Concerns about off-balance-sheet AI commitments grew sharply this past week, driven by headline estimates of $3–4 trillion. But those totals say little about how big the risk is or where it sits.
Not every commitment is a loan waiting to go bad. These obligations have very different purposes and economics.
Purchase commitments: building capacity. Committing to equipment or projects under construction isn’t a red flag by itself. What matters is whether cash flow can cover the commitments and whether the investments will earn attractive returns.
Google stands out here, well ahead of Amazon and Microsoft. That suggests more upside to Google’s capex.
Leases not yet commenced: renting capacity. Hyperscalers can build or lease. A company with fewer construction commitments may have more future lease obligations. Neither route makes growth free.
Microsoft leans toward leasing. Meta is aggressive on both fronts, which fits its AI agent ambitions. Companies with their own internal use cases, like Google and Meta, have more confidence to build. That’s a positive sign.
Guarantees and backstops: underwriting someone else’s business. This is where the “Bank of AI” concern has more weight. Nvidia is putting its balance sheet behind customers’ expansion, helping them get financing while taking on risk itself.
Two deals stand out. Nvidia has guaranteed up to $105 billion for an OpenAI data center in Ohio, an amount that shrinks as lease payments are made. It also has a CoreWeave backstop, first valued at $6.3 billion, that requires it to buy qualifying unsold computing capacity through April 2032.
Still, these guarantees and backstops add up to less than one year of Nvidia’s EBITDA, and the company has only $30 billion in debt. That’s hardly a stretched balance sheet.
The debate also leaves out Nvidia’s equity wins. Its $30 billion OpenAI stake is worth $45 billion at the latest valuation. Its $10 billion xAI investment, now part of SpaceX, was worth $21 billion at June 30. Its $10 billion Anthropic stake would be worth about $40 billion at the proposed IPO valuation.
That’s $50 billion-plus in gains on three deals in under two years. The Bank of AI may not be doing so badly after all.















