A Trillion-Dollar Story the Market Is Still Betting On
The US stock market, and by extension a large share of ordinary retirement accounts and index funds, currently rests heavily on one specific story: that American AI companies will eventually generate trillions of dollars in durable profit because the rest of the world will have no real alternative to their technology. It’s a story large enough that some analysts now argue the current AI infrastructure buildout is, as a share of the US economy, considerably larger than the telecom and dot-com buildout of the late 1990s.
That story is increasingly being tested from three directions at once: growing distrust of AI vendors among the businesses paying for their tools, an AI business model that behaves nothing like traditional software economics, and a lower-cost Chinese AI ecosystem offering a real alternative to the “no other choice” assumption the whole trade depends on. For a broader look at how this AI-driven rally compares directly to the dot-com crash and what it’s meant for the S&P 500, see our related breakdown: AI Stock Bubble 2026: S&P 500 and the Dot-Com Crash Comparison.
Problem One: Businesses Are Losing Trust in Their AI Vendors
A recurring theme among enterprise leaders, according to commentary from Palantir CEO Alex Karp, is a basic pricing objection: if AI companies were as confident in the value of their technology as their marketing suggests, they’d price it around outcomes rather than usage. Instead, most AI companies charge per token, essentially per word processed or generated, regardless of whether the output was actually useful. That pricing structure stands in contrast to how most professional services work: a lawyer or contractor is typically paid in relation to a specific result, not simply for hours of activity regardless of outcome.
Part of the reason outcome-based pricing isn’t realistic yet is a well-documented technical limitation: large language models can “hallucinate,” confidently generating incorrect information, and neither outside users nor the companies that built these systems have fully solved when or why it happens.
Claim: Enterprises are also increasingly wary of a deeper structural risk: that the AI vendors they pay end up learning from their proprietary data and becoming a future competitor.
Evidence: This concern isn’t purely theoretical. Design company Figma’s CEO has publicly described being caught off guard when Anthropic, a company Figma had a working relationship with, launched its own competing design product, Claude Design.
Interpretation: For any business whose competitive advantage depends on proprietary processes or data, sending that information through a third-party AI system carries a real risk that the vendor could, intentionally or not, absorb insights that end up commercially competing with the client.
Limitation/counterpoint: Not every business relationship with an AI vendor carries this risk equally, and vendors generally have contractual and reputational incentives not to weaponize client data directly. Still, the concern has been significant enough that some of the most technically sophisticated companies are now moving toward running their own AI models entirely in-house, on owned infrastructure and proprietary data, specifically to avoid this exposure. It’s worth noting that companies promoting this “own your own AI” approach, including Palantir, have a direct commercial interest in that shift, which doesn’t make the underlying trust concern invalid, but is a relevant conflict of interest worth factoring in when evaluating who’s making the argument.
Problem Two: The AI Business Model Doesn’t Behave Like Traditional Software
Traditional software has historically been one of the most profitable business models ever built, because the cost of producing one additional copy of a product is essentially zero. Once a company builds a piece of software, selling it to another customer costs almost nothing extra, which is why software companies have dominated the most valuable companies in the world for decades.
Generative AI breaks that pattern. Every single query costs real money in electricity and computing hardware wear, meaning additional customers add cost roughly in proportion to revenue rather than arriving as nearly free incremental profit. That’s a fundamentally different economic structure, closer to a restaurant that has to buy fresh ingredients for every meal served, except one that currently loses money on every meal it serves.
The scale of those losses is significant: OpenAI reportedly burned through more than $20 billion in 2025 alone, according to audited financials reported by the Financial Times. Unlike earlier unprofitable tech growth stories, where costs typically flattened out as the customer base scaled (a pattern often cited from Amazon’s early years), AI companies’ costs have continued rising roughly in line with revenue rather than leveling off, since each new, more capable model tends to cost more to run than the one before it.
This financial pressure extends well beyond OpenAI itself. Oracle has reportedly committed to building over 7 gigawatts of data center capacity largely tied to a single customer relationship, an exposure large enough that Oracle’s own annual report reportedly flagged the risk of not being paid. There are also concerns about circular financing arrangements in the AI supply chain, where major chipmakers have both sold hardware to smaller cloud infrastructure companies and then rented back capacity from those same companies, a structure some analysts argue can make underlying demand appear stronger than it organically is.
Adding to the opacity, several of the largest technology companies investing heavily in AI infrastructure, including Microsoft, Google, Meta, and Amazon, disclose granular revenue figures for most of their business lines but have generally not broken out specific AI revenue separately, making it difficult for outside investors to independently verify the return on that spending.
Problem Three: The World Increasingly Has a Cheaper Alternative
The assumption underpinning much of the AI trade is that even if profits take time to materialize, American companies will eventually capture them because the rest of the world has no real substitute. China’s AI ecosystem complicates that assumption directly.
Claim: China is competing effectively in AI despite spending a fraction of what the US spends.
Evidence: Estimates place US AI infrastructure spending at roughly $764 billion this year, climbing toward $1 trillion next year, representing around 3% of the entire US economy. China’s comparable spending is estimated at roughly $102 billion this year, around 0.6% of its economy, meaning the US is outspending China by close to a ten-to-one ratio on AI infrastructure. Despite that gap, independent benchmark comparisons have shown leading Chinese open-source models completing coding tasks at a small fraction of the cost of top American models, with one comparison finding a roughly seven-to-twelve-times price difference for functionally similar output quality.
Interpretation: China’s cost advantage appears to come substantially from a technique called distillation: rather than training a frontier model from scratch at enormous expense, developers can train smaller, cheaper models by studying the outputs of an already-existing frontier model, then release those smaller models freely as open source. Because American labs are the ones spending the most on original frontier research, this dynamic means much of that spending indirectly subsidizes cheaper Chinese alternatives once those frontier capabilities are replicated and compressed.
