AI's Real Bottleneck Is Power, Not Chips

You have probably heard that AI is stuck waiting on GPUs. The newest numbers point at a different wall: electricity. The chip shortage is no longer the thing holding data centers back.

AI capital spending has reached $750 billion, and more than 23 gigawatts of data center capacity is under construction worldwide right now. What is slowing it down is high-voltage transmission and grid permits, not silicon.

What Happened

The constraint has moved off the chip and onto the power line. It is not about GPUs anymore, and it is not about data — it is about high-voltage transmission and getting grid permits approved.

Amazon, Meta and Microsoft have started contracting directly for modular nuclear reactors. In practice, they are turning into energy utilities rather than software companies, building fission infrastructure.

• One AI data center draws 100-200 MW.
• Modular nuclear, or SMRs, is the option they treat as a stable power source.
• 1 GW of capacity powers about 1 million households. The 23+ GW in the pipeline works out to roughly 23 million households.
• The public is still largely anti-nuclear, so permits move slowly.

Why It Matters

Where AI gets built is now an energy decision as much as a hardware one. If you work anywhere near energy, grid planning or data centers, the contracts worth watching are the power deals, not the model releases. A company that can lock in steady generation gets to keep building; one that cannot waits in a permit queue like everybody else.