The AI Economy
Priced for Forever
America is spending more than half a trillion dollars a year building artificial intelligence. The story that justifies that spending rests on three assumptions, and in the past month all three took visible damage. Here is what the data actually shows.
Key Takeaways
- A June 12 US export order briefly shut down the world's most advanced AI models for every foreign user, and allied governments responded by moving away from US technology vendors within days.
- Leaked audited financials show OpenAI lost $20.9 billion from operations in 2025 on $13.1 billion of revenue, and the biggest builders are increasingly funding the boom with debt, equity raises, and vendor financing.
- Chinese labs now field open models at roughly 85 to 90 percent of top US benchmark scores while Chinese firms spent about 82 percent less on AI infrastructure from 2023 to 2025.
- Bond markets remain calm, with credit spreads near historic lows, but the same gauge was calm in early 2007 too.
- Nobody can time a bubble. The data suggests watching five specific signals rather than predicting a date.
The United States stock market is carrying an unusual amount of weight on a single narrative. The narrative goes like this: American companies are building the most important technology in history, the world will have no choice but to buy it from them, and the trillions being spent today will come back as profits tomorrow. Retirement accounts, index funds, and a meaningful share of US economic growth now lean on that story being true.
This report examines that story with the best available data. It was prompted by "China Is About To Pop The AI Bubble," a video essay by the finance creator Andrei Jikh1, which argued that the AI trade rests on cracked foundations. We checked the video's central claims against primary sources, regulatory filings, and reporting from Bloomberg, the Financial Times, Fortune, Semafor, and others. Most of those claims held up. Some needed context. What follows is the verified picture, the strongest case against it, and our conclusion.
Part One
The night Washington switched off its own champion
On June 12, 2026, Commerce Secretary Howard Lutnick sent a letter to Anthropic, the San Francisco company behind the Claude models. Citing national security and export control powers, the letter directed Anthropic to suspend access to its two most capable models, Claude Fable 5 and Claude Mythos 5, for every foreign national in the world2.
The order did not distinguish between China and America's closest allies. Users in France, Germany, and Japan lost access. So did Anthropic's own employees who were not US citizens. Because the company could not reliably verify citizenship across hundreds of millions of accounts in real time, the practical result was a hard global shutoff of the most advanced AI systems on the market2. The restrictions were partially eased in late June, first for a group of roughly 100 companies and federal agencies, then more broadly3. But the damage to the narrative was done, because the world had just watched Washington demonstrate that access to American AI can be revoked by letter, overnight.
Four days later, France's domestic intelligence agency, the DGSI, announced it would drop the American data firm Palantir in favor of the French supplier ChapsVision. Prime Minister Sébastien Lecornu said: "We cannot depend on the goodwill of some partners who are capable, as we've seen in recent days, of cutting off access to Anthropic's models." He pledged 655 million euros for domestic AI through 2030, pointing directly at the Anthropic restriction as evidence that dependence on US platforms carries unacceptable strategic risk4. Germany's federal domestic intelligence office had already selected ChapsVision over Palantir, and Germany's military said it would stop using Palantir products4.
This matters for the money story. If the thesis is that the world will be forced to buy American AI forever, June 2026 was the month major allies began building the exit.
Part Two
The trust problem: paying for tokens, training your replacement
Even before the export order, enterprise customers had grown uneasy about the basic bargain of the AI business. Alex Karp, the chief executive of Palantir, put the complaint bluntly in a recent interview featured in the Zeck video: "I am paying for tokens that create no value. These people are stealing the weights and alpha of my business"1.
Karp's argument deserves unpacking, and also a disclosure. The pricing point first. AI companies charge by the token, which is roughly a charge per word processed, whether the output is brilliant or useless. Karp's provocation is that a vendor truly confident in its product would price on outcomes instead: pay us nothing unless we make you money. No major lab offers that deal, partly because models still hallucinate, meaning they confidently produce false output in ways their own creators cannot fully predict or prevent.
