Ares Just Declined AI Chip Financing. The Money Still Exists — It Now Costs SOFR+550

The replacement trade is not exotic. Arougheti described the target asset in unusual detail: large metropolitan sites — he named Tokyo, Osaka, London, Atlanta, Amsterdam and São Paulo — pre-let to large investment-grade hyperscale counterparties on 12-to-15-year leases with no termination rights and inflation escalators. The building is drawn against a contract that already exists, so the lender never has to forecast a technology curve. That contract, not conviction, is what makes the $2 trillion fundable.

The franchise can absorb the refusal. Ares raised more than $31 billion of gross capital in the quarter, closed a first close of over $8 billion in equity commitments the week before the call, and guides to 16-to-20% fee-related earnings growth for 2026. Read the \"simpler underwriting\" line as both analysis and marketing: Ares earns two-thirds of its fees in credit, so its incentive is to keep pointing at the bucket it can win.

There is no shortage of money to fund AI. Goldman Sachs Research estimates roughly $500 billion of AI-related debt was issued in 2026 so far, that debt should cover about 33% of hyperscaler capital spending this year and roughly 35% in 2027, and that private markets hold about $4.5 trillion of dry powder. The scarce input is not capital. Scarcity has moved from money to balance sheet willing to warehouse obsolescing silicon.

The evidence that this is a price problem, not a money problem, is who is walking away and why. Insurers are close to their data-center exposure caps and less willing to take risk off banks' books. KKR is steering clear of remote sites that would be hard to re-lease, and Carlyle has passed on several transactions. Those are not statements about the supply of capital. They are statements about the price of holding GPU collateral, and they set the terms for everyone who does write the check.

Four channels are absorbing the demand. Each one moves the risk somewhere different, and the destination matters more than the headline amount.

This is what makes Ares' refusal look disciplined rather than fearful. A lender that bought compute-backed paper at 74 basis points of spread is now asked to mark it at roughly twice that level, with 17 of 23 comparable joint ventures already trading wide. Ares declined to compete for the paper that was repricing fastest, and its fee engine compounded anyway.

Nvidia's accounts show why it can afford to be the backstop. Trailing interest expense of $463 million sits against $192.9 billion of trailing net income on $302.97 billion of trailing revenue, and second-quarter fiscal 2027 revenue of $96.22 billion came with a 75.0% gross margin. The market's circular-financing worry is therefore not that the vendor is hiding losses. The vendor is quietly becoming the guarantor of the asset class it sells, which turns GPU financing into a vendor-balance-sheet decision.

The second-order effect runs through the supply chain. If the marginal GPU financing increasingly depends on the vendor's backstop, then GPU demand becomes a function of the vendor's credit decisions rather than of a private lender's appetite. That converts a funding cycle into a vendor-balance-sheet cycle, and it is why memory, foundry and power suppliers should be tracked on order-book sensitivity rather than on headline capital spending.

Upstream is being paid in cash today, so the risk is not that demand stops but that financing terms change first and orders follow. Downstream, the neoclouds and hyperscalers already carry the cost: CoreWeave's trailing interest expense of $1.87 billion sits against $7.59 billion of trailing revenue, and it paid $640 million of interest in a single quarter. At the far end of the chain, TSMC builds the accelerators and SK hynix supplies the high-bandwidth memory stacked alongside every financed GPU — and both now sell into order books whose pace depends on credit approvals.

The thesis resolves one of two ways. If GPU residual values hold, the depreciation curve becomes modellable, private credit re-enters a $3 trillion bucket, and the spread on compute-backed credit compresses toward the 74 basis points of a year ago — with the most price-sensitive lenders on the street suddenly underwriting the asset class Ares just passed on. If residual values disappoint, the $3 trillion stays financed by bank syndicates, vendor guarantees and neocloud balance sheets, and the loss is recognised in bank credit and equity rather than inside a $691 billion private credit manager.

So the question behind the headline has a precise answer. Bank balance sheets and vendor guarantees fund the AI buildout, while the private credit funds that built the AI lending playbook are paid to distribute other people's capital rather than to warehouse the collateral themselves. The risk that matters most is now the least visible — sitting in loan books, residual-value guarantees and non-bank private placements.