2 Comments
User's avatar
Rohan Jaiswal's avatar

The Applied Digital 1.2 GW contracted plus 1.7 GW available number is the line that stopped me. Running theaifounder.substack.com pulls me into builder-side worries, and the inference cost curve depends on whether TPU-based and modular CDU stacks like LiquidStack's 14 MW unit hit price parity with the NVIDIA-anchored sites. The $7.5B to $18.5B option ladder on a single 300 MW lease suggests hyperscalers are paying optionality premiums rather than commodity prices. What does the secondary market for unused option capacity look like once one of these tenants pulls back?

Rich Miller's avatar

Hi Rohan. Thanks for the comment. We don't yet have precise data about the scenario you describe. But it's worth noting that we are in the midst of a capacity crunch for wholesale data center capacity, with vacancy rates at about 1 percent. There have been several instances in recent years when a tenant has backed out of a large hyperscale deal, and it appears others quickly stepped up to take that capacity. Going back to cloud precedents, the ability to resell vacated space has been surprisingly strong in major markets.

Right now hyperscalers are glad to pay that optionality premium because many of them are currently capacity-constrained, and none of them like it. So the recent trend has been that large players are less worried about overprovisioning than running out of compute. If demand falters, those options for future space might look different. But if agentic AI continues to push demand, planning ahead will likely seem like a good bet.