The numbers in an AI company's investor presentation and the numbers buried in the footnotes of its regulatory filings cover the exact same period. Both are public. Neither is false. But the pictures they paint can look strikingly different.
When CoreWeave listed on the Nasdaq in March 2025, its IPO priced at the bottom of the expected range. The company, which operates GPU rental infrastructure, had booked roughly $1.9 billion in revenue in 2024 — a number that reads like a growing AI infrastructure business. What investors dug into instead were the footnotes beneath the balance sheet: data center space agreements, power supply contracts, and cloud computing commitments requiring payment whether or not the contracted capacity was used. Disclosed separately in the filing, the total value of these commitments exceeded annual revenue.
CoreWeave isn't an isolated case. Across the AI industry today, it's common to find the same pattern: a company's headline financial figures tell one story, while the commitments disclosed in its filing footnotes tell another.
Operating Leases Moved Onto the Balance Sheet. Cloud Commitments Didn't.
U.S. accounting rules changed starting in 2019 under FASB ASC 842. Operating leases — building and equipment rentals that had previously stayed off the books — began appearing on the balance sheet as assets and liabilities. It's part of why accounting standards can seem to have already closed off most avenues for financial opacity.
Cloud computing service agreements take a different path. Contracts to purchase GPU compute as a service — particularly those with take-or-pay terms guaranteeing a minimum usage level — aren't currently recognized as liabilities under accounting standards. They're classified instead as commitments to receive goods or services in the future.
Disclosure is still required. But it shows up not on the balance sheet, but in the "Commitments and Contingencies" section of the filing's footnotes — a different location than the financial metrics that press coverage and investor presentations tend to cite, like debt ratios and cash on hand.
OpenAI's compute commitment to Microsoft and Anthropic's cloud credit agreement with Amazon, worth up to $4 billion, both fall into this category. Both are matters of public record, yet neither shows up in the total liabilities on either company's balance sheet.
When the Same Commitment Becomes an Asset — or a Liability
Whether a take-or-pay commitment strengthens a company's financial position or weighs it down depends on timing — and the exact same contract terms can cut either way.
In the second half of 2023, Nvidia's H100 GPU supply couldn't keep up with demand. Companies that had locked in compute commitments with cloud providers ahead of time scaled up capacity faster than their competitors. In that environment, the commitment functioned as a hedge — a way to lock in reliable supply ahead of the pack.
The reverse case exists too. AI inference costs have fallen sharply since 2023 — multiple AI infrastructure analyses estimate that the cost of GPT-4-class inference has dropped more than 90% since then. When the compute volume a company assumed it would need at the time of signing starts to look like excess capacity because of technological progress, a take-or-pay clause — payment due whether or not it's used — turns into a fixed cost.
CoreWeave's S-1 filing shows that its total non-cancelable commitments tied to data centers, power, and equipment exceed its annual revenue. For a company whose commitments are growing faster than its revenue, even as revenue rises, unit-economics metrics can end up pointing in opposite directions.
What the Footnotes Actually Tell You
For publicly traded AI companies, the "Commitments and Contingencies" section of the footnotes in the annual 10-K filed with the SEC lays out take-or-pay commitments, non-cancelable operating leases, and other future spending obligations.
When there's a large gap between the total in that section and total liabilities on the balance sheet, the financial health metrics a company presents publicly may look better than the scale of its actual obligations would suggest. A large gap doesn't necessarily signal trouble on its own — sometimes demand growth is keeping pace with rising commitments. But without reading the total commitments alongside current revenue, it's hard to even begin forming a judgment.
Private AI startups don't disclose this information at all. There are indirect signals worth watching instead: whether a major service has been scaled back, pricing structures have changed, or a key cloud partnership has shifted within the past 12 months. When changes like these appear without accompanying growth metrics, it's hard to rule out that the burden of compute commitments is affecting operations.
For a solo operator paying a monthly subscription for an AI service, it's worth checking three things about any service tied to a core workflow, roughly once every six months: whether the company is publicly traded, who its main compute partner is, and any recent financial news about it. Just tracking these together changes what you have to work with when judging the risk that a service might not stick around.