Limitation/counterpoint: The US still holds a meaningful lead in raw model capability at the very top tier, according to industry benchmark rankings like the Artificial Analysis Intelligence Index, where the leading US model outscored the leading Chinese open model. However, for the substantial majority of practical business use cases, like customer service automation and other routine tasks that make up most of how companies actually deploy AI, the difference in real-world usefulness between the top and near-top models is often small enough that the price gap matters more than the capability gap.
When Geopolitics Becomes a Direct Business Risk
The tension between US AI dominance and global adoption isn’t purely theoretical. In June 2026, US Commerce Secretary Howard Lutnick sent a letter to Anthropic, the AI company behind the Claude models, warning that the company would need explicit government permission before making its most advanced models, Fable 5 and Mythos 5, available to foreign nationals anywhere in the world, not limited to adversarial nations, citing export control laws typically applied to sensitive dual-use technology. Anthropic disabled access to both models in response, while stating publicly that it viewed the government’s action as disproportionate to the underlying concern, which involved a narrow potential security vulnerability rather than evidence of actual misuse. Access to both models was later restored on July 1, 2026, after the relevant export controls were lifted.
The episode illustrates a real risk sitting underneath the “the world has no choice but American AI” assumption: if access to leading US models can be abruptly restricted by regulatory action, even temporarily and even for allied countries, it strengthens the case for governments and companies elsewhere to prioritize AI infrastructure and models they fully control, reducing exposure to decisions made by another country’s regulators.
Early Warning Signs Analysts Are Watching For
If this dynamic does eventually unwind, historical precedent from the dot-com era suggests it likely won’t wait for companies to formally announce spending cuts. During that earlier bubble, the Nasdaq peaked in March 2000, but the companies laying fiber-optic infrastructure at the time continued spending heavily for another year before the buildout actually slowed, well after the broader market had already turned. Sentiment, in other words, shifted before spending did.
A few specific indicators are being watched as potential early signals this time around:
- A major tech company publicly moderating its infrastructure spending. Because the technology sector tends to move as a group, the first major company rewarded by investors for pulling back AI capital expenditure could give other companies effective permission to follow, similar to how competitors quickly matched Microsoft’s early move to integrate AI into its Bing search product.
- Difficulty issuing new data center debt. With an estimated 100-plus gigawatts of data center capacity reportedly under construction or in planning, funding that buildout requires enormous ongoing borrowing; strain in that debt market, or a shift toward large equity raises instead, as seen with a large recent equity raise from Google, is considered a meaningful signal to watch.
- Widening credit spreads. Corporate bond investors price in a premium above the “risk-free” rate of government bonds to compensate for the risk that a company might default. That premium, called a credit spread, tends to widen sharply when lenders grow nervous about a sector’s financial health. As of the data referenced here, spreads remain historically tight, among the calmest readings on record, which either suggests bond markets see limited risk in the current AI buildout, or reflects the same kind of complacency credit markets displayed in early 2007, months before the 2008 financial crisis became apparent.
- Divergence between AI infrastructure spenders and AI equipment sellers. Analysis attributed to investor Michael Burry has highlighted a pattern where stock prices for companies selling AI chips and equipment have risen sharply, while the stock prices of the large technology companies actually spending the money on AI infrastructure have moved comparatively little, suggesting the market may not fully believe that spending will be recouped. A separate pricing index tracking the cost of AI tokens has also reportedly declined by roughly 20% from its recent high, which analysts say could reflect either a healthy shift toward cheaper, more efficient models, or softening demand at current prices, with the signal itself considered ambiguous.
Where This Analysis Has Real Limits
A few things are worth stating plainly. First, predictions of an imminent AI bubble burst have been made before without materializing on the predicted timeline, and some of the specific analysts and investors whose views are referenced in this kind of analysis, including Michael Burry, have a track record that includes both accurate and premature market calls. Second, credit spreads and other market-based indicators reflect what investors currently believe, not necessarily objective underlying risk, and history shows credit markets can remain calm right up until a crisis becomes visible. Third, this analysis draws heavily on one particular synthesis of commentary from several named figures (Alex Karp, Ed Zitron, Michael Burry) whose individual incentives and track records vary considerably, and their views shouldn’t be treated as a unified, independently verified consensus.
The Takeaway
None of the three structural problems outlined here, vendor trust, broken unit economics, and a credible lower-cost competitor, guarantee that the current AI investment boom collapses on any particular timeline, and reasonable analysts disagree sharply about whether today’s spending will eventually be justified by real, durable profit. What the available evidence does suggest is that the market’s underlying assumption, that American AI companies will inevitably capture outsized global profit because no real alternative exists, is being tested more seriously than it has been at any point since the current AI investment cycle began, and the credit market, capital expenditure trends, and China’s cost position are all reasonable places to keep watching for early signs of that assumption breaking down.
Frequently Asked Questions
Is the AI bubble bigger than the dot-com bubble?
Some analysts argue that, measured as a share of the overall US economy, the current AI infrastructure buildout is larger than the telecom and dot-com infrastructure buildout of the late 1990s and early 2000s.
Why is China able to compete with the US in AI while spending far less?
A significant factor is distillation, a technique where smaller, cheaper models are trained by studying the outputs of existing frontier models rather than being trained from scratch, allowing lower-cost replication of much of a frontier model’s practical capability.
What would actually signal the AI bubble is bursting?
Analysts point to several potential early indicators: major tech companies publicly pulling back AI infrastructure spending, difficulty issuing new data center debt, widening corporate credit spreads, and a growing gap between the stock performance of AI equipment sellers versus the companies actually spending on AI infrastructure.
