Is the AI Bubble About to Burst? China's Challenge Explained

The $1 Trillion Bet That May Not Pay Off

The US stock market's recent strength rests heavily on a single narrative: American AI companies will generate unprecedented profits for decades, and the rest of the world will have no choice but to pay for access. That story is now under serious pressure — and the numbers behind the AI bubble are worth examining closely before assuming your 401(k) is safe.

US companies are on track to spend approximately $764 billion on AI infrastructure in the current fiscal year, scaling toward $1 trillion annually — roughly 3% of the entire US economy. Meanwhile, China is spending an estimated $102 billion, rising to $123 billion next year, representing just 0.6% of its economy. America is outspending China nearly 10-to-1. Yet the outputs are increasingly comparable, and in some cases, the Chinese models are a fraction of the cost to run.

This is not a geopolitical opinion piece. It is a financial analysis of a potentially mispriced market narrative — one that affects index fund investors, retirement savers, and institutional allocators alike.

---

The Business Model Is Structurally Broken

Software has historically been the most profitable business model ever invented. The economics are elegant: build something once, then sell it to millions of customers with near-zero marginal cost. Microsoft's Excel, Adobe's Photoshop, Salesforce's CRM — each new user adds almost pure profit. That is why tech stocks have dominated equity markets for 25 years.

Generative AI broke that model entirely.

Every time a user submits a query to ChatGPT or Claude, OpenAI or Anthropic incurs a real, measurable cost — electricity consumption, chip wear, server load. More customers do not mean more margin. They mean more cost, dollar for dollar. The analogy that captures it best: AI is not a software business. It operates more like a restaurant where every meal costs money to prepare — except this restaurant consistently loses money on every dish served and believes the solution is to serve more dishes.

The reported financials support this concern. OpenAI burned through approximately $20.9 billion in 2025, according to figures reported by the Financial Times. More troublingly, analysts note that costs are scaling linearly with revenue — meaning the gap between the two lines, where profit is supposed to live, is not opening up. For 25 years, investors have been conditioned to be patient with loss-making tech companies citing the Amazon precedent. But Amazon's losses were tied to logistics infrastructure that generated real margin improvement at scale. With generative AI, each successive model generation costs more to run than the last.

OpenAI has reportedly delayed its IPO to 2027, in part because it could not justify a $1 trillion valuation. That detail alone signals that sophisticated institutional money is beginning to ask harder questions.

---

Hidden Revenue, Hidden Risk: What Big Tech Isn't Disclosing

Microsoft, Google, Amazon, and Meta collectively represent a substantial portion of the S&P 500. All four are spending aggressively on AI infrastructure. None of them disclose AI-specific revenue as a separate line item.

They report cloud revenue. They report advertising revenue. They report subscription revenue. But AI revenue — the actual return on the hundreds of billions being deployed — remains conspicuously absent from earnings disclosures.

This matters for a specific reason: public companies are highly incentivized to share good news. When a segment is performing well, it gets its own slide in the investor deck. The absence of AI revenue disclosure, combined with continued strong performance in legacy business lines like cloud and advertising, creates a misleading impression that AI is driving growth. The more likely interpretation, based on available evidence, is that legacy businesses are growing despite AI spending, not because of it, and that AI is currently a net cost center across the board.

Oracle's situation illustrates the downstream risk. The company is reportedly building 7.1 gigawatts of data center capacity — the equivalent of multiple nuclear power plants — largely for a single customer: OpenAI. Oracle's own annual report acknowledged the risk of not being paid. If OpenAI's financial position deteriorates further, the ripple effects through Oracle's balance sheet — and through the margin loans that Oracle's executive leadership holds against company stock — could be significant.

Nvidia faces a structurally similar concern. A portion of its GPU sales flow through smaller cloud companies called NeoClouds. These firms borrow billions to purchase Nvidia hardware, which Nvidia then reportedly pays to rent back. When a supplier is partially financing its own demand, that is not organic customer acquisition. It is a sign of insufficient genuine end-market demand.

---

Why Enterprises Are Starting to Push Back on AI

Beyond the macro financial picture, there is a ground-level trust problem developing between AI vendors and corporate customers.

The pricing model itself is a source of friction. AI companies charge per token — essentially per word read or generated — regardless of whether the output created any value. A lawyer charges to win a case. A contractor charges to complete a renovation. The price is tied to a result. AI companies charge for the process, successful or not, including the instances where models "hallucinate" — generating confident, plausible-sounding but factually incorrect outputs with no reliable way to predict when or why it will happen.

For enterprise CFOs trying to justify $50 million in annual AI subscriptions to their boards, the inability to quantify ROI is a serious problem. There is no standard metric, no industry benchmark, and no vendor willing to put forward a pay-for-performance model. If the technology genuinely delivered the transformative productivity gains being marketed, a pay-on-results pricing structure would be the easiest sale in enterprise software history.

The IP leakage concern compounds this. When companies use third-party AI models, proprietary workflows, trade secrets, and competitive differentiators flow through those systems. Anthropic's launch of a design product called Claude Design — while maintaining an active relationship with design platform Figma — caused Figma's CEO to publicly express shock. It raised a legitimate question for every enterprise buyer: are you paying a vendor, or are you training your own replacement?

The logical response, increasingly voiced by sophisticated enterprise buyers, is to run open-source models locally on owned infrastructure, where no external party can access or learn from proprietary data, and crucially, where no one can switch off access.

