China Is Giving AI Away. America Bet the Economy on Selling It. A $589 billion dress rehearsal, and the bet it threatened.

On January 27, 2025, Nvidia lost about 589 billion dollars in a single trading day. Not over a quarter. In one day. It was the largest one-day loss of market value any company had ever recorded, and it happened because a Chinese lab most Americans had never heard of released a model called DeepSeek R1 that did roughly what OpenAI's best could do, except it had reportedly cost under 6 million dollars to train and was free to download. The Nasdaq fell 3.1 percent. By the time the dust settled, something like a trillion dollars had evaporated from US tech.

Then it came back. Within weeks Nvidia had recovered most of the loss, Wall Street talked itself back into the story, and Wedbush's Dan Ives summed up the consensus that saved the bulls: no Global 2000 company, he argued, was going to build its AI infrastructure on a Chinese startup. For a year, he was right.

That day was a dress rehearsal. The thing worth paying attention to now is that the play has not changed, only the production values. The models that spooked the market in early 2025 have gotten cheaper, better, and far more numerous. And the bet the rehearsal threatened is now a much larger share of the American economy than it was then.

The case has four parts. None of them is crazy.

What rattled people in January 2025 was the price, not the achievement itself. If a frontier-class model can be trained for a fraction of what everyone assumed, and then handed out for anyone to run on their own machines, the core financial premise of the American AI build-out starts to wobble. That premise is roughly this: companies will pay enormous, recurring sums to access closed models from OpenAI, Anthropic, and Google, and those payments will eventually justify the hundreds of billions being spent on chips and data centers to serve them.

Free, capable, self-hostable models poke a hole in it. If the model is good enough and costs a fraction to run, why sign a multimillion-dollar annual contract with a closed lab? And the moment a critical mass of large buyers starts answering that question with "we won't," the market has to ask whether the valuations built on those contracts make any sense.

This is no longer a single DeepSeek story. By spring 2026, eight of the top ten Chinese models were open-weight, released under permissive Apache 2.0 or MIT licenses that let any company download them, fine-tune them, and run them commercially. Alibaba's Qwen family passed Meta's Llama in cumulative downloads on Hugging Face, with more than 170,000 derivative models built on it. Chinese models went from roughly 1 percent of global open-model downloads in late 2024 to about 30 percent a year later.

The cost gap is the headline. Chinese open models routinely run at a fraction of the price of US frontier APIs, and several of them are seriously competitive. Models from Zhipu (GLM), Moonshot (Kimi), and DeepSeek have each, at various points this year, beaten or matched GPT and Claude on specific coding benchmarks. Zhipu trained a frontier model entirely on domestic Huawei chips, with no Nvidia hardware at all. The gap between the best American model and the best Chinese one has narrowed from months to, on some tasks, days.

Adoption is the part the popular thesis gets slightly wrong, and the correction matters. It is not that a third of big companies have torn up their OpenAI contracts. The reality is messier and, for the build-out, arguably worse. Surveys through early 2026 suggest roughly half of enterprises still use closed commercial models only, about a third now mix commercial and open source, and roughly a fifth have gone fully open.

One enterprise forecast had open-weight, self-hosted models rising from about a quarter of production AI workloads at the start of 2026 to around 40 percent by the third quarter. The default posture now is multi-model: most large companies run several model families and route each job to whichever is cheapest for the task. Cheap open weights do not need to win outright. They just need to cap what anyone can charge for the closed alternative.

“

A ceiling on pricing is enough to break a valuation built on the assumption of no ceiling.

The word "dumping" is doing a lot of work, and it deserves an honest look. Dumping has a specific meaning: selling below cost, often underwritten by the state, to capture a market and kneecap competitors. People reach for it here because the pattern rhymes with what China did in steel, solar panels, and electric vehicles. Massive state support, cheap and often coal-fired power, and products pushed onto the world market at prices nobody else can match. Steve Forbes has called it exactly that, and a Stanford analysis credits the Chinese government's backing as a substantial factor in the open-weight surge.

Whether you call it dumping or competition, the effect on pricing is identical, and that effect is what the market has to swallow.

If AI were a contained sector, a repricing would sting some rich shareholders and we would move on. It is not contained. By late 2025 the ten largest companies in the S&P 500 made up around 40 percent of the entire index, roughly double their share a decade earlier, and those names are overwhelmingly the AI trade: the chipmaker, the hyperscalers, and the supply chain of power and hardware feeding them. If you own a plain index fund, you own a concentrated bet on AI whether you meant to or not.

