Cloud backlog far exceeds capex and total debt for all 3 major public cloud providers. But what if planned data centers can't be built and backlog can't convert? Or, what if all the data centers hyperscalers have agreed to lease DO get built and connected? Special purpose vehicles are borrowing against the credit of Amazon, Microsoft, Alphabet and Oracle to finance data centers for these cloud customers. As hyperscalers take possession, these leases will start to appear on their balance sheets.... will earnings grow fast enough to keep debt ratios the same?

​

This week Oracle sent a force majeure letter to Stack Infrastructure, a Blue Own Capital subsidiary that is building a data center for Oracle and OpenAI in New Mexico as part of the Stargate project. Oracle's management is concerned about making lease payments before it can earn any revenue from the data center. Revenue is delayed because Stack Infrastructure can't power the data center. The team is not waiting for a grid connection but will instead use natural-gas powered fuel cells made by Bloom Energy. But they can't get the natural gas because New Mexico has rejected their plan toi build a pipeline through state-owned lands. They have a backup plan, but it involves building on land administered by the federal government.

​

Oracle is not the only hyperscaler struggling to find capacity to convert backlog to revenue. Microsoft has chronic capacity problems, and AWS and Google Cloud both have to balance their backlog commitments with the needs of their own cloud-based businesses, which provide the earnings that allow them to keep borrowing for AI builds. 

​

With communities across the country pushing back against data centers builds, and grid/transport infrastructure tapped out in many regions, data center build timelines are lengthening. At the same time, development times for AI models and AI chips are compressing. Data center operators worry about designing facilities for AI accelerators that will be obsolete by the time the facility comes online. And enterprises using AI worry about their token bills. Increasingly, they are turning to smaller, open weight models and to hardware that can run these models on-prem, bypassing data centers altogether.