can't plug them in because there isn't enough power. The thing throttling AI capacity right now is electricity and grid permits. Powered rack space has a lead time measured in years, and the bottleneck is a substation. The GPU-rental firms, CoreWeave and NScale among them, are the fragile layer. Gartner says 80% die by 2030, and there's a $4.9 trillion gap between what infrastructure owners need to earn and what AI revenue is projected to deliver.

The numbers that actually got corrected matter. Anthropic's "$80 billion loss" that went viral was an accounting figure; actual cash burn was $8.7 billion. NVIDIA's top two customers are 39% of revenue; Anthropic's top two are around 20%. Those concentration figures come from a leaked IPO prospectus, which means for the first time there's a document, not a vibe. The $740 million in CoreWeave bonds sitting in Blackstone-managed insurance portfolios is a rounding error against those balance sheets, so the subprime comparisons are overcooked. Data-center debt is priced as corporate bonds that everyone knows are risky, not AAA paper pretending to be safe.

What changes once Anthropic goes public. The opacity that supports a $1.5-to-2 trillion valuation disappears the day audited financials hit. You'll see the gross margins. You'll see what inference actually costs. Once a public market can see thin margins, the pressure to raise prices or tighten free tiers becomes visible and real. If your product depends on the Anthropic or OpenAI API, get multi-year pricing where you can, and keep a second model wired up so you're not captive when terms tighten. The case for running an open-weight model, Llama, Qwen, Mistral, on hardware you control is that it doesn't get repriced when CoreWeave's bonds trade distressed or when Anthropic's IPO disappoints.

The call: When Anthropic's prospectus goes public, its gross margin will come in under 60%, confirming that frontier-lab economics are thinner than the valuation implies. The company is reportedly negotiating a $45 billion compute contract with NScale. A company paying that much to rent GPUs cannot carry software-style margins, and nothing in the reporting suggests otherwise. We're wrong if audited financials show 60% or higher.

The scarce resource in AI right now is electricity. If you're buying compute, the substation sets your deadline.

Nathaniel Whittemore spent an episode walking through every loud AI bubble warning and concluding that most of them don't add up to a system-wide crash yet. He's probably right on the macro. The useful part is the plumbing he exposes: who's actually paying the frontier labs, who's funding the data centers, and what breaks first if the money gets nervous.

What's actually being decided. Nothing you can undo or not undo. This is a read on the health of the companies you rent models and compute from. But it names three things that matter if you buy AI: who OpenAI and Anthropic's real customers are, whether the GPU-rental firms you might depend on survive, and what happens to pricing once Anthropic's finances go public. No deadline, except the Anthropic IPO, which turns a pile of speculation into audited numbers.

The Skeptic. Whittemore's rebuttals are mostly sound, but he waves off the one signal that should worry a buyer. Ramp's data shows the top 1% of businesses drive 80% of OpenAI and Anthropic spend, and 80% of those are venture-backed startups burning other people's money. That's not a healthy revenue base. That's a cohort that stops spending the quarter their next round doesn't close. Ken Griffin's point lands harder than Whittemore admits: enterprises are staring at their Anthropic and OpenAI bills and asking where the productivity went. "Enterprise spend is growing 50% a quarter" is true and also exactly what you'd see right before a venture-funded demand cliff.

The Compute Pragmatist. The AWS warehouse story is the clearest thing in the episode. Amazon is taking delivery of $5 billion in NVIDIA GB300 racks, $3 billion more a month, and can't plug them in because there isn't enough power. Whittemore calls that rational inventory hoarding, not weak demand. He's right. The thing throttling AI capacity right now is electricity and grid permits, not appetite. For anyone planning compute, that means powered rack space has a lead time measured in years, and the bottleneck is a substation, not a chip order. The NEO clouds, meaning the GPU-rental firms like CoreWeave and NScale, are the fragile layer. Gartner says 80% die by 2030. The math is a $4.9 trillion gap between what infrastructure owners need to earn and what AI revenue is projected to deliver.

The Researcher. Watch the numbers Whittemore corrects, because the corrections are the story. Anthropic's "$80 billion loss" that went viral was an accounting figure; the actual cash burn was $8.7 billion. NVIDIA's top two customers are 39% of revenue. Anthropic's top two are ~20%. Those concentration figures are real and they come from a leaked IPO prospectus, which means for the first time there's a document, not a vibe. The $740 million of CoreWeave bonds sitting in Blackstone-managed insurance portfolios is a rounding error against those balance sheets, so the 2008 subprime comparison from Americans for Financial Reform is overcooked. The data-center debt is priced as corporate bonds that everyone knows are risky, not AAA paper pretending to be safe. That distinction matters.

The Enterprise Buyer. Here's what changes for anyone signing AI contracts. Once Anthropic files real audited financials, the opacity that lets people argue about a $2 trillion valuation disappears. You'll see the gross margins. You'll see what inference actually costs them. And the moment a public market can see thin margins, the pressure to raise prices or tighten free tiers becomes visible and real. If your product depends on an Anthropic or OpenAI API, that's a commercial risk you can plan for now: get multi-year pricing where you can, and keep a second model wired up so you're not captive when terms tighten.

The Open-Source Advocate. This is the quiet case for open-weight models that the episode never makes. If the frontier labs are a duopoly funded by venture-backed startups, propped up by NVIDIA guarantees, and headed for a margin reckoning the day they go public, then the single-vendor API bet is the fragile one. A Llama, Qwen, or Mistral model you can run on hardware you control doesn't get repriced when CoreWeave's bonds trade distressed or when Anthropic's IPO disappoints. It's not about matching the top benchmark. It's about not having your cost structure hostage to someone else's fundraise.

The tensions. Whittemore and the Skeptic split on the revenue base: he treats 50%-a-quarter growth as proof of health, the Skeptic treats it as proof of a venture-funded sugar high. The Compute Pragmatist and the bubble callers split on the AWS warehouse: idle GPUs are either channel stuffing (NVIDIA forcing chips out to book revenue) or rational hoarding against scarcity. And the Researcher and the doom crowd split on contagion: $740 million in insurance portfolios is either a subprime echo or a rounding error. In each case Whittemore has the better argument on the facts available today. The part he underweights is the one that matters most to you: who's paying, and whether they keep paying.

What it hinges on. One belief: is frontier-lab revenue resting on companies that will still be funded in twelve months? If venture funding stays open, the growth is real and the bubble talk is noise. If it tightens, the demand cliff hits API revenue fast, because the spend is concentrated in exactly the cohort that loses access to cash first. The Anthropic IPO is the first hard data point either way.

Prediction: When Anthropic's IPO prospectus financials become public, its reported gross margin will come in below 60%, confirming that frontier-lab economics are thinner than the valuation implies.

Confidence: Medium. The leaked prospectus already shows heavy cash burn and a $45B compute commitment.

Why: The leaked prospectus already shows $8.7 billion in cash burn and the company is reportedly negotiating a $45 billion compute contract with NScale, which means compute cost sits as a massive line against revenue. A company paying that much to rent GPUs cannot carry software-style margins, and the opacity that supports a $1.5-to-2 trillion valuation only holds while nobody can see the cost of goods. The opposite outcome, a fat margin that justifies the valuation, would require inference to be far cheaper than the known compute commitments suggest, and nothing in the reporting points that way.

Revisit by 2027-06-30: We're right if Anthropic's first public prospectus or audited financials show gross margin under 60%. We're wrong if they show 60% or higher.

One more thing the episode gets right and is worth repeating plainly: the scarce resource in AI right now is electricity, not chips. If you're buying compute, the substation sets your deadline. The invoice is the easy part.