Is there an AI bubble?

I

The claim.

The video is "China Is About To Pop The AI Bubble" by Andrei Jikh, a personal-finance channel with a large audience and a habit of building each video around one market story. This one is built around two interviews — tech critic Ed Zitron on CNBC, and Palantir CEO Alex Karp — plus a stack of charts, including three posted by Michael Burry, the investor famous for shorting the 2008 housing market.

The thesis, compressed: the US stock market — and by extension a lot of ordinary retirement money — is being held up by a story. The story is that American companies will eventually collect trillions in AI profits because the world will have no choice but to buy American AI. The video argues the story is breaking in three places at once: businesses don't trust the AI vendors, the vendors' own economics don't work, and China is giving away a nearly-as-good version of the product for a tenth of the price.

It opens with a geopolitical scene-setter: a June order from the US Commerce Department that, in the video's telling, forced Anthropic — the company behind Claude — to cut off its most powerful models from foreign nationals worldwide, after which European governments began walking away from American AI vendors, with the French prime minister warning against depending on partners "capable of turning off the tap." We could not independently confirm the details of that account, so we note it as the video's framing rather than established fact — but the conclusion it is used to support is checkable, and we check it below.

II

Problem one: nobody trusts it.

The first leg of the argument comes mostly from Alex Karp, and it is about how AI is priced. When you hire a lawyer or a contractor, the price attaches to a result — the case, the kitchen. AI companies instead bill by the token: every word the model reads and writes is metered, whether the answer was brilliant or useless. Karp's point is that if these companies truly believed their product built billion-dollar businesses, they would price like partners — take a cut of the upside — instead of selling metered words. Nobody offers "pay us only if it works," he argues, because nobody can promise it works: models still confidently make things up, and nobody, including their builders, can fully explain when or why.

That leaves corporate buyers in an awkward spot. A CEO who spent $50 million on AI this year has no clean ROI number to show the board — the video's claim is that nobody does. And the fear underneath the spending is sharper: when your company's data flows through a vendor's model, the vendor is learning your business. The video cites the moment Anthropic launched a design product while partnered with Figma — whose CEO said publicly he was shocked — as the cautionary tale: you may be paying to train your own replacement.

Karp's prescription is for companies to own their AI outright — their own chips, their own data, their own model weights, nothing a vendor can switch off or learn from. The video, to its credit, flags the conflict: days before that interview, Palantir announced a partnership with Nvidia selling exactly that kind of self-hosted setup. The most-quoted critic of renting AI was, in that moment, launching the product you would buy instead.

III

Problem two: the math doesn't work.

The second leg is Ed Zitron's, and it is the strongest structural point in the video. Software became the best business model on earth because of a gap: you build the thing once, and each new customer costs you almost nothing. Costs stay flat, revenue climbs, and the space between those lines is why tech companies became the most valuable on the planet.

Generative AI breaks that gap. Every question a user asks burns electricity and wears down chips. More customers no longer means free money; it means more cost, roughly dollar for dollar. The video's analogy: it is not a software business, it is a restaurant — one that loses money on every meal and plans to fix that by serving more meals. Investors have been trained by two decades of "Amazon lost money for years too" to wait for margins to appear at scale. Zitron's claim is that with AI the margins are moving the wrong way, because each new model costs more to run than the last.

• $20.9B OpenAI's reported cash burn in 2025 alone, per the video, citing Financial Times and The Information reporting
• 7.1 GW Data-center capacity Oracle is building for essentially one customer — with its own annual report flagging the risk of not getting paid
• ~$1T/yr The US AI buildout's approximate annual pace — about 3% of the entire US economy, per the video

Then comes the part that should give anyone pause, whatever they think of AI: the financing has started to circle. The video describes Nvidia selling chips to a ring of smaller "neocloud" companies who borrow billions to buy them — and Nvidia then paying to rent its own chips back, with the whole loop reported as demand. A dealership that lends you money for the car, borrows the car back on weekends, and books it all as sales. Meanwhile, the video notes, the giants funding the buildout — Microsoft, Google, Amazon, Meta — itemize every revenue line in their earnings except AI revenue. Public companies love good news. The inference the video draws from the silence is that there isn't any yet.

IV

Problem three: the China discount.

