A million GPUs sitting in a data center. Ready to go. And completely useless, because the local power grid is maxed out. That's the bottleneck no one saw coming. And it's why Microsoft just bought the entire output of a restarted nuclear plant for the next 20 years.

In 2026, the biggest bottleneck in artificial intelligence isn't chips or algorithms. It's energy. Data center power consumption is on track to double this year, and the US grid can't keep up. So Big Tech is going nuclear.

In this video, we break down the full nuclear AI energy stack: why NVIDIA's Blackwell GB200 made the power problem worse even as it got more efficient, what the FERC co-location ruling means for your electricity bill, how Small Modular Reactors (SMRs) and TRISO fuel work, and why the DOE wants states to host nuclear AI campuses. Plus, what developers need to know about joules per inference, carbon-aware scheduling, and why Rust and C++ are making a comeback.

This is the space race of our generation. The race for AGI is a race for atoms.

00:00 Microsoft Is Reopening Three Mile Island for AI
00:36 Why AI Hit an Energy Wall in 2026
01:03 NVIDIA Blackwell GB200 Energy Problem
02:24 The Secret Grid War: FERC Ruling Explained
03:42 Small Modular Reactors: Nuclear Power for Data Centers
05:13 DOE Nuclear AI Innovation Campuses
06:09 Carbon Aware Scheduling and the Developer Energy Stack
07:22 Why the AI Race Is Really an Energy Race

Tiff In Tech

Microsoft is reopening Three Mile Island to power its AI data centers, signaling a broader industry shift from algorithmic optimization to energy infrastructure. Big tech companies are hitting an 'energy wall' as data center power consumption is set to double, with GPU clusters rendered useless by grid capacity limits. The FERC issued a 2025 ruling allowing data centers to draw up to 900 MW directly from nuclear plants while sharing only 100 MW with the public grid. Companies are betting on Small Modular Reactors (SMRs) — factory-built, truck-deliverable units using advanced fuels like TRISO pebbles and molten salt cooling — placed directly next to server farms. The DOE is pushing 'nuclear AI campuses' that co-locate enrichment, SMRs, data centers, and waste reprocessing. For developers, this means new considerations: energy efficiency metrics (joules per token), carbon-aware batch scheduling via cloud APIs, and a resurgence of Rust and C++ for efficiency gains at scale.

Microsoft Bought a Nuclear Power Plant for AI

<p>Good. Nuclear power got demonized for political reasons and shelved because misinformation scared the public. Even if there are many other concerns and objections about Ai, returning nuclear to the front of the conversation can help us reduce or eliminate our overreliance on fossil fuels. And that’s a silver lining, even for the Ai skeptics and opponents.</p>
