#106 - Dinesh Ganesan - Arctern - Climate Investing in the Age of Compute

Okay, I think we've got a good understanding of what the energy underneath compute thesis is and how you're looking at it.

And I think we've got a good understanding of what climate investors should look for in that space.

Let's continue to go through the different layers.

What's the next step?

Yeah.

So then you then go inside the building.

So you go from the energy, which is powering all of this compute, all of this AI, and then you go into the power electronics, the cooling.

Grid power today gets converted through multiple stages before it becomes useful voltage at the chip level.

And so every loss becomes heat rather than compute.

And as rack densities rise, the cooling and the heat loss has to be managed well.

And so just to frame it, like data centers, I think the power requirement ranges from 5 to 50 megawatts.

A specific rack is like 7 to 10 or 7 to 15 kilowatts.

But anywhere from like 25% to 50% of the power does not actually go into actual computing capabilities.

So there's this term called PUE, power usage effectiveness.

It's the total facility energy divided by the energy used strictly for the compute.

And so PUEs today sit at like 1.5-ish.

And so that means that for every 100 watts delivered for compute or IT servers, an additional 50 watts are either consumed or lost.

But obviously, you have the best data centers, which has PoE of closer to 1.1, right? And I would say that a lot of the non-compute goes to cooling systems, power distribution, things like conversion losses, lighting, security, things like that.

So there's that.

And then you have also within AI, which is different from CPUs, you have these high spikes on compute.

So at a specific time, a specific minute, or a specific hour, you have heavy stress on the power requirements.

And that has a huge impact back to the grid, huge impact on all the cables and wires that you're dealing with.

So it's not just about energy efficiency around the cooling and the power conversion and bring it down to the chip level, but it's also the ability to dynamically manage energy.

And so that is one portion of that sort of second bucket as it relates to compute.

But then you can go even further, right? You can go then on the micro level interconnect.