When Will the AI Bubble Pop? Signs to Watch in 2026

Q: "When will the AI bubble pop?"

" No one can time it precisely. The historical pattern suggests it unwinds when enough investors stop believing US firms will recoup their spending — not when spending physically stops. Watch for the first hyperscaler capex cut, widening credit spreads, and sustained declines in AI token prices as early tells."

Q: "Is China winning the AI race?"

" According to the analysis, the US still has the single smartest models, so it leads on capability and on total dollars spent. But China is winning on cost and adoption for everyday business tasks — customer service, claims processing, routine automation — where a cheaper open model within a few quality points is good enough. On that basis, China is \u0026ldquo;winning the customer.\u0026rdquo;"

Q: "What is AI distillation?"

" Distillation is training a smaller model by having it learn from the outputs of a larger, already-trained model, compressing that knowledge cheaply instead of paying billions to train from scratch. Chinese labs use it to ship low-cost open-source models, which the video argues turns every dollar of US frontier research into a partial subsidy for global competitors."

Q: "Why do businesses want to own their own AI model?"

" Executives worry that routing proprietary data through a vendor\u0026rsquo;s model exposes their \u0026ldquo;alpha\u0026rdquo; and can turn that vendor into a competitor, and that a foreign government or supplier could cut off access. Owning an open model on your own hardware keeps data private, controllable, and impossible for a third party to switch off — the same self-sovereignty logic that links AI to crypto rails."

Nobody can say exactly when the AI bubble will pop, but the data points to a trigger that arrives before companies stop spending: the moment enough investors stop believing the story. As of July 2026, the clearest early signs are a hyperscaler pulling back capital expenditure, credit spreads widening, falling AI token prices, and China’s cheap open-source models eroding the profits the whole trade is priced on.

That framing comes from a July 2026 analysis by finance creator Andrei Jikh, who stitched together interviews with tech critic Ed Zitron and Palantir CEO Alex Karp to argue the American AI story is more fragile than the stock market assumes. Below is what the argument actually says — and the specific market signals it says to watch.

Key takeaways

The bull case for US tech rests on one assumption: American companies will eventually earn trillions in AI profits because the world has no alternative.
China is that alternative. According to the video, the US is set to spend roughly $764 billion on AI in 2026 versus China’s ~$102 billion — an almost 10-to-1 gap — yet Chinese open models are within a few quality points at a fraction of the price.
The AI business model is unusual for software: costs rise roughly in line with revenue, so more customers means more losses, not more margin.
History suggests the bubble pops when belief breaks, not when spending stops — during the dot-com crash, the Nasdaq peaked in March 2000 while infrastructure spending ran into 2001.
The signals to watch: the first hyperscaler capex cut, widening credit spreads, and a sustained drop in AI token prices.

Why people ask when the AI bubble will pop

Investors ask when the AI bubble will pop because a large share of the US stock market — and by extension index funds and retirement accounts — now rests on the expectation of enormous future AI profits. Jikh’s core claim is that valuations assume “American companies are going to make trillions of dollars in profits forever because the world will be forced to use America’s technology.”

The video frames three problems with that story. First, enterprises don’t fully trust the tools. Second, the economics don’t work like normal software. Third, and most important, the world now has a cheaper option in China. Each weakens the assumption that US firms will capture the eventual payoff.

Problem one: enterprises don’t trust the AI model

The first crack is trust. In the cited interview, Palantir CEO Alex Karp argues that AI vendors price by token usage — you pay per word read and written, regardless of whether the output was useful — rather than charging for outcomes. His rhetorical question: if the value were as large as marketed, why not offer to build a customer a billion-dollar idea for a cut of the revenue, “pay us nothing unless we make you money”?

Karp’s second concern is data. When a business runs its proprietary workflows through a vendor’s model, that vendor can learn from the “alpha” that makes the business profitable. The video points to Anthropic launching a design product while it had a relationship with design company Figma as the kind of vendor-becomes-competitor dynamic executives fear. The proposed fix — echoed by Palantir’s own new offering — is ownership: run an open model on your own hardware and data so no one can see it, learn from it, or switch it off. This is the same “own your stack” logic driving monetizing open-source AI with crypto tokenomics.

