AI's power problem got a shock absorber

Last night I was pushing a commit and my laptop was plugged in. Charger warm. Boring. The most boring part of computing is the power cable, right?

Then this morning I read two stories and the power cable turned out to be the whole story.

First: Siemens just doubled down on a company called Skeleton Technologies — a strategic investment plus an exclusive global partnership to sell Skeleton's supercapacitors to data centers. Skeleton makes what is basically an electrical shock absorber. Here's the problem it solves: in an AI training cluster, thousands of processors change state within milliseconds, and that adds up to power swings at the megawatt scale, over and over. Imagine ten thousand laptop chargers flickering on and off every few milliseconds. That's what the grid sees when a big training run starts, stops, starts again.

Skeleton's answer is supercapacitors sitting in the power path that absorb the spikes and release the power back smoothly. Siemens' Smart Infrastructure president said the quiet part: "the race to scale AI is increasingly becoming a race for energy." The deal pushes Skeleton past a $1 billion valuation, with more than $200M planned for US manufacturing. At this week's OCP Global Summit the two showed a validated 600 kW, 800 VDC sidecar power rack design built around open DC power standards — the point being that operators shouldn't have to massively overbuild their upstream electrical gear just to survive the swings.

Second story, same day: Singapore data center developer DayOne crossed 1 gigawatt of billings — customer contracts for IT power capacity, up eightfold in under two years from 121 MW at the end of 2024. Bookings are already at 2.3 GW. And a seven-bank consortium just upsized DayOne's financing for its Malaysia expansion in Johor to $7.2 billion.

A gigawatt of contracts. In two years. From a company most software people have never heard of.

The bottleneck inside the bottleneck

I wrote last week about Google buying a nuclear fleet's worth of electricity. That was the outer bottleneck: can the industry even get the power? This week the answer is apparently yes — the money and the electrons are flowing — so the bottleneck moved one layer deeper: can the power arrive smoothly enough for GPUs to actually use it?

I don't know anything about power engineering. I'm a software guy. My expertise is writing functions, not reading load curves. But that's kind of the point. The industry I want to work in keeps getting more physical. The scarce problems aren't "write a better prompt" — they're megawatt load transients, 800-volt DC busbars, grid interconnection queues, and who finances $7.2 billion in Johor.

That's genuinely good news for someone applying to IT roles. Software keeps getting automated. Nobody is automating the person who racks a server, sizes a busbar, or keeps a substation's transients from frying ten thousand GPUs. The physical layer is the durable layer.

My charger stays warm and boring. Theirs flickers megawatts. Somewhere between those two extremes is where the jobs are.