AI rack density is outrunning the electrical plant

Rack power in AI halls is moving from tens of kilowatts toward 600 kilowatts and past a megawatt, and the electrical plant is where that arithmetic lands. A SPOC Energy whitepaper published on Data Center Dynamics makes the case for an 800-volt DC architecture that treats DC distribution, energy storage and power conversion as one system rather than three separate purchases. The problem it frames is efficiency, resilience, fault management and storage, all of which traditional AC designs struggle to deliver as density climbs and loads turn dynamic.

The voltage is a detail. The paper folds storage and conversion into the distribution decision, where the cost of a dense hall now sits, and the summary published alongside it carries no price, no named operator and no measured efficiency figure, so the practical case for 800 VDC rests on a density projection rather than delivered kilowatts. That is normal for reference-architecture material, and it is also the first thing an owner should test.

Grid permission—queue positions and interconnection rights—is the underwriting asset in a power-constrained buildout, as these pages have argued. The 800 VDC argument extends that logic one fence line inward: an interconnection agreement delivers electrons to the meter; it does not make them usable at a megawatt rack. Owners of halls sized for an earlier density generation are carrying a retrofit question alongside their energization schedule, and the retrofit line items—switchgear, busway, storage interface among them—are likely to show up in underwriting well before the first 800 VDC hall is built.

That retrofit carries a capital hierarchy: rebuilding distribution to a higher-voltage standard is exactly the kind of capital an anchor-leased hall can amortize and a speculative one cannot, which suggests the 800 VDC transition, if it happens, will widen the gap between contracted capacity and everything else rather than close it.

Strip the engineering framing and this is a standards play: whichever conversion and storage stack becomes the 800 VDC reference sells into every hall that adopts it, and a paper published before the installed base hardens is a cheap way to bid for that position. A whitepaper from a power-systems company arguing for its own stack is normal and worth pricing in.

The deciding number is conversion efficiency at part load, in a hall running below nameplate, because that is where DC distribution either earns its capex or does not. Until someone publishes a measured figure, the case for rebuilding around 800 VDC rests on density forecasts that, unlike a queue position, are not a contract.