The first two problems could still be survived if the endgame held — if, when profits finally come, American companies collect them because the world has no alternative. The third leg of the argument is that the world has an alternative, and it is priced like a loss leader.

How does China compete while spending a tenth as much? The video's answer is distillation: rather than doing the hardest research in human history from scratch, Chinese labs train their models on the outputs of the frontier models that already exist — copying the homework, compressing the result into smaller, cheaper models, and then open-sourcing them so anyone can download and run them. On this telling, every American research dollar is part donation to the Chinese AI industry.

The sharpest version of the point is not the benchmark gap but the question businesses actually ask: do I need the smartest AI in the world to answer my customer-service email? No. Ninety percent of what companies use AI for is boring work, and a model at 90% of the quality for 10% of the price wins boring work. The video's emblem for this is LongCat — a model that goes toe-to-toe with older American flagships and was built by Meituan, a Chinese food-delivery company, as a side project. If a DoorDash can do it, what exactly justifies a trillion-dollar valuation for doing it first?

V

How a pop would announce itself.

Here the video makes a genuinely useful historical point. The NASDAQ peaked in March 2000 — but the companies laying fiber-optic cable, the data centers of that era, kept spending billions well into 2001. The market did not wait for the builders to stop building. It broke when enough investors stopped believing the story. So the trigger, if it comes, will likely be something small and boring. The watchlist, as the video lays it out:

1. The first hyperscaler to blink gets rewarded

A CEO on an earnings call saying something as mild as "we are moderating the pace of our infrastructure investment" — and the stock going up. The video cites Zitron relaying a Goldman Sachs view that the first company to pull back on capex will be rewarded by the market. The tech industry copies itself; the first rewarded pullback is permission for everyone.

2. The debt window closing

The remaining buildout — allegedly around 100 gigawatts of planned data-center capacity — needs trillions in financing. When data-center debt stops being issued, or a major AI lab's funding round falls through, the video calls that "bedtime for the industry." Google's unusual $85 billion equity raise gets read as an early strain sign.

3. Credit spreads — with a warning label

Bond investors charge a premium over the risk-free rate to lend to companies; that premium — the spread — is the market's fear gauge. Today spreads sit around 2.6%, near the calmest readings ever recorded, which either means everything is fine or means nothing at all: the video's own caveat is that spreads were just as calm in early 2007, months before Bear Stearns. Spreads measure what lenders believe, not what is true.

4. Burry's three charts

Chip stocks trading at the top of their fifteen-year valuation range. AI suppliers up roughly 200% while the hyperscalers doing the actual trillion-dollar spending sit barely above zero — the market rewarding the companies receiving the money and ignoring the ones spending it, which is the market quietly saying it doubts the spenders will earn it back. And a token-price index — the price of AI itself — down about 20% from its May high, with demand shifting toward cheaper models. That last one, the video notes, is the China theory showing up in the data. It also notes Burry has been early — the polite word for wrong — before.

And the honest ending, which we want to preserve because headlines rarely do: after twenty-four minutes titled "China Is About To Pop The AI Bubble," the video's actual conclusion is that no one knows if or when it pops. These are early-warning signs, not a countdown.

VI

Our read.

Remember the disclosure in section I: we build with these tools and sell the results. With that on the table — here is where we think the argument is strong, where it is thin, and the distinction that we think does most of the work.

What holds up

The unit-economics point is real. Serving AI answers costs real money per answer, and that is a genuinely different business from classic software — anyone who has watched their own API bill knows it. The circular financing is worth taking seriously on its face: when an industry's demand is partly funded by its own supplier, history says look closer. The silence about AI revenue on hyperscaler earnings calls is a fair observation. And the "you don't need the smartest model for boring work" point is not a theory to us — it is how we run this studio. We use frontier models where judgment matters and cheap ones where it doesn't, and the cheap ones get better every quarter. The video treats that as a threat to America's AI industry. It is. It is also exactly what being a customer of a maturing technology feels like.

What's thin

Almost every voice in the video is arguing its own book — and the video, to be fair, says so. Karp's rent-nothing sermon doubled as a product launch. Zitron is tech media's most committed AI bear; his predictions deserve the same scrutiny as the boosters'. Burry has predicted several crashes that didn't come. The geopolitical opener — the export-control letter, Europe walking away — is the part we could least verify and the part doing the most emotional work. And distillation cuts both ways: a competitor whose product depends on copying your homework is a fast follower, not a leader, and fast followers inherit the ceiling, never set it.