The scarier number is what AI spending has done to growth. Here the original thesis is directionally right and worth stating carefully. AI-related capital spending, mostly data centers, has been carrying an outsized share of US GDP growth. Barclays estimated it added about a full percentage point to growth in the first half of 2025, when total growth was only 1.6 percent, meaning the data-center build-out alone accounted for roughly half of it. Renaissance Macro's Neil Dutta noted that in a couple of recent quarters, AI capex contributed more to growth than all consumer spending combined. One estimate cited in a congressional letter on bubble risk put non-data-center GDP growth in a recent stretch at around 0.1 percent. BCA Research's chief strategist said flatly that without the AI boom, it is plausible the economy would already be in recession.

So the chain is real. A large share of the index depends on the AI trade. A large share of recent growth depends on the spending behind it. And the spending depends on a belief that the models will eventually be worth paying for. Cheap open weights attack that last link directly. One note of discipline, though: this is mostly a 2025 phenomenon, not a steady feature of "the last few years." Read it as a sudden and recent dependence, which if anything makes it more fragile, not less.

There is a second crack that has nothing to do with China, and in the near term it may matter more. A widely cited MIT study found that 95 percent of organizations investing in generative AI reported no measurable return, against 30 to 40 billion dollars spent. The sample was small, around 52 organizations, so do not treat it as scripture, but the broader pattern shows up everywhere: most enterprises are stuck in pilots and cannot point to revenue. Big Tech is on track to spend somewhere around 650 to 700 billion dollars on AI infrastructure in 2026 to serve demand that, so far, mostly is not paying for itself.

Then there is the money going in circles. Nvidia committed roughly 100 billion dollars to OpenAI, which turned around and committed to buying Nvidia chips. Nvidia took a stake in the cloud provider CoreWeave and agreed to buy back its unused capacity. The asset manager GMO described arrangements like these as reminiscent of the circular financing of the dot-com era, where money shuffled between a small group of players manufactures the appearance of demand.

Chips depreciate fast. The contracts assume they will not. You can see why people are nervous.

Nobody has pressed that depreciation point harder than Michael Burry. If the name does not ring a bell, the movie will: he is the investor played by Christian Bale in The Big Short, the one who saw the subprime mortgage collapse coming in 2007 and bet against it while half of Wall Street told him he had lost his mind. In late 2025 he resurfaced through his fund, Scion Asset Management, with what people promptly nicknamed "Big Short 2.0." His regulatory filings revealed put options worth roughly a billion dollars against Nvidia and Palantir, the two cleanest proxies for the AI trade, one in chips and one in software.

His most interesting argument is not the obvious "valuations are too high," though he makes that case too. It is an accounting one, and it lands directly on the depreciation problem. The hyperscalers, meaning Meta, Google, Microsoft, Amazon, and Oracle, are writing down their AI hardware over five or six years. But the Nvidia chips inside those data centers run on a two-to-three-year product cycle, after which a faster generation shows up and the old silicon is worth far less.

Stretching the assumed lifespan of an asset that ages out in two or three years shrinks the annual depreciation charge, and a smaller charge makes profits look bigger than they are. Burry called this one of the more common frauds of the modern era, and estimated the maneuver would understate depreciation across the big spenders by about 176 billion dollars between 2026 and 2028. By his math, that leaves Oracle on track to overstate earnings by roughly 27 percent and Meta by roughly 21 percent by 2028.

Plenty of people think he is wrong, or at least early. Palantir's CEO called the bet "super weird." Others argued the depreciation question is more a quirk of accounting rules than a fraud, and that older chips keep doing useful work on lighter jobs long after they leave the training cluster, so a longer paper life is defensible. They are not obviously wrong. But the reason to take Burry seriously is not a perfect record on timing, because he does not have one. It is that he reads the footnotes, and the footnotes are where the last crisis was hiding while everyone else watched the headline numbers. When the person with that particular résumé starts circling the exact depreciation schedules holding up reported AI profits, it is worth at least checking his math.

Because the bears have been wrong for two years running, and the bulls have reasons. The most important one is counterintuitive: cheaper AI might grow the market rather than shrink it. This is the Jevons paradox, the old observation that when something gets cheaper to use, total consumption tends to rise rather than fall. Cheaper coal did not reduce coal demand. Cheaper inference may simply mean far more of it, which is good for the people selling the picks and shovels even if the margin on any single model collapses. Capex estimates have come in too low, not too high, for two straight years. And the DeepSeek rehearsal cuts both ways: it proved the mechanism is real, but it also proved the market can absorb the shock and keep climbing.

