Wall Streets Quiet Panic Over AIs $10 Trillion Gamble

The Math That Should Keep Wall Street Up at Night

Anthropic went public asking for $518 billion in future investment. Its current annual revenue sits somewhere around $10 billion. That is a hundred times the company’s existing income statement. Not a typo. A hundred.

This single data point captures the entire absurdity of the AI infrastructure boom — and why Wall Street’s recent warnings carry more weight than the usual cyclical skepticism. The investment community is not just questioning whether AI will pay off. It is quietly adjusting valuations downward because the revenue math simply does not close.

According to a Brookings Institution report released last month, cumulative US spending on AI data centers and supporting infrastructure between 2025 and 2032 will reach $10.3 trillion. That equals roughly 3.63 percent of annual US GDP every year for eight years. To put that in historical perspective: the Transcontinental Railroad era absorbed 2.24 percent of GDP. The Interstate Highway System, 1.10 percent. The fiber-optic telecom buildout of the 1990s, a modest 0.66 percent. None of those legendary infrastructure waves came close to what the AI economy is now demanding.

PwC estimates global data center spending alone will exceed $30 trillion by 2050 — a figure roughly equal to the entire current outstanding amount of US Treasury debt. Four of the five hyper-scalers — Google, Amazon, Meta, Microsoft — plus Oracle are tracking toward $4.2 trillion in combined capital expenditure through 2029, according to FactSet.

No company on earth generates revenue at anything approaching the scale needed to service that level of spending. And yet the money keeps flowing.

The Corporate-Bond Trap Nobody Is Talking About

Here is where the story gets uglier, and where English-language financial coverage has largely missed the signal. Korean financial media — notably AI Times — has been warning about something more structural than a stock-valuation bubble: a corporate-bond liquidity trap being engineered by SPV structures built around AI infrastructure. The analogy to subprime mortgage-backed securities is not decorative. Steen van Nieukergh, a Columbia Business School professor who authored the Brookings report, told Reuters directly that the opacity surrounding these special-purpose vehicles recalls the pre-2008 era.

What happens when demand for AI services falls short — and every signal so far suggests it might — is that the debt layered onto these SPVs has nowhere to roll. Corporate bond markets are already showing signs of crowding. A sudden repricing event would not be contained to tech equities. It would move through the fixed-income system, which is where most institutional exposure to AI infrastructure actually lives.

This is the risk that makes the AI bubble different from the dot-com crash or the late-1800s railroad mania. In those episodes, equity holders took the hit first. Today, the leverage sits predominantly in the bond market, where diversification is thinner and fire-sale dynamics are faster. The Brookings report does not dwell on this link. But it is the quiet center of the Korean regulatory warning, and it deserves far more attention than it is getting in New York or London.

Who Is Wrong About Everything

The revenue gap is so enormous that even optimistic assumptions collapse under scrutiny. Van Nieukergh calculated that for the US AI sector to achieve a reasonable 10 percent return on the projected $10.3 trillion investment, it would need to generate between $3.55 trillion and $3.7 trillion in annual revenue by 2032. Current combined revenue for OpenAI and Anthropic is estimated at roughly $100 billion. That means AI revenues would need to grow roughly thirty-five-fold in seven years — an 80 percent compound annual growth rate sustained against an economy that, according to the Congressional Budget Office, is projected to see productivity increase by only 1.75 percent per year over the same period.

JP Morgan’s own analysis lands on the same contradiction. The bank concluded that US productivity would need to grow at 3 to 5 percent annually over the next decade just to justify the valuations being assigned to AI companies today. The CBO benchmark sits at 1.75 percent. The gap is not narrow. It is structural.

This is not a new argument. Diane Coyle, an economics professor at Cambridge, noted that historically, the productivity benefits of genuinely transformative technologies — the steam engine, electrification, the internet — took anywhere from ten to fifty years to materialize in aggregate economic data. The financial schedules of AI companies, built on quarterly revenue targets and multi-year capex commitments, are colliding with a reality that moves on a different timescale.

DeepSeek and the Open-Source Pivot

Adding further pressure to the thesis that AI infrastructure spending is disconnected from commercial reality is the rapid acceleration of open-source models. According to OpenRouter data, DeepSeek now commands 25.3 percent of text model requests, pulling ahead of OpenAI’s 18.6 percent and Anthropic’s 2.9 percent. This is not a marginal shift. It signals that a significant portion of the market is already choosing cost-effective inference over proprietary ecosystems — exactly the dynamic that undermines the revenue projections underpinning trillions in capex.

Scott Wilson, chief investment officer of the University of Washington’s endowment and an early investor in SpaceX, told Business Insider bluntly that companies like OpenAI and Anthropic are “not worth the risk they’re taking.” His broader point was that the emerging trillion-dollar AI firms are assuming responsibilities far beyond their actual economic value. The open-source pivot, he argued, is not just a technical choice — it is a commercial indictment of the closed-model pricing strategy.

Vinod Khosla, an early OpenAI investor, pushed back hard. He called Wilson’s view foolish and argued that open-source models represent only the software layer, while the real costs — electricity, server maintenance, optimization, energy — make proprietary stacks far more cost-effective at scale. Whether Khosla is right or wrong depends on the time horizon. His argument holds in the long run. It does not help Anthropic raise $518 billion today.

The Divergence That Matters Most

The most useful framing for understanding where this is heading comes from the split between two institutions that otherwise agree on the basic facts. JP Morgan warns that if infrastructure spending fails to generate sufficient returns, the technology boom will end abruptly — and the shock will move through the financial system. Cambridge’s Coyle, by contrast, argues that as long as the physical infrastructure is built, the productivity payoff will eventually arrive; short-term losses are the price of long-term capability.

Both can be true simultaneously. The question is timing, and timing is where the corporate-bond risk lives. If the productivity gains take ten to twenty years to emerge, the SPVs and leveraged structures built around AI infrastructure will need to refinance or roll over debt through periods of negative cash flow. That is precisely the environment in which corporate bond markets seize up.

The $10 trillion number is not a forecast. It is a commitment. The question is whether anything close to $3.7 trillion in annual AI revenue exists anywhere in the foreseeable path — and whether the bond market will notice before the equity market does.