Accent colours the marks — active tab, underline, the main button. The frame, grid lines and body text stay graphite.
Blog
Notes · Money
The restaurant that loses money on every plate
A finance YouTuber lays out the bear case for the AI bubble in three moves: nobody really trusts the technology, the unit economics are broken in a way software has never been, and China distills the American frontier for cents and gives it away. Here's what's structurally solid, where it leans on people who are selling something, and the handful of signals actually worth watching.
9 July 2026 · 8 min read · Andrei Jikh
Written with AI tools, reviewed by me on its substance — how these are made →
Andrei Jikh is a finance YouTuber, not a lab insider, and this is a bear video — it's built to make you nervous. But strip the thriller framing and there's a real argument underneath, one I keep hearing from more careful people too. His claim is that the whole US market is propped up by a single story: American companies will make trillions from AI forever, because the world will have no choice but to rent it from them. He thinks that story has three holes.
Problem one — nobody trusts it
The trust argument comes mostly from Alex Karp, the Palantir CEO, and it's sharp even though he's talking his book. If AI were as valuable as the marketing says, he asks, why sell it by the token instead of taking a cut of the value it creates? You pay a lawyer to win the case, a contractor to finish the house — the price is attached to a result. AI charges you per word whether the answer was brilliant or garbage. Add hallucinations nobody can fully fix, and the real fear for enterprises: your data flows through the model, and the vendor learns your business. Jikh's example is Anthropic launching a design tool while partnered with Figma — you pay millions to train your own replacement.
"Why are they charging for tokens if it's so valuable?"
Worth naming out loud: Karp's answer to all this is to buy Palantir's sovereign, self-hosted models. The interview was, in effect, a product launch. The question is still good; the messenger is selling the cure.
Problem two — the money doesn't add up
This is the strongest part. Software is the best business ever invented because you build it once and every extra customer is nearly free money — costs flat, revenue up, and the gap is profit. AI breaks that: every single query burns electricity and wears down chips, so more customers means more cost, dollar for dollar. It's not software; it's a restaurant that loses money on every meal and plans to fix it by serving more meals. The receipts pile up — OpenAI's 2025 burn (the video says ~$21B; the CNBC clip on screen shows a $38.5B net loss on $13B revenue — different metrics, both ugly), Oracle building 7 gigawatts for a single customer it admits might not pay, Nvidia quietly financing the 'NeoClouds' that buy its chips and rent them back. And the tell: hyperscalers report cloud, ads, YouTube revenue to the penny, but never break out AI revenue. Public companies love good news.
"It's a restaurant that loses money every time it serves food, and its plan to fix it is to serve more food."
Problem three — the world has another option
The moat depends on there being no alternative, and China is the alternative. America is spending toward a trillion a year, ~3% of its economy; China is spending a fraction, ~0.6% of its. It wins on price through distillation — you train a small cheap model on the answers of an existing frontier model, like copying the homework once someone else has done it. So the US funds the hardest research in history, China compresses the result and open-sources it for free. On the same task, the video shows a top Chinese open model running roughly 7–12× cheaper than Claude Opus at close quality. The kicker: one of those competitive models, LongCat, was built by Meituan — a food-delivery company.
"Every dollar of US AI spending is kind of like a donation to the Chinese AI industry."
So when does it pop?
His honest answer is: nobody knows — but he points at where to look. Not at capex stopping (in the dot-com crash, fiber spending ran a full year past the 2000 market peak). The real trigger is subtler: the first hyperscaler that cuts AI spending and gets rewarded by Wall Street for it, giving every other CEO permission to follow. Watch data-center debt drying up, and credit spreads — today a calm 2.6%, which sounds safe until you remember spreads were just as calm in early 2007, months before the crash. Plus Michael Burry's charts: chip stocks at the top of their 15-year range, and the price of AI tokens down ~20% from its May high, even mid-boom.
"The real answer to how long it will take for the AI bubble to pop, if ever, is that no one knows."
Where I'd push back
Two of the three loudest voices here are selling the thesis. Karp sells sovereign models; Ed Zitron is a professional AI skeptic whose whole brand is that there's no business here; Burry has called crashes that never came — 'early,' in markets, is a polite word for wrong. The numbers get slippery too — the spoken $21B burn and the on-screen $38.5B loss get blurred together. And 'China is 90% as good for 10% of the price' quietly assumes the cheap distilled model stays 90% as good as the frontier keeps moving, which isn't guaranteed. But the structural spine survives all of that: AI genuinely isn't a software margin business, the circular Nvidia financing is real, and open Chinese models really are eating the low end. Those don't need Karp or Burry to be true.
- Keep one line: AI isn't software. Every query has a marginal cost, so scale doesn't automatically print money the way it did for Excel.
- The moat is 'the world has no choice.' China's cheap open models are the counter-evidence — watch the low end, not the frontier.
- Distillation means US research partly subsidizes its competitor. Every open frontier result is a free training set.
- The pop signal isn't capex stopping — it's the first hyperscaler getting rewarded for cutting it.
- Discount the messengers, keep the mechanism. Karp, Zitron and Burry all have an angle; the unit economics don't.