Major hyperscalers (Microsoft, Meta, Oracle, Amazon) are shifting AI infrastructure capex off balance sheet through alternative financing structures as credit markets scrutinize the spending boom.

The AI capital cycle's latest evolution is not about spending less—quite the opposite. Money continues pouring into chips, data centers, power, networking, and cloud capacity at a pace that would have seemed absurd just years ago. What is changing is the route the money takes, and that shift may ultimately matter as much as the capex bill itself.

Amazon is reportedly exploring an arrangement that would move roughly $8 billion of Nvidia Grace Blackwell GPUs into a special purpose vehicle, allowing outside investors to finance and own the chips while Amazon leases them back. Simultaneously, Broadcom-linked financing for Anthropic and other AI customers is moving toward a $60 billion structure, with banks preparing a large senior secured tranche and Blackstone positioned further down the capital stack. Neither transaction signals that the AI boom is running out of money. If anything, both show how determined the ecosystem has become to keep spending accelerating even as conventional balance sheets grow heavier.

The shift toward SPVs, leases, guarantees, and private credit does not erase these obligations—it simply relocates them. It changes where they sit, who ultimately owns them, and how clearly equity investors can see them. Hyperscaler bond issuance is already exploding, and the most consequential portion of the financing stack may now be the piece sitting outside conventional corporate debt.

The true danger does not require an AI bust. It emerges if growth simply stops accelerating while financing structures built around scarcity, utilization, and generous residual values continue multiplying. The AI trade spent two years teaching investors to watch the chips. The next phase may demand watching the financing behind them.