That is the genuine uncertainty. The thesis is not wrong. It is just not yet proven, and it has been wrong before.

There are three different audiences for that question, and they need three different answers.

For the country, the live fight is over export controls and whether to compete in open models or wall them off. The enforcement hawks, including Anthropic's Dario Amodei, argue that DeepSeek strengthens rather than weakens the case for controls, and a recent Justice Department indictment over 2.5 billion dollars in smuggled AI servers suggests the controls bite hard enough that people break the law to get around them. The opposing camp, including voices at Brookings and Chatham House, argues that controls mostly push China to innovate around the constraint, which is more or less what happened with DeepSeek itself. There is wide agreement on one point: trying to ban open-source models outright, as some in Congress have floated, is both impractical and self-defeating, and the smarter move is for the US to lead in open weights rather than cede that ground. A trusted American open model in the world's hands is a more durable answer than hoping the Chinese ones disappear.

For the companies, the uncomfortable truth is that selling access to a raw model is becoming a commodity business, and commodities do not support trillion-dollar valuations. The defensible position sits everywhere except the model itself: the products built on top of it, the proprietary data, the distribution, the workflows a customer cannot easily rip out. The labs that survive a price collapse will be the ones that turned a model into a business before the price collapsed.

For investors, the practical step is the boring one. If a plain S&P 500 fund now behaves like a leveraged bet on seven or eight names, then diversifying away from that concentration is not pessimism about AI, it is basic risk management. Equal-weight versions of the index exist for precisely this reason, and strategists from Ed Yardeni down have spent the past year pointing at the other 493 companies that have largely sat out a rally happening above their heads.

The rehearsal in January 2025 lasted a few weeks. Whether the real thing is as forgiving is the question worth losing a little sleep over.

One caveat worth ending on, because it is the part the headlines skip. Open weights are only cheap if you can serve them, and serving them efficiently, reliably, and fast at scale is hard, harder than the download button suggests. Inference engines like Ollama and vLLM carry a lot of the load, but they do not make the hard calls for you. Sizing the KV cache against the VRAM you have on hand, loading weights efficiently for slimmed-down and quantized variants, orchestrating several models at once, and doing all of it without an infrastructure bill that quietly erases the savings you went open to capture. Get that last part wrong and you have just swapped a foundational-model invoice for a cloud one.

• CNN Business - "A shocking Chinese AI advancement called DeepSeek is sending US stocks plunging."
• NBC News - "Tech stocks fall as China's DeepSeek sparks U.S. worries about the AI race."

• Gizmochina - "Why U.S. Startups Are Dumping Western AI for China's Open-Source Models."
• Stanford HAI / DigiChina - "Beyond DeepSeek: China's Diverse Open-Weight AI Ecosystem and Its Policy Implications."

• Steve Forbes, Fox News opinion - "The AI Cold War has begun and America cannot afford to lose."

• Lucidworks - "Enterprise AI in 2026: Adoption Trends, Gaps & Strategic Insights."

• The Motley Fool - "The Magnificent Seven's Market Cap vs. the S&P 500."
• Benzinga - "The S&P 500 Equal Weight Vs. Market Cap Weight Debate" (RBC top-10 weight data).

• Data Centre Magazine - "What Risks the US Economy Faces if AI Data Centre Boom Slows" (Barclays and BCA Research).
• Data Center Dynamics - "As AI capex leads US economy..." (Renaissance Macro / Neil Dutta).
• Goldman Sachs - "Why AI Companies May Invest More than $500 Billion in 2026."

• Yale Insights - "This Is How the AI Bubble Bursts" (the MIT 95% finding).
• NBC News - "The AI boom's reliance on circular deals is raising fears of a bubble."

• CNBC - "'Big Short' investor Michael Burry accuses AI hyperscalers of artificially boosting earnings."
• Fortune (via AOL) - "'Big Short' investor Michael Burry follows up cryptic AI bubble warning with bearish stock activity on Nvidia and Palantir."
• Benzinga - "Michael Burry Hints At Closing His Shorts As Palantir, Nvidia Sink" (early-2026 pullback).

• Dario Amodei - "On DeepSeek and Export Controls."
• Brookings - "DeepSeek shows the limits of US export controls on AI chips."
• Just Security - "Export Controls on Open-Source Models Will Not Win the AI Race."