Innovation Heroes: AI power demand is going nuclear. That’s a good thing

Every IT leader has a number in their head. Budget. Headcount. Timeline. But there’s a new number that matters more than any of them — and most organizations haven’t done the math yet. 

“If you think about a rack in a traditional data center, it was probably in the range of 20 to 40 kilowatts per rack,” says Ryan Hotchkin, SHI’s Sr. Director – Datacenter & Business Management Operations. “If you look at Nvidia’s GB200s, the NVL72s, now we’re talking about 120 kilowatts per rack. What Nvidia is doing is every year they’re doubling the power per rack.” 

That’s the infrastructure reality facing every organization with AI ambitions. And it’s exactly what host Ed McNamara, Brian Smith of Idaho National Laboratory, and Hotchkin explored on the latest episode of Innovation Heroes.

1. The power wall is already here

For years, computing felt elastic. Need more capacity? Spin it up. But AI has changed that equation by tying performance directly to physics. Power in, heat out, no exceptions. 

The challenge isn’t just electricity. It’s everything that comes with it: cooling systems designed for last-generation density, buildings with weight limits never tested by liquid-cooled racks, and power infrastructure back-ordered for years. 

“The problems that are coming up today are problems I’ve never seen before,” Hotchkin said, “and we’re all having to learn a lot of new skills that we’ve never learned before.” 

Even organizations buying cloud capacity aren’t immune. Hyperscalers are making massive infrastructure investments, and those costs flow downstream to every customer. 

2. Nuclear power is back — but not how you remember

When Brian Smith talks about nuclear power, he’s careful to separate perception from reality. Today’s innovation isn’t massive cooling towers or decade-long construction. 

“This is almost like building airplanes, not airports,” Smith said. “Rolling something off of a factory assembly line, then sending it out to its deployment location.” 

Small modular reactors are designed to be manufactured, shipped, and deployed incrementally. Google has already signed a partnership with Kairos Power for 500 megawatts of nuclear capacity by 2035. The hyperscalers aren’t betting on nuclear because it’s futuristic — they’re betting on it because it’s reliable, carbon-free baseload power at a scale renewables can’t match. 

Smith expects the first commercial SMR deployment by 2030, possibly sooner. But he’s direct with data center audiences: a microreactor going critical at INL next year won’t solve the tens of gigawatts the grid needs. It’s a step, not a solution. 

3. The stakes are career-defining

This isn’t abstract infrastructure planning. For IT leaders, it’s personal. 

“If I’m an IT director and I go buy a $10 million AI solution and I wrap power and cooling around it, and I don’t get the ROI back from it and I don’t future-proof my organization, I’m going to get fired,” Hotchkin said. “These are so expensive and it is potentially a career-limiting decision.” 

The window for planning is now — even if deployment is years away. Organizations that wait for perfect clarity will find themselves locked out of capacity, priced out of cloud, or stuck with infrastructure that can’t scale. 

“If you’re not thinking about it today, if you’re not planning for three to five years from now, you’re gonna miss it,” Hotchkin said. 

NEXT STEPS

The companies that understand AI strategy is now infrastructure strategy won’t just survive the transition. They’ll be the ones who scale while others stall. Listen to the full conversation here to understand the collision between AI ambition and energy reality, and what steps you need to take to succeed.

You can also find episodes of the Innovation Heroes podcast on SHI’s Resource Hub, Spotify, and other major podcast platforms, as well as on YouTube in video format.

Video + audio

Audio only