Tech giants shift $120bn of AI data center costs into off-book SPVs

Moving more than $120 billion of AI data center build-out into opaque SPVs is the financial mirror of the model-scaling race. The hyperscalers and frontier labs are telling the market they intend to keep ramping compute even if it means adopting the kind of structured finance that fueled earlier credit booms. In the near term, this absolutely accelerates the availability of GPU capacity for training larger models and deploying more agentic workloads—Wall Street is effectively fronting the capex in exchange for long‑dated lease payments.

The systemic risk is that we’re now tying global AI progress to complex, illiquid financing structures. If AI monetization disappoints or the cost of capital spikes, these obligations don’t vanish; they migrate into private-credit vehicles and securitizations that are hard to unwind cleanly. For the race to AGI, that means timelines are being pulled forward on the back of financial engineering. We may get more compute in the 2025–2028 window than organic cash flows would justify, at the cost of a nastier correction later if expectations prove too optimistic.