The deeper fear is competitive. When a company routes its workflows through a frontier lab's model, its processes and its data pass through a potential future rival. That fear stopped being hypothetical in April. On April 14, Anthropic's chief product officer, Mike Krieger, resigned from the board of the design company Figma. Three days later, Anthropic launched Claude Design, a product that generates interactive prototypes and marketing materials, in direct competition with Figma6. Figma's stock slid in the weeks that followed, down nearly 19 percent in the month after the launch, the activist investor Findell Capital demanded a review of the board's ties to Anthropic, and Figma's chief executive, Dylan Field, said of Anthropic: "They were not consistently candid in their communications"7.
"I want to own the GPUs. I want to own my data. I want to own the model. I want to control the alpha. Why would they get access to my data?" Alex Karp, CEO of Palantir, describing the shift toward self-hosted AI among technical enterprises1
Now the disclosure: Karp is not a neutral observer. Palantir has been losing European government contracts all year, and two days before that interview it announced a partnership with Nvidia to sell open models in sovereign, customer-controlled environments. His critique of rented AI doubles as a pitch for his own product. That does not make the critique wrong. It does mean the strongest evidence for the trust problem is not what Karp says, but what governments and enterprises are actually doing, and what they are doing is moving toward AI they can own and switch off themselves.
Part Three
The arithmetic problem: software margins without the software
Software became the world's most valuable industry because of a simple property: you build the product once, and each additional customer costs almost nothing to serve. Generative AI does not have that property. Every query burns electricity and wears down expensive chips. Costs scale with usage, which means costs scale with success.
The numbers at the industry's most important company illustrate the problem. In June, leaked audited financial statements, obtained by the critic Ed Zitron and verified by the Financial Times, showed that OpenAI lost $20.9 billion from operations in 2025, even as revenue roughly tripled to $13.1 billion5. Total costs reached about $34 billion. Research and development spending alone, at $19.2 billion, exceeded the company's entire revenue5.
OpenAI in 2025: the gap is the business model
Revenue versus total costs and expenses, in billions of US dollars, per leaked audited financials verified by the Financial Times
Operating loss: $20.9B
View data as table
| Line item, 2025 | Amount |
|---|---|
| Revenue | $13.1 billion |
| Total costs and expenses | $34.0 billion |
| Of which R&D | $19.2 billion |
| Operating loss | $20.9 billion |
The widely reported $38.5 billion net loss is the GAAP figure and includes large non-cash items. The $20.9 billion operating loss is the cleaner read on cash economics. Sources: Fortune; Where's Your Ed At5.
Investors have been trained by two decades of Amazon comparisons to be patient with losses, on the theory that margins improve at scale. The uncomfortable pattern in AI is that each new generation of models costs more to train and run than the last, so the two lines, revenue and cost, keep climbing together. As Zitron put it in a CNBC appearance: "Their costs increase linearly with their revenues"1.
Meanwhile the financing structure of the buildout is getting more creative, which is rarely a comforting sign late in a cycle. Three examples from recent filings and reporting:
Oracle bet the company on one customer. Oracle's remaining performance obligations, essentially its contracted future revenue, surged 359 percent year over year to roughly $455 billion, growth that D.A. Davidson analyst Gil Luria says came "almost entirely from OpenAI," under a reported five year, $300 billion cloud deal that ranks among the largest technology contracts in corporate history9. Oracle's own annual report warns that its cloud infrastructure growth is "dependent on our ability to retain a limited number of key customers"9. That key customer, recall, just lost $20.9 billion in a year.
Nvidia is financing its own demand. The chipmaker sells GPUs to a cluster of smaller cloud providers known as neoclouds, takes equity stakes in some of them, and has agreed to backstop their economics, including a commitment to purchase up to $6.3 billion of unsold compute capacity from CoreWeave through 2032, a roughly $1.5 billion arrangement to lease GPUs back from Lambda, and a guarantee of up to $860 million in data center lease obligations for a partner company10,11. Vendor financed demand is real revenue on paper, but it blurs the question every investor should be asking, which is how much of this demand exists without the vendor's own money behind it.