The distinction that matters

"AI is a bubble" smuggles two different claims into one sentence. One is about asset prices: are AI-adjacent stocks priced as if a trillion dollars of future profit is certain? Quite possibly, and on that question we have no expertise and no business advising you — we are a marketing studio, not financial advisors, and nothing here is investment advice. The other claim is about the technology: is the capability itself fake, a story with nothing behind it? On that one we have direct evidence five days a week, and the answer is no. The work in front of us — sites built in days instead of months, edits made in one pass instead of an afternoon of clicking — does not care what Nvidia's multiple is.

The dot-com crash is the video's favorite analogy, and it is ours too, read to the end. The NASDAQ fell almost 80%. Pets.com died. And the fiber laid during the mania became the cheap backbone on which the actual internet economy — the one that transformed retail, media, and everything else — was built by the survivors. The bubble was real and the technology was real. Both things were true at once. If the AI trade pops, the models do not un-invent themselves. The most likely aftermath is the one the video itself documents in progress: the price of intelligence keeps falling.

VII

What it means on Main Street.

Our clients are not hyperscalers. They are restaurants, churches, contractors, and shops in the Fox Valley whose entire exposure to this question is a marketing budget and maybe a retirement account they would rather not think about. For them, we think the video boils down to three practical lessons — none of which require knowing whether Burry is right.

A pop would make your tools cheaper, not gone

The doomsday scenario for AI stocks is, strangely, a decent scenario for AI users. The video's own data shows the price of AI falling 20% during the biggest buildout in history, with demand shifting to cheaper models. A shakeout accelerates that. The risk to a small business was never that AI disappears — it is overpaying, mid-mania, for AI-flavored subscriptions that a commodity model will do for pennies next year.

Own your assets; rent as little as possible

Karp's advice to enterprises — own the thing, don't rent what can be switched off or repriced — scales down perfectly to a business whose website lives on a platform. It is the same conclusion we reached in our Squarespace article from the other direction: the sites we build are plain, fast code that we hand to you outright. No CMS subscription, no AI vendor wired into the page, nothing that stops working because a company in San Francisco changed its pricing or its terms. AI builds the asset; the asset doesn't depend on AI to keep standing.

Pay for outcomes, not tokens

The most durable idea in the video is Karp's pricing test: a confident vendor attaches price to results. That test applies to marketing at least as well as it applies to AI labs. If anyone — us included — pitches you something "AI-powered," the question is not which model it uses. It is: what result am I buying, and how will we both know if I got it? A page that loads fast and ranks. A site rebuilt before the renewal letter. An update made the day you email it in. Those survive any bubble, because they were never priced on the story.

VIII

Method & sources.

What this article is. Sections II–V are our summary of the argument made in Andrei Jikh's video, working from its full transcript, with his sourcing noted where he gives it (Financial Times and The Information reporting on OpenAI's finances; Oracle's annual report; the Artificial Analysis index; Michael Burry's published charts; Bloomberg on the token-price index). Sections VI and VII are our own view, clearly separated on purpose.

What we verified and what we didn't. The Artificial Analysis rankings are public and directionally match the video's description: US frontier models lead, Chinese open models fill much of the table at far lower prices. The dollar figures in Figures 1 and 2 are the video's numbers, presented as such — we did not independently audit them. We could not verify the opening account of the June export-control order concerning Anthropic's models or the specific European government reactions, and we have flagged that in section I rather than repeating it as fact. Where a claim is one interviewee's opinion, we have tried to keep his name attached to it.

Not investment advice. We are a marketing studio. Nothing in this article is a recommendation to buy, sell, or hold anything. Our stake in this argument is disclosed in section I: we sell work made with the tools the video says are overhyped.

1. Andrei Jikh, "China Is About To Pop The AI Bubble", YouTube — the primary source throughout.
2. Artificial Analysis — the independent model-intelligence index cited in the video.
3. Ed Zitron, Where's Your Ed At — the newsletter behind the CNBC interview excerpted in the video.
4. Alex Karp interview and Palantir–Nvidia partnership announcement, as excerpted and described in the video.
5. Human AI Marketing, Ten years of Squarespace pricing — our earlier piece on owning your web assets outright.