Research Matters by Rothschild & Co Redburn

The AI infrastructure boom is materially smaller and financially more fragile than announced capacity suggests, with sustainable returns dependent on utilisation, enterprise adoption and continued funding.

• More than 200 gigawatts of AI capacity have been announced, but roughly 100 GW backed by hyperscalers such as Amazon, Microsoft and Google appear more credible, while another roughly 100 GW faces greater funding uncertainty.
• Redburn estimates that reaching the first 100 GW will take several more years, with more than 50% of data centres delayed by at least three months and permitting a recurring bottleneck.
• Around 70% of current capacity is estimated to be used for model training rather than inference, creating an oversupply risk if training progress stalls or is paused for safety reasons.
• Funding conditions are already tightening: hyperscalers initially relied on operating cash flow and debt, but the next phase may require new financing structures and backstops from chip companies such as NVIDIA and Broadcom.
• Projected returns may overstate cash profitability because technology companies can fund AI businesses that then purchase their computing capacity; substantial debt refinancing is expected around 2030 and 2031, while enterprise adoption remains at an earlier stage and open-source models could pressure prices and margins.

Takeaway: Investors should distinguish announced from financeable capacity and projected from cash returns, focusing on funding access, refinancing risk, utilisation and broader enterprise demand rather than headline buildout figures.