Even the richest companies are passing the hat. Alphabet, one of the most cash generative businesses in history, upsized an equity raise to $84.75 billion in June, including a $10 billion placement with Berkshire Hathaway, to help fund capital spending of roughly $185 billion this year12,13. Across the big five cloud companies, 2026 capital commitments total roughly $660 to $690 billion, nearly double 2025, with Wall Street projections crossing $1 trillion in 20278.
None of this proves a bubble. It proves scale, urgency, and rising financial engineering, all funding a product whose flagship maker loses about $1.60 for every dollar it earns.
Part Four
The China discount
The entire buildout is underwritten by one assumption: when profits finally arrive, American companies will collect them, because the world has no alternative. The strongest data based challenge to the AI trade is that the alternative now exists, and it is priced at a discount that no premium product has ever survived unchanged.
The spending gap
Cumulative AI-related capital spending by major cloud and internet firms, 2023 to 2025, in billions of US dollars
View data as table
| Region | 2023 to 2025 AI capex | Share |
|---|---|---|
| US hyperscalers (AWS, Microsoft, Google, Meta) | $694 billion | 100% |
| Chinese hyperscalers (Alibaba, Baidu, Tencent, ByteDance) | $124 billion | 18% |
China's top models now benchmark at roughly 90 percent of US performance despite the 82 percent spending gap. Source: Investing.com, citing industry capex data15.
The performance gap has narrowed faster than the spending gap. In June, the Beijing startup Z.ai released GLM-5.2, an open weights model anyone can download and run. It scored 51 on the Artificial Analysis Intelligence Index, the most cited independent ranking of AI models, placing fourth in the world behind only three American proprietary systems, which cluster around 6016. Chinese open models from DeepSeek, Qwen, Kimi, and MiniMax fill the middle of the global rankings.
Frontier scores: closed American lead, open Chinese pursuit
Artificial Analysis Intelligence Index, selected models, June 2026. Higher is more capable.
View data as table
| Model | Index score | Access |
|---|---|---|
| Top US proprietary model | 60 | Paid API only |
| GLM-5.2 (Z.ai, China) | 51 | Open weights, downloadable |
In one developer test cited in the source video, the same coding task cost $2.33 on a top US model and $0.31 on GLM, about 7 times cheaper, with both finishing in similar time1. Index scores: Artificial Analysis via Simon Willison and BenchLM16.
Then there is the fact that best captures the moment. On June 30, the highest profile new open model in China was released not by a research lab but by Meituan, a food delivery company. Its LongCat-2.0 is a 1.6 trillion parameter system, reportedly the first at that scale trained entirely on Chinese chips, and it edges out GPT-5.5 on the SWE-bench Pro coding benchmark, 59.5 to 58.617. When the Chinese equivalent of DoorDash can ship a near frontier model as a side project and give it away, the pricing power of a trillion dollar American AI complex is a fair question, not a fringe one.
How does China do this while spending a fraction of the money? Part of the answer is a technique called distillation: training a new model on the outputs of existing frontier models, capturing much of their capability at a fraction of the cost. Critics call it copying someone else's homework. Either way, the economics are brutal for the leader, because every dollar spent pushing the frontier also lowers the cost of following it.
Part Five
What a pop would actually look like
The dot com precedent contains a detail most people forget. 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 spending to stop. It broke when enough investors stopped believing the story the spending was based on. Bubbles end on narrative, not on capex schedules.
So the signals worth watching are the ones that measure belief. One of them started flashing this month. The Silicon Data LLM Token Expenditure Index, which tracks what buyers actually pay for AI usage, has fallen almost 20 percent from its May high after nearly doubling since its December launch14. Bloomberg's read: either demand is rotating to cheaper models, or buyers are refusing to pay more. The first explanation is the China discount showing up in market data. The second is worse.
The price of intelligence is falling at the wrong moment
Silicon Data LLM Token Expenditure Index, indexed to 100 at December 2025 launch. Illustrative recreation based on Bloomberg's reported figures, not the licensed series.
