Big Tech Parked $1.65 Trillion of AI Debt Where You Can't See It
A Nikkei investigation just put a number on something a lot of people suspected. Alphabet, Microsoft, Amazon, Meta and Oracle are carrying roughly 1.65 trillion dollars of AI-related debt that never shows up on their balance sheets. For scale, that is 122 percent of the 1.35 trillion they actually report. The hidden pile is bigger than the visible one.
The trick is old and it has a name that should make you nervous: the special-purpose vehicle. A separate legal entity borrows the money and owns the data center. The tech giant just signs a long-term lease or capacity contract, so the debt sits on the SPV's books, not its own. Oracle has moved about 66 billion this way, Meta around 30 billion with total off-book obligations near 420 billion, xAI 20 billion, CoreWeave 2.6 billion. The last company that got famous for this kind of accounting was Enron.
Why should an agent person care about corporate accounting? Because this is the bill for the buildout that runs your agents, and someone has to pay it back with tokens. When these data centers switch on, the leases roll onto the balance sheets and the debt spikes overnight. If AI revenue does not show up on schedule, the assets get written down and the losses flow straight to the lenders and bond investors who funded the whole thing.
Here is the part that connects to the work. Agent workloads are the heaviest inference there is. One autonomous agent run can burn the tokens of a hundred chat turns, looping, retrying, reading tool output. The optimistic case for all this leverage is that agents become the demand that fills these buildings and justifies the debt. The pessimistic case is that agents are still too expensive and too unreliable to generate that revenue in time. 1.65 trillion dollars is riding on which case is true, and nobody gets to opt out of the answer. https://www.tomshardware.com/tech-industry/big-tech/ai-tech-companies-have-hidden-debt-worth-around-usd1-65-trillion-report-claims
← Back to all articles
The trick is old and it has a name that should make you nervous: the special-purpose vehicle. A separate legal entity borrows the money and owns the data center. The tech giant just signs a long-term lease or capacity contract, so the debt sits on the SPV's books, not its own. Oracle has moved about 66 billion this way, Meta around 30 billion with total off-book obligations near 420 billion, xAI 20 billion, CoreWeave 2.6 billion. The last company that got famous for this kind of accounting was Enron.
Why should an agent person care about corporate accounting? Because this is the bill for the buildout that runs your agents, and someone has to pay it back with tokens. When these data centers switch on, the leases roll onto the balance sheets and the debt spikes overnight. If AI revenue does not show up on schedule, the assets get written down and the losses flow straight to the lenders and bond investors who funded the whole thing.
Here is the part that connects to the work. Agent workloads are the heaviest inference there is. One autonomous agent run can burn the tokens of a hundred chat turns, looping, retrying, reading tool output. The optimistic case for all this leverage is that agents become the demand that fills these buildings and justifies the debt. The pessimistic case is that agents are still too expensive and too unreliable to generate that revenue in time. 1.65 trillion dollars is riding on which case is true, and nobody gets to opt out of the answer. https://www.tomshardware.com/tech-industry/big-tech/ai-tech-companies-have-hidden-debt-worth-around-usd1-65-trillion-report-claims
Comments