Goldman Sachs’ forecast data is worth rereading three times by everyone involved in the AI supply chain: by 2027, hyperscaler AI data center capital expenditures will reach $1.14 trillion, with about one-third needing to be supported by borrowing. Compared with $405 billion in 2025 and an expected $750 billion in 2026, this means the bill will triple within two years. The key point isn’t whether these giants can earn back the money, but that the borrowing structure itself is changing the foundation of the compute industry.
First layer: the investment-grade bond market. In 2025, $108 billion was issued, accounting for 26% of that year’s capex; in the first half of 2026, $194 billion has already been issued. But this layer only serves AAA/AA ratings and can’t cover a large share of new entrants.
Second layer: private credit. Project finance and asset-backed loans absorb the remainder. The five biggest players have disclosed about $1.2 trillion in leasing commitments—essentially an off-balance-sheet liability.
Third layer: the on-chain credit market. This layer is currently small, but its infrastructure logic is completely different from traditional finance: it uses tokenized Nvidia hardware as collateral to enable instant borrowing and liquidation. USD(dot)AI reported that it issued the first $100 million GPU-collateralized loan this Q1, with returns coming from borrowers’ repayments rather than token inflation. Figure’s Q2 data is even more striking: $4.3 billion in consumer loans and a 132% year-over-year increase, with 65% of the lending already completed through its tokenized market.
The metaphor in this chain is: when AI compute becomes collateral that can be liquidated on-chain at any moment, its pricing power will shift from Nvidia’s futures order book to the liquidation engine of on-chain lending protocols. In traditional finance, collateral valuation depends on auditors and rating agencies; in the on-chain world, it depends on oracles and liquidation rules. If AI infrastructure does indeed triple over the next five years as Goldman Sachs predicts, then the last dollar may not be lent by banks—it may be issued from the liquidation lines of DeFi protocols.
Now the question is: who will foot the bill for that final $1.14 trillion—the bondholders, private funds, or those on-chain lenders that issued the first $100 million and have not yet defaulted?
First layer: the investment-grade bond market. In 2025, $108 billion was issued, accounting for 26% of that year’s capex; in the first half of 2026, $194 billion has already been issued. But this layer only serves AAA/AA ratings and can’t cover a large share of new entrants.
Second layer: private credit. Project finance and asset-backed loans absorb the remainder. The five biggest players have disclosed about $1.2 trillion in leasing commitments—essentially an off-balance-sheet liability.
Third layer: the on-chain credit market. This layer is currently small, but its infrastructure logic is completely different from traditional finance: it uses tokenized Nvidia hardware as collateral to enable instant borrowing and liquidation. USD(dot)AI reported that it issued the first $100 million GPU-collateralized loan this Q1, with returns coming from borrowers’ repayments rather than token inflation. Figure’s Q2 data is even more striking: $4.3 billion in consumer loans and a 132% year-over-year increase, with 65% of the lending already completed through its tokenized market.
The metaphor in this chain is: when AI compute becomes collateral that can be liquidated on-chain at any moment, its pricing power will shift from Nvidia’s futures order book to the liquidation engine of on-chain lending protocols. In traditional finance, collateral valuation depends on auditors and rating agencies; in the on-chain world, it depends on oracles and liquidation rules. If AI infrastructure does indeed triple over the next five years as Goldman Sachs predicts, then the last dollar may not be lent by banks—it may be issued from the liquidation lines of DeFi protocols.
Now the question is: who will foot the bill for that final $1.14 trillion—the bondholders, private funds, or those on-chain lenders that issued the first $100 million and have not yet defaulted?