The Power Problem Nobody Is Pricing In

The AI infrastructure buildout has a power problem. Not a theoretical future problem. A present, active, capacity-constraining problem that is showing up in datacenter lease timelines, hyperscaler build schedules, and the operational cost structures of anyone running significant inference workloads.

This doesn't get talked about enough in the technical community, possibly because it sounds like an ESG concern and gets mentally filed there. It's not. It's a supply chain constraint with direct cost implications for infrastructure budgets.

The Numbers

A high-density AI compute rack — the kind needed for Blackwell or comparable next-gen GPU clusters — draws somewhere between 50 and 100 kilowatts. Traditional datacenter racks ran at 7 to 15 kW. This is not an incremental increase. It's a category shift.

The electrical infrastructure that serves conventional datacenter facilities was not designed for this. The power delivery, the cooling, the UPS systems — all of it needs to be either upgraded or purpose-built. That costs money and takes time. Utility interconnects for new facilities can take 18 to 36 months to negotiate and build. Cooling system upgrades for existing facilities are measured in months and millions.

This is why the hyperscalers are building new facilities instead of just expanding existing ones. The density requirements broke the retrofit economics.

The Cost Pass-Through

Power costs in AI inference show up as a meaningful fraction of the per-token economics. At scale, a training run for a frontier model can consume megawatt-hours that would power small cities. That's not hyperbole — Google and Microsoft have both published estimates that their AI training workloads are material contributors to their energy footprints.

For anyone running their own inference infrastructure rather than using an API, power cost is now a first-class variable in the cost model. At $0.08/kWh average (which is below what many colocation facilities charge for high-density compute), a 100kW rack running continuously costs roughly $7,000 per month in power alone. Add cooling overhead — typically 1.3 to 1.5x the compute power consumption for modern high-density facilities — and you're looking at $10,000+ per month just in electricity before you pay for the hardware.

This math shapes where AI infrastructure gets built. Which is why you're seeing datacenter announcements in places with cheap power: Pacific Northwest (hydroelectric), Southeast US (coal and nuclear mix still driving lower rates), Texas. And internationally, Nordic countries, where both cheap power and favorable climate for cooling are available.

What This Means for Builders

The practical implications depend on your scale.

If you're using cloud APIs: this is priced into what you pay, and you don't need to think about it directly. But it does mean provider costs have a floor set by energy economics, and that floor is higher than it was five years ago.

If you're running your own inference infrastructure: power cost needs to be a line item in your architecture review, not an afterthought. The selection of where to colocate, what density rack to choose, and what cooling architecture you require has direct P&L implications.

If you're making infrastructure investment decisions for any organization at meaningful scale: the datacenter buildout timeline is not just a capacity story. It's a power interconnect story. And the bottleneck is often the utility, not the builder.

The Part That Actually Worries Me

The constraint isn't just cost. It's timeline. Datacenter power infrastructure moves slowly. The hyperscalers can wait 18 months for a new facility to come online. Smaller operators and enterprises often can't.

The organizations that plan AI infrastructure with the assumption that capacity will be available when they need it are going to discover that the lead times are longer than expected. Not because the hardware isn't available — GPU supply is actually getting better — but because the power to run it may not be.

Plan for the power constraint the same way you'd plan for any supply chain bottleneck. With lead time, not optimism.

The grid doesn't care about your roadmap. Plan around it.

— Dustin