Also llms in many ways will lead to more sensible work arrangements, the generative video models are going enable dopamine hijacking apps like we can’t even begin to imagine (but will also be useful for training and control of robots)

key environmental culprit will likely end up being generative video models that will demand huge amounts of memory and storage in addition to being very resource intensive, and it is less likely that breakthroughs significantly reduce the energy footprint of diffusion models relative to llms, imo

Also i think the reasonable assumption is that most of the marginal increase in compute in the next 5 years will largely be attributable to AI.

I think the conclusion of “can we live with ai” is yes, but there are a few things here that are quite… optimistic?

Caterpillar doubling gas turbine production means that these turbines are going to power data centers and will be burning gas likely for decades after they are deployed.

So coming back to this, if currently there is about 415 TWh of demand, the and the projections are that the tech industry will build about 32GW more, this is 280320 GWh or 280 TWh, expansion in next 5 years.

E.g. not quite doubling all compute on earth, but more than 50% increase.

People don’t always win…

What you are seeing now is all the tech companies are making big contracts with “neocloud” providers which are people who have spent last 5 years setting up build bitcoin mining operations, that are switching over to running ai workloads cause it pays better.

Sorry just to be clear, this will almost certainly be more than a doubling of *compute* but less than a doubling of energy draw

The primary trouble is that a shit ton of this is going to be gas, nuclear on this scale i guess is a decade out at least, and places like Google have already spent the last decade trying to find all the good renewable sources to collocate with, so the pickins already a bit slim

So coming back to this, if currently there is about 415 TWh of demand, the and the projections are that the tech industry will build about 32GW more, this is 280320 GWh or 280 TWh, expansion in next 5 years.

E.g. not quite doubling all compute on earth, but more than 50% increase.

www.lockportjournal.com/news/local_n...

www.btpm.org/local/2025-0...

Whoops

Like oops we decided to roll in generators on containers that pollute a bunch cause it’s fast and convenient but the neighbors hate it.

www.tomshardware.com/tech-industr...

One trouble is i guess they want to have reliable uptime. Unless we figure this out super fast, they going to buy a shit ton of gas generators which is going to cause a fracking boom

Just remember sir. You can count on coal.

www.countoncoal.org/2025/07/the-...

If only we were so lucky

Hmm thanks. This doesn’t sound right. I think here they are reporting total number of megawatt hours used per year, whereas, e.g. Rick Perry kicking off the construction of a 6GW power plant in Texas is going to output 6GW per hour??? E.g. 2x as the energy from the hydro station at Niagara Falls

It’s satya nadella’s fault to no small degree. Google will obviously win the race, but openAI pushed them to release too early and kicked off this unconscionable arms race. He even bragged “remember who made them dance”. I hope people do when the fracking and tar sands start popping off.

It feels meaningfully different than 2008 and perhaps more akin to the ghost cities that they built in china where they put up a bunch of skyscrapers and infrastructure in hopes of people moving in, and created financial collapse when they didn’t?

Ooops, will not do anything….

The next few years of this madness seem to already be locked in. Many of these projects have broken ground