The AI Bubble WILL Burst: Half the Neoclouds WILL Die | Jerry Murdock

Why does it matter? Because the infrastructure being built today is already transitional

Jerry Murdock has managed over $90 billion through every major tech cycle of the last 25 years. His read right now isn't bullish excitement — it's a clear-eyed map of what's about to break. The models everyone is betting on will be made architecturally obsolete. Half the capital flooding into GPU clouds is going to burn. And the credit markets underpinning all of it carry the same complacency signature that preceded 2008. Here's what he actually believes:

• Every current AI model — including Anthropic's and OpenAI's frontier models — will be completely replaced by continuous learning architectures, not upgraded • At least half of all Neoclouds disappear within 36 months; economic disruption kills them on contact • Open Router's 5% inference markup has months, not years, before blockchain exchanges compete it away • SaaS companies without a genuine AI strategy have roughly two years before they become unviable

Every AI model alive today — frontier and open source alike — is already on death row

Every model being trained, fine-tuned, and deployed right now will be thrown away. Not upgraded — replaced entirely.

The mechanism is continuous learning: the ability to update dynamically rather than being trained once and frozen. Physical AI and robotics need this to handle unpredictable real-world events. Frontier labs have it as an explicit goal. The problem is you can't retrofit it. "I don't think you can bolt on continuous learning into an existing frontier model," Murdock says. "I think ultimately it'll call for a new form of architecture and completely new training."

The timeline is uncertain in a cancer-curing kind of way. We've been on the cusp of solving it for 15 years. "We got a little bump with drugs like Keytruda — but we haven't solved cancer yet." Sample-efficient models are the early signal — learning usefully from small datasets — but when the genuine breakthrough arrives, "every model we have today dies, goes away."

SSI, Ilya Sutskever's company valued at $30 billion, reportedly shipped its first continuous learning attempt in late August. Murdock wouldn't invest at that valuation without knowing the founders personally. But the category is real, and when it lands, the hundreds of billions being spent training today's static architectures will look like dotcom fiber — valuable eventually, but the companies that built on it mostly won't survive.

At least half of all Neoclouds are gone within 36 months — a recession kills them on impact

The Neocloud graveyard is coming. "I think at least half of them go away within 36 months, but if there's an economic disruption, a lot of them go away right away." This isn't a prediction about market timing — it's a structural argument about how most of these businesses are built.

GPU capacity is a commodity. The only thing separating surviving Neoclouds is who's running them and whether they're building proprietary operating leverage. Murdock's example: Fireworks versus Base10. Fireworks was built by engineers who understood PyTorch most deeply; they're moving up the stack into fine-tuning and customization. Base10 won the Cursor contract but did it at near-zero margin — "they got revenue and they got scale from it, but they didn't get a lot of earnings." The verdict: "If you're not making a lot of money and you're putting up a lot of money, you're at risk. You're absolutely at risk."

The dotcom parallel is instructive. Fiber laid in the ground was always valuable — the companies that laid it all went bankrupt. "They're going to go away with massive amounts of money being burned as part of it." A credit dislocation doesn't just pressure these businesses; for the undifferentiated ones, it ends them. Scrutinize capital efficiency and gross margin trajectory before valuation. The business model of reselling commodity GPU hours is not a moat.

Gavin Baker says a token is a token — Murdock disagrees, and Fireworks is the proof

Baker's framing: Jensen Huang wins either way, because a token is a token regardless of which model or provider generates it. Murdock pushes back directly: "I disagree with a token is a token."

The logic turns on customization. When you fine-tune a model for a specific task, the output changes in kind, not just quality. Verbosity shifts. Style becomes consistent. "The more you customize the model, the more the token changes its value, right? Because the more you customize what's being produced." Models also vary dramatically in how much they talk — a verbose model burns far more tokens to accomplish identical work compared to a tuned-for-brevity alternative. "Some models they talk a lot more than other models and so they produce a hell of a lot more tokens."

Fireworks makes "a lot more money than Base10" not through volume but through the economics of customization. An enterprise willing to spend a million dollars fine-tuning for a specific workflow extracts far more value per dollar than one routing generic requests to a frontier model. The inference commodity race to zero is real — but only for providers selling undifferentiated throughput. Premium economics flow to whoever enables deep model specialization: style, verbosity, task specificity. That layer isn't a race to zero. It's a margin expansion story.

Developers are running AI agents in 'YOLO mode' inside containers — and containers aren't safe

The most underestimated risk in the entire AI stack isn't model quality or data access. It's sandbox security. "Everybody in my opinion is underestimating it" — including teams that believe they've already addressed it.

The specific failure: developers putting agents and their tool sets inside containers and assuming that's sufficient isolation. "I don't know how many developers are running their models in YOLO mode, but probably a lot." Docker itself said containers aren't safe and launched Docker Sandboxes in response. E2B built its entire business around cloud sandboxes for the same reason. Agents are probabilistic in ways developers aren't — one might open a hundred different sandboxes testing a hundred different libraries simultaneously, generating behavior patterns no static security model anticipates.

