Platformer's Casey Newton published a piece on April 30, 2026 arguing that the AI bubble, if it is one, looks more like the 19th-century railroad boom than the 2021 crypto bust. The framing matters because it changes what investors, regulators, and operators should expect on the other side of a correction. Railroads bankrupted their financiers and still left a continent wired together. Crypto bankrupted its financiers and left behind mostly memes and lawsuits.
The distinction is not academic. If AI is a railroad bubble, the GPUs, data centers, fiber, and trained models survive the equity wipeout and get repriced into the hands of operators who can run them profitably. If it is a crypto bubble, the spend evaporates and the infrastructure turns out to be load-bearing on nothing. Newton's case is that the underlying usage — coding agents, customer support, search, drug discovery — is real enough to make the railroad analogy the better fit.
That framing is being tested in real time this week. The Musk v. Altman trial opened in Oakland with OpenAI's corporate structure as the central question, and the testimony has spent days picking at how much of the company's value is tied to its for-profit conversion versus its underlying technology. A railroad-style bubble can survive a governance scandal at one operator. A crypto-style bubble usually cannot.
Key facts
- 01Platformer's Casey Newton published the framing on April 30, 2026, arguing AI looks more like a railroad bubble than a crypto bubble.
- 02The piece lands in the same week as the opening of the Musk v. Altman trial over OpenAI's corporate structure.
- 03Newton ties the framing to the ongoing Mythos policy fight, where regulators have struggled to settle on a response.
- 04The railroad analogy implies durable infrastructure survives the crash; the crypto analogy implies almost nothing does.
The Mythos fight is the other backdrop. Regulators have not landed on a coherent posture toward Anthropic's restricted-access model, and the inconsistency was sharpened when OpenAI gated GPT-5.5 Cyber after publicly mocking Anthropic for doing the same. Newton's argument is that policy ambiguity is itself a feature of railroad-style booms — the rules get written after the track is laid, not before.
“The railroad analogy says the capital gets torched but the rails remain. The crypto analogy says the capital gets torched and there is nothing underneath worth keeping.”— Jaeden Schafer
Railroad bubbles also share a specific financial pattern: capital is raised against future traffic that does not yet exist, the traffic eventually shows up, and the original financiers are wiped out anyway because they overpaid for the option on it. Substitute "tokens" for "traffic" and the shape rhymes. The compute being built today will likely be used. The people writing the checks may still lose.
The counter-case is that GPUs depreciate on a timeline railroads did not. A rail line laid in 1870 was still useful in 1920. An H100 bought in 2024 is already a generation behind, and the models trained on it can be retrained on better silicon in 18 months. If the useful life of the infrastructure is short enough, the railroad analogy weakens and the bubble starts to look more like fiber in 1999 — overbuilt, eventually absorbed, but a brutal decade for the people who funded it.
Newton's piece, which sits behind Platformer's paywall, frames this as the live debate inside the AI investor class right now. The bull case is that even a 70% drawdown in AI equities leaves behind a compute base that powers a decade of products. The bear case is that the depreciation curve is steep enough, and the model commoditization fast enough, that the surviving infrastructure is worth a fraction of what was spent to build it.
There are reasons to be skeptical of any tidy historical analogy. The railroad boom played out over decades with a fragmented financial system and no central bank backstop. The current AI buildout is concentrated in a handful of hyperscalers — Amazon, Microsoft, Google, Meta — whose balance sheets can absorb a correction that would have flattened a 19th-century rail baron. The bubble, if it pops, may not pop the same way.
The useful thing about Newton's framing is that it forces a more specific question than "is AI a bubble." The real question is what survives the drawdown. If the answer is data centers, trained model weights, and a cohort of engineers who know how to operate them, the analogy holds. If the answer is a pile of stranded GPUs and a lot of unwound vendor financing, it does not, and the comparison shifts toward telecom in 2001 or, in the worst case, crypto in 2022.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
