The AI Debt Bomb: $700 Billion In, $30 Billion Out, Who Pays When the Math Breaks?

Goldman Sachs is worried. Sequoia Capital is worried. OpenAI's own CFO is worried. The companies building AI are not. One of these groups is wrong, and the timeline is narrowing.

Five of the largest US tech companies are spending close to $700 billion this year building AI data centers. Total AI-attributable revenue across those same platforms sits somewhere around $30–80 billion. Goldman Sachs has published a formal warning, their "Tracking Trillions" report, saying AI won't meaningfully impact the broader economy until 2027. Sequoia Capital ran the numbers and asked: where is the $600 billion in AI revenue that would justify the current buildout? Nobody answered. Meanwhile, OpenAI's own CFO has flagged fears about servicing $1.5 trillion in funding obligations. The insiders are nervous. The marketing departments are not.

Section 1: The Numbers That Don't Close

Start with the arithmetic. The Big Five US hyperscalers, Amazon, Microsoft, Meta, Google, and Apple, have collectively committed hundreds of billions to AI infrastructure spending in 2025 and 2026. The numbers are not secret. They are in earnings calls, analyst reports, and infrastructure announcements. Microsoft has committed to $80 billion in capex. Meta to $65 billion. Amazon to over $100 billion. Google to $75 billion. Combined, this is a bet approaching $700 billion on a technology that has not yet produced proportional returns.

Goldman Sachs' "Tracking Trillions" report made the institutional alarm public. The data processing sector accounts for only 4% of US GDP, but captured 92% of GDP growth in H1 2025. Read that twice. Every other sector of the American economy, manufacturing, retail, healthcare, construction, is effectively subsidizing AI's growth while producing none of the returns. Goldman didn't say AI is a fraud. They said: don't expect significant economic impact before 2027. That's a four-year payback window on a $700 billion bet made right now.

The Economist's analysis adds a supply-side complication: AI model providers are already throttling access due to demand they cannot physically meet. This is not a sign of success: it is a sign that the infrastructure buildout is not keeping pace with the compute requirements, even as the bills for that infrastructure keep growing. The companies paying for data centers that are not yet fully operational are already behind on both the cost side and the capacity side simultaneously.

The OpenAI CFO situation is, by itself, a news event that should have generated more coverage than it did. The company's CFO publicly flagged that OpenAI's funding obligations, commitments made to investors as part of its ongoing capital raises and restructuring, may not be serviceable under the current revenue trajectory. ChatGPT processes 2.5 billion queries per day and requires the energy equivalent of one nuclear reactor's daily output to run at that scale. The cost structure is not shrinking. The revenue must catch it. If OpenAI's own CFO is worried, the "AI is on a sustainable trajectory" narrative deserves serious scrutiny.

Section 2: The Circular Financing Loop

Understanding why the capex keeps growing despite the revenue gap requires understanding the mechanism that sustains it. It is not organic demand. It is a financing loop.

Nvidia invests in AI companies. AI companies use that capital to buy Nvidia GPUs. Nvidia's stock rises: it hit a $5 trillion market cap in 2025. Nvidia uses that equity wealth to invest in more AI companies. Those AI companies raise more capital, partly on the strength of their Nvidia partnerships, and buy more Nvidia GPUs. The loop runs.

The Cursor/SpaceX deal is the purest single expression of this dynamic. A two-year-old coding tool, Cursor, was the subject of a deal in which SpaceX provided compute resources in exchange for acquisition rights at up to $60 billion by end of 2026. Sixty billion dollars. For a coding tool that started in 2024. The valuation is not a product valuation. It is a compute placement fee. SpaceX has GPU commitments it needs to utilize. Cursor provides a workload that consumes those GPUs. The $60 billion acquisition price is SpaceX paying for future utilization rights, not for the coding tool's product quality or user base. This is the circular logic made explicit: capex creates compute, compute needs somewhere to go, so infrastructure buyers acquire software companies at prices that would otherwise be absurd.

Sequoia Capital's "$600 Billion Question" analysis put the aggregate version of this problem in print: the AI ecosystem requires roughly $600 billion in additional annual revenue to justify its current cost structure. No such revenue exists. No credible model shows it appearing before 2027. The loop sustains itself only as long as equity markets believe the end state is worth what the participants are claiming, and as long as Goldman Sachs reports are treated as cautionary warnings rather than exit signals.

Section 3: The AI Debt Ledger

Abstract numbers become real when you see them in a table. Below is the AI Debt Ledger: each major AI platform's declared capex commitment, their AI-attributable revenue, and the resulting coverage ratio. A coverage ratio below 1.0 means the company is spending more building AI than it earns from it. A coverage ratio below 0.2 means the gap is so severe that the company's AI division could not fund itself for five years even if revenue stopped growing at all.

The supply chain adds a further constraint that the table cannot fully capture. The entire AI infrastructure buildout runs through a single island: TSMC in Taiwan produces the advanced logic chips that power every major AI data center. ASML in the Netherlands produces the only lithography machines capable of manufacturing those chips at the required scale. Even if demand is real and revenue is imminent, the physical supply chain cannot scale fast enough to close the capex gap in the 2025–2026 investment window. A Bloomberg Economics model from February 2026 estimated that a Taiwan conflict would cost the global economy $10 trillion: a number that puts the AI buildout's $700 billion in sharp perspective. The people making capex decisions know this. The Goldman Sachs report exists because they know this.

Section 4: The Human Ledger

The AI capex discussion focuses almost entirely on institutional capital. But there is a second ledger, one denominated in careers, not dollars, and it is equally important to understand.

Klarna is the most documented case study. The Swedish fintech froze hiring, cut its workforce from 5,500 to 3,400 employees, and deployed an AI chatbot it claimed could replace 700 customer service agents. The cost savings were real, in the short run. The customer trust damage was also real. The chatbot handled simple, scripted interactions well. On complex issues, disputes, fraud, credit questions: it failed consistently. Customer satisfaction dropped. The company was forced to reassign engineers and marketing staff to answer customer calls, a profound signal that the ROI calculation had missed something fundamental. Klarna has since begun rehiring.

The scale is not trivial. Total US layoffs in 2025 hit 1.17 million: the worst since COVID, according to Challenger, Gray & Christmas tracking. Goldman Sachs' April 2026 labor market report was blunt: workers displaced by AI face lower wages, longer unemployment periods, and harder conditions than previous technological displacement waves. The cost of the experiment is not being borne by the investors making the capex decisions. It is being borne by the workers told their roles were being automated, many of whom will be called back to those same roles once the automation fails.

5 Signals That a Company's AI Experiment Is Failing

• Customer satisfaction metrics decline in the 90 days after AI deployment, especially for complex or high-stakes interactions
• Engineers or non-support staff are being redirected to handle customer inquiries: the AI cannot own the edge cases it was supposed to replace
• The company stops publishing specific metrics about AI performance, shifting to vague "productivity gains" language
• Rehiring begins in the exact roles that were cut, often framed as "expanding" rather than reversing
• The CEO or CFO stops making claims about AI replacing headcount and begins describing AI as a "force multiplier" for existing teams

Related on Aether Intel