Big Tech has funded $120bn of AI data centres with off-balance-sheet SPVs. What does this mean?

The CapEx is large

AI’s capital expenditure is estimated at $400bn for 2025, $500bn for 2026, and $1.3 trillion by 2030. As a reference, the entire US GDP is expected to be $36–37 trillion by then. This level of CapEx is enormous. It impacts society in multiple ways, notably infrastructure, with electricity demand and costs. Moreover, electronic hardware is not standard infrastructure: technology depreciates rapidly and must be renewed continuously.

SPVs

Now, Enron had $27bn of its $38bn debt in SPVs, leaving shareholders unaware of the firm’s true leverage and expenses. We know how that story ended. SPVs deliver leverage and opacity. We do not yet know who will ultimately fund this massive CapEx, but it is safe to say it will not be an institutionals-only problem. Since 401(k) plans can now allocate to private equity and private credit funds, it is the general public that will increasingly fund this debt. Cross-shareholdings between Big Tech firms, and with the government, only increase systemic risk.

Insufficient revenues

Meanwhile, OpenAI generated $5bn of losses versus $3.7bn of revenues in 2024, and does not expect to become cash-flow positive until the end of the decade. Current AI companies are nowhere near the profitability required to support these costs, even under optimistic scenarios. Add to this the risk that new entrants or technologies can disrupt the landscape at any time (remember DeepSeek?), and this becomes a major gamble on a fast-moving but not yet fully proven technology.

The “Magnificent Seven,” representing approximately 30% of the S&P 500, trade at 25x–30x multiples. That leaves plenty of room for a downward repricing.

What about the efficiency gains for society? There is no guarantee that AI’s expected productivity improvements will justify these expenditures. Job displacement is already visible, notably across consulting firms.

What would happen

If AI fails to deliver sufficient value, large bankruptcies would likely follow, most likely leading to bailouts, at a time when the country is already heavily indebted. 

• Side note: taxing the rich, often presented as a solution to government debt, does not work well when asset values collapse and jobs disappear simultaneously.
• Second side note: the Fed’s balance sheet already stands at $6.5 trillion.

If AI disappoints, the dot-com bubble and the real-estate bubble may look like children’s games by comparison.

Reference: 

• Tech groups shift $120bn of AI data centre debt off balance sheets (Financial Times), article behind paywall