> All economic growth is traded off against future global temperature increases to some extent, but this is even more acute in this electricity-intensive industry. How many degrees of temperature increase is worth one .. whatever the unit of AI gain-of-function is?

The premise here is that if we use more electricity then we burn more carbon. And the media hates AI, so if anybody restarts any coal-fired power plant to run a data center anywhere, that's the story. But then there's this:

https://electrek.co/2026/01/28/eia-99-of-new-us-capacity-in-...

Nobody actually wants coal because solar is cheaper.

And data centers are a pretty good combination for this because the biggest problem with solar and wind is what to do during multi-day periods of low generation, but data centers have backup generators and would be willing to turn them on whenever the cost of grid power is higher than the cost of operating them. Running some gas turbines for a week every two years in exchange for stabilizing the grid and being able to run on renewable power for the other 103 weeks is a pretty good outcome for everybody, not least because that amount of grid stabilization would exceed their consumption, i.e. allow more renewables to be added to the grid than they're using. If they can shed 1GW of load when a 2GW (long-term average) solar farm is generating at 50% of typical capacity for a week, you can add that 2GW of solar to the grid and remove 1GW of fossil fuels even while the data center is increasing consumption by 1GW.

one.. fully autonomous, self improving, replicating, general intelligence.

> How many degrees of temperature increase is worth one .. whatever the unit of AI gain-of-function is?

Billionaire. And they are definitely willing to make the trade.

My post argues that the demand has permanently shifted.

By the way, plenty of people on HN and Reddit ask if the demand is real or not. They all think there's some collusion to keep the AI bubble going by all the companies. They don't believe AI is that useful today.

I actually agree: a spike in prices due to bumping against capacity limits is way better than a downturn in the market. But this is only really true if AI hyperscalers are incented to space out their big buildouts over time (while raising their prices enough to ration current demand) so that suppliers can have some guarantee that their expanded capacity will be used.

Optical interconnect in the rack is a thing already. It's just a matter of time until it moves to single-PCB scale. And most persistent memories (competing with DDR memory for speed, and with far lower wearout than NAND) are of the "3D storage" type.

If we want reproducible output we already have conventional software. Stop using a hammer on screws.

So it's like humans then

This is wrong for all LLMs which have a temperature setting.

And even if there were guaranteed to be non-deterministic, there is still lots of value in many aspects of content generation.