This chart captures the point at which data centre debt decoupled from cash flow in the AI buildout. After a decade where hyperscalers funded infrastructure from their enormous profits, 2025 marks somewhat of a watershed in the race to stay competitive in computational scale.
Hyperscalers generating $700 billion in operating cash flow are already spending $500 billion on AI, now leaving little for continued expansion. Their solution? Special purpose vehicles that keep debt off balance sheets. Meta’s $30 billion Hyperion facility sits in an SPV with Blue Owl, transforming capital expenditure into operating leases. These “debt mules” now carry obligations that parent companies can theoretically walk away from.
Adding to the precarity, Michael Burry has made the headlines this week with his Palantir Big Short 2.0 and accusations that the hyperscalers are extending depreciation schedules to 5-6 years for GPUs that can burn out in 18-24 months running intensive AI workloads. When debt servicing relies on profits against understated depreciation, the financial engineering becomes more critical than the technical innovation it funds. Technology, real estate and complex debt… what could possibly go wrong? But, as we’ve said before, no one can stop spending, or now borrowing, because what if a competitor achieves AGI first?