View data as table
| Month | Index (approx.) |
|---|---|
| December 2025 | 100 |
| January 2026 | 118 |
| February 2026 | 135 |
| March 2026 | 158 |
| April 2026 | 181 |
| May 2026 (peak) | 198 |
| June 2026 | 172 |
| Early July 2026 | 160 |
Monthly values are approximations drawn to match the publicly reported shape of the series: near doubling from December launch to a May peak, then a decline of almost 20 percent. For the actual index, see Silicon Data18. Reporting: Bloomberg, July 3, 202614.
A second gauge is the bond market, and right now it is asleep. US high yield credit spreads sit near 289 basis points, in the bottom decile of history, and investment grade spreads recently touched their tightest levels since 199719. On its face, that says lenders see almost no risk in the corporate landscape, even as AI related investment grade issuance hit records this year19. The caution: spreads measure what lenders believe, not what is true. In early 2007, months before Bear Stearns collapsed, spreads were roughly this calm. Calm credit markets did not predict 2008, and they are not proof of safety now.
The bearish celebrity data point comes from Michael Burry, the investor who famously shorted the housing bubble. He has disclosed bearish positions against Nvidia, Micron, and the broader semiconductor index, and has circulated charts showing chip stocks at the top of their 15 year valuation range while the companies doing the heaviest AI spending badly lag the companies collecting it1. Burry's record demands the standard caveat: he has called many crashes that never came, and being early in markets is a synonym for being wrong. Bloomberg itself describes the token index signal as ambiguous14.
The Watchlist: five signals that would matter
- A hyperscaler cuts capex and gets rewarded. Ed Zitron reports Goldman Sachs analysts have suggested the first big cloud company to pull back on spending will be rewarded by markets1. If Wall Street cheers a cut, every other CEO gets permission to follow.
- Data center debt stops getting issued. Roughly 100 gigawatts of capacity is under construction or in planning. If lenders balk, the buildout stalls regardless of anyone's beliefs.
- Credit spreads widen from historic tights. A move in high yield spreads from under 3 percent toward 4 or 5 percent would signal lenders repricing AI risk19.
- Token spending keeps falling. A continued slide in the Silicon Data index would confirm eroding pricing power rather than a one time rotation to cheaper models14.
- A major AI financing falls through. A failed raise at a frontier lab, or trouble in the OpenAI and Oracle relationship, would test the market's patience directly9.
Part Six
The other side of the trade
An honest version of this story has to include the data pointing the other way, and there is more of it than the bears usually admit.
First, the revenue is real and growing fast. Nvidia's most recent quarter brought in $57 billion of revenue, up 62 percent year over year, and enterprise spending on AI reached an estimated $37 billion in 2025, up from under $2 billion just two years earlier20. Whatever this is, it is not the dot com era's revenue free speculation.
Second, most of the buildout is still funded by operating cash flow rather than debt. Analysts note that four of the five biggest spenders could cover their capex from cash generation alone, with Oracle the leveraged exception20. Companies deliberately running negative free cash flow because they believe in the return is a risk, but it is a different and smaller risk than a system built on borrowed money.
Third, falling prices can grow markets. Economists call it the Jevons paradox: when something useful gets cheaper, total consumption often rises. Token prices have collapsed more than 90 percent since 2023, yet total spending on AI roughly doubled over the past year because cheap intelligence unlocked new uses14. On this reading, China's cheap models do not destroy the market. They expand it, and the biggest, most demanding workloads still land on premium American systems.
Fourth, "good enough" has a ceiling. Enterprises processing insurance claims may not need frontier intelligence, but the frontier keeps finding new tasks that only the best models can do, and those tasks command premium prices. The US lead at the top of the rankings is intact.
Conclusion
Our read of the data
The Bottom Line
The evidence does not support the comfortable version of the AI story, the one where American companies simply collect trillions because the world has no alternative. The alternative exists, it is improving quickly, and in June the United States government itself handed every foreign buyer a reason to want it. At the same time, the evidence does not support a confident crash call. Revenues are large and accelerating, most of the spending is funded by real cash flows, and the historical record shows that people who put dates on bubbles are almost always wrong, including the ones who turn out to be right about the bubble.
