AI Will Save Us All (Not How You Think)

AI might be the only force we’ve ever wanted badly enough to finally build hybrid clean energy systems at scale.

The Irony Is the Point

We’re not going to convince people to build nuclear to save the coral reefs. We’ve been trying that for decades. It doesn’t work.

But build nuclear so ChatGPT doesn’t crash? So data centers don’t brown out? So AI investments don’t become stranded assets?

That might actually work.

The irony isn’t lost on me. The technology that keeps me up at night might be what finally breaks the political gridlock on clean energy. Not because people suddenly care about climate, but because trillion-dollar companies need reliable power, and they need it clean.

Why This Matters Now

Data center power demand could grow 160% by 2030 (Goldman Sachs Research, 2024). US power demand has been essentially flat for a decade as efficiency gains offset growth. That’s over. AI just broke it.

The math:

Nuclear runs 93% of the time. Wind produces power about a third of the time. Solar about a quarter. Data centers need power 24/7/365. You can’t run continuous operations on 25% availability without massive storage or backup generation.

Source: EIA Electric Power Monthly, 2024

Some argue “baseload” is an outdated concept. AI just brought it back. Data centers run flat. There’s no peak to shave.

I know this because I recently co-authored a national standard for location-based hourly carbon emission accounting (RESNET/ICC 1580). When you analyze grid emissions hour by hour, the pattern is clear: baseload hours, the middle of the night and the dead of winter, are when we’re still burning coal. Using NREL’s Cambium database, we can see exactly when and where marginal emissions are highest. It’s not peak demand. It’s baseload. That’s exactly when decarbonization matters most.

It’s Not Nuclear vs. Renewables

Most of this debate goes wrong because people treat it as a binary choice.

I’ve published research on hybrid energy systems (Kelly et al., The Electricity Journal, 2014). The answer isn’t nuclear OR renewables. It’s intelligent systems that match each technology’s strengths to actual grid needs. Solar and wind are cheap and fast to deploy. Nuclear provides reliable baseload. Storage bridges the gaps. We already have tools like virtual power plants and demand response that shift load to match supply, but they can’t conjure electrons that aren’t there.

The tech companies understand this. Microsoft, Google, and Amazon have signed vastly larger renewable PPAs than their nuclear deals. Nuclear represents maybe 5% of their clean energy commitments. It’s a hedge, not a replacement.

But it’s a necessary hedge. Because when you need power at 3 AM in January and the wind isn’t blowing, you need something that runs 93% of the time.

The Tale of Two Countries

France gets 67% of its electricity from nuclear. Grid intensity: ~85 grams CO2/kWh.

Germany spent €400+ billion on its Energiewende (renewable transition) while closing nuclear plants. Grid intensity: ~350 grams CO2/kWh.

That’s 4x worse. Germany’s grid emissions have declined, but nowhere near as fast as France achieved decades ago with nuclear.

The lesson isn’t “nuclear good, renewables bad.” The lesson is: don’t close your clean baseload before you’ve replaced it with something equally reliable.

Yes, Nuclear Has Problems

I’m not going to pretend otherwise.

Cost: Vogtle Units 3 & 4 came in at $35 billion instead of $14 billion. That’s 2.5x over budget, 7 years late.

Failed projects: The NuScale SMR project in Idaho was cancelled in November 2023 when costs rose to $89/MWh.

Timelines: Microsoft’s Three Mile Island Unit 1 restart targets 2028. Google’s Kairos deal won’t deliver power until the early 2030s. None of this is fast.

Technology details: Kairos Power is building fluoride-salt-cooled high-temperature reactors (FHRs) with TRISO pebble fuel, often incorrectly called “molten salt reactors.” The salt is coolant, not fuel carrier. This matters for understanding waste profiles.

So why am I still optimistic?

How Money Changes Things

The tech companies aren’t ignoring these problems. They’re routing around them.

Buying existing plants: Microsoft’s TMI deal restarts a plant that already exists. No construction risk, just restart and regulatory approval.

Risk-sharing PPAs: Google’s Kairos deal is a power purchase agreement. Google pays for power when delivered. Construction risk stays with Kairos.

Building where power is: Amazon bought a data center campus next to a nuclear plant in Pennsylvania. Skip the 5-year interconnection queue entirely.

When trillion-dollar companies want something, they find ways. The question is whether their motivation is enough to rebuild an atrophied nuclear supply chain.

The Global Stakes

Tech companies buying nuclear for their data centers doesn’t directly shut down coal plants in India. But rebuilding the global nuclear workforce does.

The US has spent decades letting its nuclear industrial capacity atrophy. We’ve lost welders, engineers, manufacturing expertise. Restarting that engine, proving that nuclear projects can be completed, that SMRs can work, that the economics can pencil out, creates capacity that travels.

Supply chains are global. Expertise is exportable. When South Korea builds reactors in the UAE, they’re drawing on decades of domestic experience. When China deploys nuclear across Asia, they’re leveraging an industrial base they never let die.

Someone has to go first. The tech companies are volunteering.

Because lead times are long, starting today matters more, not less. India building nuclear in 2035 requires someone rebuilding the industrial base now.

The Timeline Paradox

The honest problem: climate physics says we need emissions cuts by 2030. Nuclear projects deliver in the 2030s.

My answer: we need nuclear for the 2035-2050 period, while renewables and storage handle the near-term. This isn’t either/or. It’s sequencing.

Battery storage costs have dropped 90% in a decade (BloombergNEF) and are still falling. Grid storage is growing 25-30% annually. By 2035, these resources will be vastly more capable than today. But we’ll still need baseload for the loads that can’t flex, and nuclear ages better than the gas plants we’d build otherwise.

The Path Is Open

Given the timeline physics gives us, I’m betting on the fastest path. Right now, that runs through markets.

AI is a capitalist juggernaut that’s going to demand clean power whether we like it or not. The tech companies are buying nuclear alongside renewables, not instead of them. They’re investing in grid-scale storage. That’s the hybrid approach, and it actually works.

The timeline is slower than I’d like. The costs are higher than they should be. Most utilities are still defaulting to gas.

But someone has to prove the model. The tech companies are going first.

The thing that keeps me up at night might be what finally lets me sleep.

Sources: IEA (2024), Goldman Sachs Research (2024), EIA Electric Power Monthly (2024), BloombergNEF, Lawrence Berkeley National Lab interconnection queue data (2024), European Environment Agency grid carbon intensity, corporate announcements (Microsoft/Constellation Sept 2024, Google/Kairos Oct 2024), Georgia PSC filings (Vogtle), UAMPS (NuScale cancellation Nov 2023).