Murdock's view on the legal AI sector makes the stakes concrete. Harvey and Agora are burning enormous capital competing for the same customers. "The first one of those guys that has a security problem — and it will happen — is going to wreck their market opportunity." His bet: back the next highly specialized entrant targeting one narrow legal workflow, not either combatant. The broader investment thesis is direct: "If you don't get the sandbox right, forget everything else." Any agentic product shipping without proper sandbox isolation is carrying existential technical debt, not a temporary gap.

The AI credit bubble has a specific burst window — and it rhymes with 2008

October 2026 to March 2027. That's the window Murdock marks if the Iran conflict worsens and credit markets crack. The analogy isn't vague — it's technical: private debt spreads are too narrow for the actual risk being carried, the same complacency signature that preceded the housing collapse. Back then, credit rating agencies were asleep at the wheel. "I see the same thing today."

The specific triggers: Japan holds a trillion dollars in US Treasuries. If they need to sell $300 billion to support the yen, "we would have a real problem on our hands immediately. Immediate global problem." Meanwhile hyperscalers have taken on more debt than at any prior point in their history, with free cash flow at historic lows relative to capex.

The counterintuitive read: hyperscalers survive a dislocation. They have recurring revenue and financial depth to absorb it. "If it happens, it'll be good for them because everybody else gets wiped out and then assets become cheaper for them to acquire." The demand for AI compute doesn't disappear — only the ability to fund it short-term does. For everyone else on the stack, a credit crack isn't a correction. It's a culling. Hold dry powder for the window; if it opens, infrastructure assets become acquirable at discounts.

Open Router's 5% markup is a dead man walking

Five percent on every inference transaction is "a crazy amount of money. That's not going to last." It worked because developers defaulted to convenience — routing through Open Router was easy, and enterprises didn't scrutinize the toll. That window is closing.

Blockchain inference exchanges are already launching. Aki Naki just deployed its Dodex exchange on mainnet with model routing built into the product — eliminating the need for a standalone router entirely. Venice IO is building the same thing. "There's two or three other guys that are now in the process of building exchanges that you can go directly get the inference you need without paying the 5% markup." Murdock's timing: "You're going to see a big disruption in that model in the next 3, four, five months."

The board-level call is blunt: if a $10 billion acquisition offer from Stripe materializes, take it without hesitation. "You say, 'Fuck yeah, this is great.' You take the money when you can." Credit to the team for identifying the opportunity early and executing on developer laziness — but they built a toll booth on a road that's about to be bypassed, not a durable business.

SaaS companies without a real AI strategy have two years — and that clock is running now

Most of the SaaS generation is applying "lipstick on the pig" — a co-pilot feature, an AI toggle, the kind of cosmetic addition that generates a press release without changing the underlying product. The threat they're facing isn't cosmetic.

Agentic systems are beginning to replace the workflows those tools were built to support. "The co-work era it begins the threat — so the threat to the SaaS world is just starting right now." A company selling project management or document review now competes with agents that do the actual work, not just the interface. "I don't believe that you're going to have an opportunity to do much of anything with the company in two years."

Companies sitting at 4-6x revenue multiples that haven't rebuilt core products around agentic workflows are value traps by Murdock's reckoning. The pivot window still exists — barely — for those with a system of record and the conviction to act on it. PE firms with levered SaaS portfolios face the worst of both worlds: the AI threat compressing EBITDA while any credit market disruption triggers margin calls simultaneously. "If you're not doing that now, good luck."

Blockchain's real killer app was never Bitcoin — it's agent payment rails, hiding in plain sight

"Blockchain for agent payments" is Murdock's answer when asked what seems crazy today but will be obvious in five years. He separates the signal from the noise deliberately: Bitcoin's quantum vulnerability, the speculative greed narrative, the hype cycle collapse — that's noise. Solana and Ethereum have genuine long-term potential. Venice IO and Aki Naki are building inference exchange infrastructure on chain. Robin Hood Chain already generated a billion in revenue through tokenized equities.

"Whether you're tokenizing stocks or you're going to do payment rails or you're going to use blockchain for buying inference like on Akinaki — those things are going to be real innovations. People are going to be blown away that blockchain has found true utility other than some supposed form of utility."

The structural case is simple: autonomous agents transacting at machine speed need settlement infrastructure that existing financial rails can't support at scale. Programmable settlement is exactly what blockchain does well. The category is building quietly while Bitcoin noise absorbs all the attention. Separate them.

The next two years will reveal who was building for where AI is going — not where it was

What Murdock describes — across models, infrastructure, routing, SaaS, and security — is a stack that's mid-transition at every layer simultaneously. Continuous learning hasn't arrived. Neoclouds haven't consolidated. Credit markets haven't cracked. Sandbox security is unsolved. All of it in motion at once, with a narrow window before a forced reset arrives from either a financial dislocation or a technical breakthrough that makes today's bets look like dotcom fiber.

The investors who survive won't be those who picked the hottest 2025 infrastructure play. They'll be those who recognized that today's AI buildout is transitional by definition — and held dry powder for what comes next.

Builders who treat current model architectures as durable moats are building on sand.

---

Topics: AI infrastructure, venture capital, Neoclouds, inference providers, credit markets, cybersecurity, open source models, frontier models, continuous learning, blockchain, SaaS disruption, AI bubble, model customization