Kimber: AI inference will ultimately draw far more power than training
Sheldon Kimber · The case for colocating data centers and generation · Feb 13, 2025 · at 44:04
Sheldon Kimber, CEO of Intersect Power, discusses the long-term energy drivers of artificial intelligence.
0:00 / 0:09exact quote · 9.4s
720p mp4 · rendered on demand · StarZero watermark
“I think now what we're seeing is that inference in the aggregate is actually probably a much larger draw than the training in the long run.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Sheldon Kimber
Kimber: Data centers are entering an era where grid connection is optional
“Generally speaking, I think we're moving into an era where grid connection is optional and should be optional. And that, you know, we should be able to, you know, operate these things fully off grid if necessary.”
Kimber: Hybrid renewables and gas engines beat new combined-cycle plants
“If you go build a CCGT right now for 2000 dollars KW and, you know, like you can't even find sort of like You know, labor unions and pipe fitters and whatnot that have ever even built these things in many cases, right? You know, it's going to take five, six ye…”
Kimber: AI data centers will never offer flexible power load
“Every second you aren't running that at a hundred percent capacity factor is enormous amortized cost, right? So that it's just, it's one of those things where it's like, you'd be a fool to turn it down. You have to at some points because operationally, you kno…”
Kimber: Edge LLMs trigger recursive queries, massively multiplying datacenter power demand
“Every smaller distilled LLM that is power efficient doing inference is going to launch thousands of queries to bigger, more capable models that can actually reason their way through buying something on Amazon and reason. So it's almost like increasing the worl…”
Kimber: Intersect pivoted clean power assets from hydrogen to AI
“We thought maybe hydrogen would be first because of the supply side supports under the IRA. The demand just didn't materialize, and now all of a sudden, AI has just swept up all of the assets that we developed for, you know, behind the meter And co-located loa…”
Kimber: Co-located renewables can power Texas data centers carbon-free 75%–80% of the year
“75 to 80% of the hours of the year, we can run fully carbon free with, you know, wind, solar, and batteries all on the same site. And then, you know, 20%, maybe a little over 20% of the time, we need to have some gas or we need to use the grid.”