Crypto Bros: The AI Bubble Comparison Everyone Gets Wrong

Two weeks ago the Bun team announced they’d finished rewriting the entire runtime in Rust: about 535,000 lines of Zig, ported in 11 months, with Claude doing a large share of the work. Depending on which corner of my feed I was reading, this was either the future of software arriving on schedule or the 2026 equivalent of Long Island Iced Tea renaming itself Long Blockchain.

The dismissive take has a name, and I hear it weekly: the AI boom is crypto bros v2. Same hustlers who shilled tokens in 2021 now selling prompt-engineering courses. Same thought-leader threads, same FOMO shipping, same “ngmi if you’re not building with agents.”

So I went looking at what the credible skeptics actually say. The surprise: almost none of them say crypto.

The Cultural Case Is Real, and It’s Vibes

Concede the cultural half up front: it’s true. Cory Doctorow made the canonical version in 2023, and John Gorman’s title said it plainer: “All These Insufferable AI Bros Were CryptoBros Last Year.” The migration was literal. The accounts that sold NFT drop calendars now sell agent-stack cheat sheets. The course economy re-skinned itself in six weeks.

But culture is the weakest form of the argument. Grifters chase every boom; they chased the real internet in 1999 too. The question is whether the economics look like crypto. That’s where the serious skeptics part ways with the meme.

What the Receipts Say

• The revenue gap is genuine bubble evidence. GMO (January 2026): under $50 billion of AI application revenue against a trillion-plus of investment, market CAPE around 40 versus the dot-com record of 43.5. CB Insights counts 498 AI unicorns worth a combined $2.7 trillion.
• The circular deals are unprecedented. OpenAI’s 2025 spree: $250 billion of Azure from investor Microsoft, an up-to-$100 billion staged Nvidia investment tied to 10 gigawatts of Nvidia systems, a roughly $300 billion Oracle deal that moved Oracle’s stock 36% in a day. Money in, money out, same names on both sides.
• The spend is going on the credit card. The four hyperscalers spent $433.9 billion in the year to Q1 2026, with about $700 billion planned for 2026, increasingly bond-funded. Capex runs at three times reported depreciation. Michael Burry alleges $176 billion of understated depreciation through 2028 via stretched GPU useful-lives; contested, but the ratio itself is arithmetic.
• Enterprise ROI is missing. MIT’s NANDA study: 95% of enterprise GenAI pilots delivered no measurable P&L return, even as individual adoption runs hot. People use it constantly. Companies can’t turn that into margin.

You can see why “crypto” feels close. Circular money, valuations detached from revenue, losses growing with scale: OpenAI made about $12 billion in 2025 and lost about $8 billion doing it.

Where the Analogy Breaks

Bloomberg’s deep dive on the circular deals drew the line the meme skips: these aren’t fraudulent round-trips. Real compute gets built, real workloads run on it. The risk is concentration, not fakery. Dave Karpf, no AI booster, put the disanalogy in one sentence:

The fundamental value of bitcoin is basically zero. AI has several uses that generate economic value.

— Dave Karpf, 'What Sort of AI Bubble Are We In?'

OpenAI’s $12 billion of revenue is the tell. The problem is the $8 billion loss, not the product. Twelve billion dollars of people paying for a thing is something no crypto project ever produced.

The Rewrite Mania

Which brings me back to Bun, because the builder scene has its own version of the capex problem, and I don’t think we’re honest about it.

The rewrite is a real artifact: it exists, it passes 99.8% of the test suite, 22 million monthly downloads run on it. But “real artifact” and “worth building” are not the same claim. What did users get? A runtime that does what the old runtime did. The stated payoff is memory safety and stability; the demonstrated payoff, so far, is a press release about what AI can port in 11 months. The value question gets answered with an activity metric: lines migrated, tests passing, tokens burned. That’s capex-as-proof, the hyperscaler move at hobbyist scale.

And Bun at least had a reason. Scroll the feed: rewrites of working tools in new languages, frameworks reimplemented “agent-first,” codebases regenerated because regenerating them became cheap. The justification is always capability: look what it can do. Just because you can doesn’t mean you should, and “should” is exactly the question nobody’s answering. Where are the receipts, the measured payoff that justified the spend? Still nowhere. That MIT 95% number isn’t just an enterprise problem; it’s the same gap between usage and value, and we’re on both sides of it.

The Scarier Comparison

Here’s why the crypto-bros framing, satisfying as it is, is a comfort blanket. If AI is crypto, the tech is fake, the bust is deserved, and those of us building real things are safe. The skeptics with receipts tell a worse story: Bloomberg’s fiber-swap framing, Ed Zitron’s Nortel references, VC Paul Kedrosky’s numbers. Kedrosky’s is the sharpest: AI capex is approaching 1.2% of US GDP, above the roughly 1% peak of the fiber buildout, except fiber stayed useful for thirty years and GPUs may be obsolete before they pay back.

Crypto’s bust validated the skeptics: the thing was hollow, and it popped. A telecom-style bust validates nobody. The models keep improving, your workflow survives, and the financial layer above it detonates anyway, taking jobs, startups, and subsidized API pricing with it.

What I’m Doing With This

• Stop arguing about whether the tech is real. It is. So was fiber. The question is whether the financing survives the gap between capability and cash flow.
• Assume the subsidy ends. Everyone shipping on discounted inference is a customer of the burn. When the money-go-round slows, your unit economics change before your model quality does.
• Demand receipts, including from yourself. Before the next rewrite or regeneration: what does anyone get that they didn’t have? If the honest answer is an activity metric, it’s capex cosplay.
• Keep mocking the grifters. Just don’t let the mockery do your economic analysis for you.

The crypto bros comparison flatters everyone: skeptics get to dismiss the whole thing, builders get to say “but my thing works” and stop thinking. Telecom 1999 flatters nobody, which is usually the sign you’ve found the right analogy.