unknown

Major tech companies are investing billions in nuclear power to fuel AI infrastructure—Meta signed a 20-year deal for 1,121 MW, Microsoft is restarting Three Mile Island for $1.6B, and Amazon invested $1.15B in nuclear technology. This represents a fundamental failure in optimization: instead of making AI models more efficient, Big Tech is scaling power infrastructure. Meanwhile, DeepSeek trained a competitive 671B parameter model for $5.5M, Berkeley replicated OpenAI's reasoning model for $450 in 19 hours, and Stanford did it for $50 in 26 minutes. The real competition isn't about who has more compute power—it's about who can achieve the same results with dramatically less energy through smarter engineering and architectural decisions.

Big Tech’s Nuclear Play

Leif Latiff

“We have no idea how to make our technology more efficient. So we’re just going to throw a nuclear reactor at the problem.”

“We have no idea how to make our technology more efficient. So we’re just going to throw a nuclear reactor at the problem.”

Very interesting. Do you think a new, 10-100x more efficient model could get this massive AI-bubble to pop?
The big few seem like they’re too big to fail. Or are they?

We need more energy anyway. That will push nuclear development too

I think one of the next steps we may start seeing is “model-as-a-service” and “Model Hosting + Federated Fine-Tuning” as a managed service.
Things like this will be needed to address similar SPoF AI model issues like we recent seen with AWS and Azure outages.
I am currently working on an article/report to talk about this.

Billions on power vs. thousands on smarter design. The real breakthrough isn’t scale; it’s efficiency.

nah, to get the best you have to max out on all fronts.
<a href="https://epoch.ai/blog/can-ai-scaling-continue-through-2030" target="_blank" rel="noopener nofollow">https://epoch.ai/blog/can-ai-scaling-continue-through-2030</a>