[LA:SA] 10GW AI training run before 2029?

In "Situational Awareness: The Decade Ahead", Leopold Aschenbrenner predicts that the largest AI training clusters will consume 10GW of electricity in ~2028.

This market resolves YES if a training run of a single AI model consumes 10GW+ of power sustained through most of the training run. This power cost includes overhead to run the data center, such as cooling.

This is one of a series of markets on claims made in Leopold Aschenbrenner's Situational Awareness report(s):

Other markets about Leopold's predictions:

Get Ṁ600 play money
Sort by:

bitcoin mining already uses like twice that much. Less of an infrastructure challenge if they figure out a way to efficiently parallelize the training across sites. I don't have a good inside view of whether training parallelism is likely.