Problem two: the AI business model breaks software economics

The second problem is that AI inverts what made software the best business model ever invented. Traditional software costs a lot to build once, then each additional customer is nearly free — the gap between flat costs and rising revenue is the profit. AI breaks this because every query consumes electricity and wears down chips, so serving more customers adds cost dollar-for-dollar.

The video likens it to a restaurant that loses money on every meal and whose plan is to serve more meals. It cites reporting that OpenAI burned roughly $20.9 billion in 2025, and Ed Zitron’s argument that AI margins are getting worse, not better, because each new model costs more to run than the last. That matters for the “be patient, margins will improve at scale” thesis that investors extended to Amazon for years — here, scale isn’t fixing the gap. The strain shows up across the buildout: the compute-and-energy bill behind it is the subject of our look at the AI supercycle in compute, energy and crypto.

Problem three: China is the cheaper option

The third and most decisive problem is that the world has an alternative, and it is China. According to figures cited in the video, the US is on track to spend about $764 billion on AI in 2026 (around 3% of its economy) versus China’s ~$102 billion (about 0.6% of its economy) — the US outspending China nearly 10 to 1.

China closes the gap through distillation: instead of training a frontier model from scratch for billions, a smaller model studies the outputs of an existing model and compresses the results — effectively copying the homework for a fraction of the cost — then open-sources it for free. On the Artificial Analysis Intelligence Index the video references, the top US model scores around 60 while the best Chinese open model (GLM) scores about 51, with DeepSeek, Qwen, Kimi and MiniMax filling the global middle. On a matched coding task, the video claims the American model billed $2.33 versus $0.31 for the Chinese one. You can check the live rankings at Artificial Analysis. The punchline: you cannot make back a trillion dollars selling what a competitor gives away at ~90% of the quality for ~10% of the price.

The signs the AI bubble is popping

The clearest lesson is that the bubble pops when belief breaks, not when capex stops. In the dot-com bust, the Nasdaq peaked in March 2000, yet fiber-optic buildout — that era’s data centers — continued well into 2001. So the video points to earlier, subtler triggers:

The first hyperscaler capex cut. Per Ed Zitron, Goldman Sachs has suggested the first hyperscaler to pull back spending will be rewarded by markets — which would give every other CEO permission to follow.
Credit spreads. Corporate bond spreads over the risk-free rate measure lender fear. As of mid-2026 they sit near record lows (~2.6%), but the video warns spreads were similarly calm in early 2007 before the 2008 crisis — they measure belief, not truth.
Falling token prices. Investor Michael Burry reportedly flagged that an AI token-price index was down nearly 20% from its May high, consistent with demand shifting toward cheaper models — the China effect showing up in pricing.

None of these is a guaranteed top. The video is careful to note the signal is ambiguous and that Burry has been early (often a polite word for wrong) before. But taken together they describe how a repricing could begin quietly, on an unremarkable earnings call, rather than with a dramatic crash. For a broader take on what a 2026 downturn might actually look like, see our next financial crash 2026 analysis.

Frequently asked questions

When will the AI bubble pop?

No one can time it precisely. The historical pattern suggests it unwinds when enough investors stop believing US firms will recoup their spending — not when spending physically stops. Watch for the first hyperscaler capex cut, widening credit spreads, and sustained declines in AI token prices as early tells.

Is China winning the AI race?

According to the analysis, the US still has the single smartest models, so it leads on capability and on total dollars spent. But China is winning on cost and adoption for everyday business tasks — customer service, claims processing, routine automation — where a cheaper open model within a few quality points is good enough. On that basis, China is “winning the customer.”

What is AI distillation?

Distillation is training a smaller model by having it learn from the outputs of a larger, already-trained model, compressing that knowledge cheaply instead of paying billions to train from scratch. Chinese labs use it to ship low-cost open-source models, which the video argues turns every dollar of US frontier research into a partial subsidy for global competitors.

Why do businesses want to own their own AI model?

Executives worry that routing proprietary data through a vendor’s model exposes their “alpha” and can turn that vendor into a competitor, and that a foreign government or supplier could cut off access. Owning an open model on your own hardware keeps data private, controllable, and impossible for a third party to switch off — the same self-sovereignty logic that links AI to crypto rails.