Our conclusion is narrower and, we think, more useful. The AI trade has quietly changed from a story about unlimited demand into a bet on pricing power, and the early data on pricing power is going the wrong way: token spending down 20 percent from its peak, open models at 85 percent of frontier scores for a fraction of the price, allies procuring away from US vendors, and the industry's flagship losing $1.60 on every dollar of revenue. That is not a popped bubble. It is a story running out of margin for error, held up by a bond market that has been this calm before, in 2007. Watch the five signals above rather than anyone's predictions, ours included.
Sources and citations
- Andrei Jikh, "China Is About To Pop The AI Bubble," video essay, YouTube (@AndreiJikh), July 2026, including interview clips of Ed Zitron (CNBC) and Alex Karp. youtube.com/watch?v=siazPdsZHuI
- Bloomberg, "Lutnick's Letter to Anthropic Warned of Curbs on Top AI Models," June 16, 2026. bloomberg.com; see also Forbes, "Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order," June 16, 2026. forbes.com
- CNN Business, "White House lifts export control on Anthropic that froze its most advanced models," June 30, 2026. cnn.com; CNBC, "Trump admin allows Anthropic to release Mythos AI model to some companies, government agencies," June 26, 2026. cnbc.com
- Semafor, "France's spy agency drops Palantir," June 16, 2026. semafor.com; SOFX, "France Cuts DGSI's Palantir Contract Six Months After Renewal, Selects ChapsVision," June 2026. sofx.com; EU Today, "France Moves Away From Palantir in Test of Europe's Digital Sovereignty," June 2026. eutoday.net
- Fortune, "OpenAI's financials have leaked, showing $21 billion in losses against $13 billion in revenue," June 16, 2026. fortune.com; Ed Zitron, "OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion," Where's Your Ed At, 2026. wheresyoured.at
- TechCrunch, "Anthropic CPO leaves Figma's board after reports he will offer a competing product," April 16, 2026. techcrunch.com
- Upstarts Media, "How A Board Departure And Product Launch Sparked An Anthropic And Figma Feud," 2026. upstartsmedia.com; Yahoo Finance, "Figma (FIG) Faces Activist Push Over Anthropic Tie As CEO Defends AI Plan," June 18, 2026. finance.yahoo.com
- Futurum Group, "AI Capex 2026: The $690B Infrastructure Sprint," 2026. futurumgroup.com; Goldman Sachs Insights, "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out," 2026. goldmansachs.com
- Silicon Report, "Oracle's AI Bet Hinges on a Handful of Large Customers," 2026. siliconreport.com; TechBuzz, "Oracle Stock Tumbles 7% as Analysts Flag OpenAI Risk," 2026. techbuzz.ai; Eagle Point Capital, "AI Skepticism and Oracle's Big Risk," 2026. eaglepointcapital.substack.com
- Forbes, "Why The Neocloud Gold Rush Is Now Vendor-Financed," July 3, 2026. forbes.com
- Data Center Dynamics, "Nvidia backs $860m lease obligations of partner data center," 2026. datacenterdynamics.com
- Bloomberg, "Alphabet Upsizes Equity Offering to $85 Billion for AI Spending," June 3, 2026. bloomberg.com
- CNBC, "Alphabet seeking $85 billion with stock facing 4-week losing streak," June 5, 2026. cnbc.com
- Bloomberg, "AI Token Prices Drop, Raising Questions on Sector's Pricing Power and Growth," July 3, 2026. bloomberg.com
- Investing.com, "China's AI models achieve 90% of U.S. performance with fraction of capex," 2026. investing.com
- Simon Willison, "GLM-5.2 is probably the most powerful text-only open weights LLM," June 17, 2026. simonwillison.net; BenchLM, "Artificial Analysis Intelligence Index Benchmark 2026." benchlm.ai
- South China Morning Post, "China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0," June 30, 2026. scmp.com; VentureBeat, "Chinese food delivery app Meituan's open source AI model LongCat rivals GPT-5," 2026. venturebeat.com
- Silicon Data, "LLM Token Expenditure Index (SDLLMTK)." silicondata.com
- AllianceBernstein, "2026 Credit Outlook: Growing Divergence Amid AI's Big Build-Out," 2026. alliancebernstein.com; Man Group, "H2 2026 Credit Outlook: The AI Buildout, Boom or Bubble?" 2026. man.com; FRED, ICE BofA US High Yield Index Option-Adjusted Spread. fred.stlouisfed.org
- Investing in AI, "What To Watch in 2026 To Evaluate The AI Bubble," 2026. investinginai.substack.com; Guinness Global Investors, "Are we in an AI bubble?" 2026. guinnessgi.com
