Hi Dotikers!

Everyone loves to talk about how AI needs more electricity. More gas plants, more nuclear, more solar farms. But the most interesting problem right now is not how much power we generate, it is how badly we handle it once it arrives at the data center door.

A recent piece in MIT Technology Review makes this point convincingly. The grid failures we saw in Virginia, where dozens of data centers disconnected almost simultaneously and nearly destabilized the regional network, were not caused by a shortage of megawatts. They were caused by architecture. AI training workloads swing from near idle to full throttle in milliseconds, sometimes by gigawatts at a time. Traditional data center power systems, built around UPS units designed for steady, predictable enterprise loads, simply were never meant to absorb that kind of electrical whiplash. Asking a legacy UPS to smooth out an AI training run is a bit like asking a garden hose to handle a fire truck.

The proposed fix is elegant: move power infrastructure to medium voltage, bring it closer to the substation, and place it inline so it absorbs the chaos internally and presents a calm, stable load to the grid. Done right, this could improve efficiency, shorten permitting timelines and even turn data centers into grid assets rather than grid liabilities.

This is exactly the kind of unglamorous engineering the AI industry needs to take seriously, and fast. The race for compute has been dominated by chips and models, while the electrical plumbing underneath has been treated as someone else's problem. It is not. If the industry keeps bolting hyperscale AI onto power architectures from another era, the next Virginia incident will not be a warning, it will be a headline about a blackout. The companies that solve power architecture will quietly win as much as those who ship the next frontier model.

Aym.