Power is the new compute constraint

The bottleneck moved outside the server hall

For the first wave of generative AI, the scarce asset was the accelerator. Operators competed for allocation, accepted long delivery windows and optimized clusters around the hardware they could obtain. As supply expands, the limiting factor moves upstream: a site must secure firm electricity, transform it to the required voltage, manage harmonic loads and remove the resulting heat.

A data center with reserved equipment but no energization date is not productive capacity. That makes utility coordination, substation design and grid interconnection part of the compute product rather than background real estate work. The relevant metric is no longer announced megawatts. It is usable megawatts delivered on a credible schedule.

Four power questions that matter

A headline power figure can hide the conditions that determine whether a project works. Investment analysis should separate the physical connection from the commercial right to consume energy and from the redundancy needed to serve contracted workloads.

• When can the site be energized, and which milestones are controlled by the developer rather than the utility?
• Is capacity firm, interruptible or dependent on future transmission upgrades?
• What redundancy level is required, and how much capacity is stranded by that architecture?
• Can the load participate in demand response without breaching customer service levels?

Location economics are being rewritten

Latency still matters, but not every AI workload needs to sit next to a major population center. Training, batch inference and model evaluation can move toward power-rich regions if network connectivity is adequate. This widens the investable map while increasing the importance of transmission, fiber routes and local permitting.

The best site is therefore not simply the one with the cheapest electricity. It is the site that combines a dependable energization path, a stable tariff framework, sufficient fiber, available water or an alternative cooling design, and a construction ecosystem capable of delivering at cluster speed.

Where durable value can emerge

Scarcity favors assets that shorten deployment time: entitled land with a real interconnection position, modular substations, transformers, switchgear, on-site generation, energy storage and software that orchestrates flexible loads. These are less visible than the GPU, but they decide whether expensive compute can produce revenue.

The investment lens should follow the critical path. If power delivery determines the opening date, then schedule certainty and execution capacity deserve the same attention as chip performance.