Disclosure Statement
No financial, investment, legal, or tax advice
This article is published for general informational and educational purposes only. Nothing in it constitutes financial, investment, legal, accounting, or tax advice, and it must not be relied upon as such. No content herein is a recommendation, solicitation, or offer to buy, sell, or hold any security, digital asset, or financial instrument. Investing involves substantial risk, including the possible loss of principal. Past performance does not guarantee future results. Before making any financial decision, consult a licensed financial advisor, attorney, or other qualified professional who can evaluate your individual circumstances. The publisher and any contributors accept no liability for any loss or damage, direct or indirect, arising from the use of, or reliance on, the information presented here.
Forward-looking statements and opinion
Portions of this article discuss possible future events, including the possibility of a market correction. These are inherently speculative, constitute opinion and commentary rather than statements of fact or predictions, and may prove incorrect. No one, including the publisher, can predict the timing or occurrence of market events.
Accuracy, sources, and third-party content
Facts and figures were compiled from publicly available third-party reporting and filings, cited in the Sources section, and reflect information available as of July 8, 2026. We do not guarantee the accuracy, completeness, or timeliness of third-party information, and figures such as leaked financial statements are attributed to the outlets that reported them. Certain quotations are drawn from interviews as excerpted in the cited video essay. Some charts are original illustrative recreations based on figures described in cited reporting; they are approximations for editorial purposes and are not reproductions of any proprietary index or dataset. If you believe any statement is inaccurate, contact us and we will review and correct it promptly.
Use of artificial intelligence
This article was researched, drafted, and produced using artificial intelligence (Claude, developed by Anthropic), operating under human editorial direction. A human editor selected the topic, directed the structure, sourcing standards, and design, and reviewed the output before publication. Because Anthropic and its products are discussed in this article, readers should know the drafting model is itself an Anthropic system; all statements about Anthropic are drawn exclusively from the cited third-party reporting, and no non-public information was used. Illustrative images in this article were generated with Google Flow, an AI image generation tool; they are artistic renderings and do not depict real people, places, products, or events.
Ownership disclosure and conflicts of interest
References to companies, products, and trademarks, including Anthropic, OpenAI, Nvidia, Oracle, Palantir, Figma, Alphabet, Microsoft, Amazon, Meta, Meituan, Z.ai, and others, are for identification and commentary only. All trademarks remain the property of their respective owners, and no affiliation, sponsorship, or endorsement is claimed or implied.
Transparency: the publisher personally owns shares of some of the publicly traded companies mentioned in this article, held as part of an ordinary, long-term investment portfolio. No position was established, increased, or reduced in anticipation of this article's publication, and the publisher has no intention of trading around it. This article is not intended to, and should not be used to, promote, inflate, or depress the price of any security, and the analysis and conclusions were not influenced by these holdings. The article presents both bullish and bearish evidence precisely so readers can judge for themselves.
The publisher received no compensation from any company mentioned. This article contains no paid placements and no affiliate links.
External links
Links to third-party websites are provided for reference. We are not responsible for the content, accuracy, or availability of external sites, and linking does not imply endorsement. Some cited articles may sit behind paywalls or become unavailable over time.