Connect your Claude, Codex or OpenClaw for full transcripts on millions of podcasts. Two week free trial, cancel anytime.

Start free trial →
RB Raghu Belur On Volts

“If I save 5% in efficiency, do I take that 5% to save energy? I think that's a very good reason if you are power constrained, which in a lot of cases we are. But in the other case, it means I can add 5% more GPUs, which means I can generate more revenue. The amount of tokens per watt improves.”

Volts · Climate & Energy · November 2026

“If I save 5% in efficiency, do I take that 5% to save energy? I think that's a very good reason if you are power constrained, which in a lot of cases we are. But in the other case, it means I can add 5% more GPUs, which means I can generate more revenue. The amount of tokens per watt improves.” — Raghu Belur, Volts

Context

The conversation was about how simplifying the power chain in an AI data center frees up space and power. Belur says the savings can either cut energy use or go into more compute, which is how he frames the pitch to hyperscalers.

Listen to the episode → More from Raghu Belur → More Climate & Energy quotes →

Transcript

Volts Around 46:26 into the episode
Raghu Belur

…contains the battery backup unit, the power supply unit, the power distribution unit, and et cetera. So if you can simplify that entire powertrain, and if the end phase can show that, hey, you don't even need that local battery anymore, hey, that could be a huge advantage to the entire system.

David Roberts

Yeah, yeah, lots more compute in the same building.

Raghu Belur

Yeah. So the question comes: if I save 5% in efficiency, do I take that 5% efficiency to saving electrons, saving energy? I think that's a very good reason to do if you are power constrained, which in a lot of cases we are. But in the other case, it means I can add 5% more GPUs. Right. I can add more GPUs, which means I can generate more revenue. I can be more efficient. The amount of tokens per watt improves.

David Roberts

So hyperscalers need this and have a lot of money. So they're the first big market that SSTs are going after approximately. This is all semi-speculative. We're talking about pilots of this next year, maybe, and then the industry shifting to SSTs 28, 29, 30, something like that. But that's the big…

Raghu Belur

Yes, I'm far more bullish as an engineer. I think about it on a first principles basis. I go, why wouldn't you put something that's more reliable, more efficient, and lower cost? And so if you believe in the AI in the AI transformation, which I'm a believer in, I think you should expect this to…

David Roberts

Yes, but it's always the engineering mindset running up against the institutional conservatism of people with lots of money who have been doing things a certain way for a long time. It always goes more slowly than one might like. But one way or the other, these are about to move into data centers…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

More from Volts

“When you rely heavily on semiconductors and software, a few amazing things happen. If you look at smartphones, at laptops, they continue to improve in performance over time and they don't asymptote in cost, unlike copper and iron, where they become incompressible. Copper and iron is copper and iron. I buy them by weight.” — Raghu Belur, Volts “If you look at old data centers, per rack, which is a seven-foot telecom rack, it used to be 30 kilowatts. Today they went to 150. Today we are already at 250 kilowatts. And now we are talking about getting them up to a megawatt. If you stay at 50 volts and it's a megawatt, you're looking at 20,000 amps. … That's just not feasible. So the push is to go to 800 volts DC, so we can drop the amount of current delivered to each server by almost 20x.” — Raghu Belur, Volts