---

China's Cost Advantage Is the Variable Nobody Priced In

The foundational assumption underpinning the AI investment narrative — and by extension, much of the current premium in US tech valuations — is that American AI is irreplaceable. The world will pay whatever is charged because there is no alternative.

That assumption is being tested in real time.

A comparison circulating among developers illustrates the gap starkly. The same coding task, submitted to Anthropic's Claude Opus (a leading US model) and to GLM (a Chinese open-source model), took both approximately 5.5 minutes to complete. Claude Opus charged $2.33 for the task. The Chinese model cost a fraction of that — estimates suggest 7 to 12 times cheaper depending on the workload.

China is spending roughly one-tenth of what the US is spending on AI, producing models that perform comparably on many benchmarks, and distributing them at minimal cost or for free. The strategic logic is straightforward: if you cannot yet win on raw capability, win on accessibility and price, and let adoption do the rest.

The geopolitical dimension adds another layer of complexity. When the US Commerce Department ordered Anthropic to cut off two of its most powerful AI models from every foreign national globally — covering not just China but France, Germany, Japan, and even Anthropic employees without US citizenship — it demonstrated that access to American AI is a policy variable, not a guaranteed utility. Within four days, France terminated a contract with a major US technology partner. Germany, Spain, and the UK each began reconsidering dependencies on US AI infrastructure.

The European response was predictable in retrospect: governments and enterprises do not want critical infrastructure that can be switched off by a letter from a foreign commerce secretary. Open-source Chinese models, run on locally owned hardware, cannot be remotely disabled.

---

What Investors Should Watch

None of this means the AI sector collapses tomorrow, or that every company with AI exposure is overvalued. But several signals are worth tracking for investors trying to assess whether the current narrative holds:

• AI revenue disclosure: If Microsoft, Google, or Amazon begin breaking out AI-specific revenue as a positive contributor, that changes the analysis materially. Continued silence is informative.
• OpenAI's IPO timeline and valuation: If the valuation at IPO is substantially below $1 trillion, it will mark a public data point on what the market actually believes this technology is worth.
• Enterprise renewal rates: Whether large corporations renew or reduce AI software subscriptions in 2025-2026 will be a leading indicator of real-world ROI perception.
• Chinese model adoption outside China: If European or Asian enterprises increasingly choose Chinese open-source models over US commercial products, the moat that justifies US AI premiums narrows significantly.
• Margin trajectory: Investors should watch whether AI-linked costs for major cloud providers grow faster than cloud revenue. If so, the AI narrative is actively destroying value in existing businesses rather than creating it.

The dot-com bubble of 1999-2000 is the obvious historical reference point. At its peak, the combined market capitalization of technology companies implied profit growth that mathematics could not support. Many of those companies had real technology. Some of them still exist and are profitable today. But investors who bought at peak valuations waited over a decade to break even. The technology was real. The prices were not.

---

The Practical Takeaway

The AI investment story is not simply true or false — it exists on a spectrum between transformative technology and overpriced narrative, and right now there is credible evidence pushing in both directions. What is clear from the financial data is that the current cost structures do not support the profit timelines priced into many tech valuations, that enterprise trust in AI vendors is fragile, and that China has emerged as a genuine low-cost alternative that the market did not fully price in.

For investors with significant exposure to US tech through index funds or direct holdings, this is not a reason to panic-sell. It is a reason to understand precisely what assumptions your portfolio is making — and to watch the specific indicators above for signs that those assumptions are breaking down.

Diversification across geographies, sectors, and asset classes remains the most durable hedge against a narrative that turns out to be wrong.

---

Frequently Asked Questions

Is the AI bubble comparable to the dot-com bubble of 2000? There are structural similarities — widespread speculative investment in a technology whose commercial applications remain unproven at scale — but also important differences. The current AI buildout represents a larger share of GDP than the TMT investment cycle of 1999-2000, and the companies involved are generally more financially robust than the speculative startups of that era. However, the lack of disclosed AI-specific revenue from major tech companies and the worsening margin trajectories at AI-native firms do echo the late stages of the dot-com expansion.

Why is China able to build competitive AI at a fraction of the US cost? Several factors contribute: open-source model architectures reduce the need to build from scratch, Chinese hardware and energy costs differ from US equivalents, and Chinese AI labs have demonstrated an ability to achieve near-comparable performance through architectural efficiency rather than raw compute scale. The release of models like DeepSeek in early 2025 provided a public benchmark showing Chinese AI performance at significantly lower inference cost than leading US commercial models.

Does China's AI advantage mean US tech stocks will fall? Not necessarily or automatically. US tech companies have durable advantages in enterprise relationships, regulatory trust in Western markets, and ecosystem integration. However, if China's models achieve sufficient adoption among non-US enterprises, the pricing power that underpins US AI revenue projections weakens, which affects the valuation multiple investors are willing to pay. This is a risk factor, not a certainty, and investors should assess their exposure accordingly rather than drawing binary conclusions.

What does token-based pricing tell us about AI companies' confidence in their product? Token-based pricing — charging per word read or generated regardless of outcome quality — is structurally different from results-based pricing. As Palantir CEO Alex Karp has noted publicly, if AI vendors were confident that their models reliably created measurable business value, a revenue-share or success-fee model would be commercially rational and far easier to sell. The persistence of token pricing suggests that vendors cannot yet consistently guarantee outcomes, and that hallucination rates and unpredictability remain significant enough to make results-based contracts financially risky for the vendors themselves.

---

This article is for informational purposes only and does not constitute financial advice. Always consult a qualified financial professional before making investment decisions.