Remember when we didn't have enough electricity for electric cars?

Question for the experts: does the power crunch mean that AI hyperscalers will turn off previous generation GPU datacenters to free up power for their new Vera Rubin GPUs?

Project Kilby is probably the most intelligent approach to this problem so far.

https://www.chevron.com/newsroom/2026/q2/chevron-signs-20-ye...

The idea is to bring the data centers, power generators and energy supply together in the ~same physical space so the only thing you have to transmit is data. Moving energy is way more expensive than moving information.

The primary bottleneck to this growth is the availability of electricity.

The bottleneck for building some AI datacentres and switching them on is electricity, sure, but that's not what drives growth. There also needs to be demand for the additional capacity; people need to be waiting for capacity to catch up so they can do the useful work that grows [society|GDP|something] that they aren't doing right now.

There's also very likely to be diminishing returns from additional capacity if we're near or over the limit of productive use. And there's the opportunity cost of what could have been done with that [money|land|electricity].

This is a much more complicated system than "people say they need more AI -> build datacenter -> power datacenter -> magical growth!"

TFA isn’t about consumer usage, it is about training the next generation models. An interesting thought I heard recently is that a SOTA model has about the same parameters as the number of synapses in a Golden Retriever’s brain that are not dedicated to biological processes like breathing. Most of that should be wrapped in double quotes, don’t take it literally! But that number is about 100x lower than the same synapses in a human brain.

If the next order of magnitude costs 40B, I wonder if it’s even possible to get to the one after.

1. People fighting back due to never ending effects caused by these massive data centers

2. Companies realising that they are burning half million to get nowhere

3. Circular investment scaring investors

4. And more recently, companies hiring people back coz the AI aimed to replace humans, created more problems than solved them

5. Memory cartel falling apart, again, they did the same thing during 2000s

6. China is making good ML free, supply and demand, destroying the US tech token business model

7. Even META has too much computer power and no enough use for them.

Those are the main reasons why AI buildout is not just slowing down but falling apart faster than expected.

I think at the very least, once the dust settles, a lot of these datacenters could become really really cool haunted houses, giant escape rooms, etc.

The real "AI" success story will be the person that makes an IRL backrooms theme park in the husk of a datacenter.

Or: laser tag park, the vests you wear are in part old tpu/gpu components.

Crusoe bypasses the grid completely by leveraging old batteries and wasted resources to power compute.

Maybe it's the same thing I heard it was six months ago when it was determined that more hardware was sold than twice the nation's supply of electric.

‘the abundance of [AI] will be limited by the abundance of energy’.

And the reason current US policy opposes clean, renewable energy is --- purely political.

This is part of why the AI bubble will burst. The only way to make the profit numbers backing the loans to AI companies is to get increased capacity, and the capacity requires energy, and the energy won't arrive in time, only partly due to all the factors here, and partly because building transmission and generation is speculative and can fail for a number of reasons.

US administration can try to pull a China and basically remove all regulatory barriers (following existing playbook of "do whatever we want and wait a year or two for the courts to catch up and stop us"). It'll create havoc that will make people very upset (more so than the people that already protest DCs in their backyards). But even then, it's construction on varied terrain and property over long distances; you can't predict exactly how that will go. Triple the estimated timeline and that is probably doable, but current AI investment likely can't wait that long, unless somebody can pull additional hundreds of billions out of a hat to extend lines of credit or a ponzi-scheme-esque paying-creditors-with-newly-lent-money. In that time the market will realize the hype was hype, the gains were modest, they'll start divesting, and then the house comes down.

One way around that might be to deploy thousands more gas turbines and make rural air quality look like 2010 Beijing. It will probably happen if things get really tight, and we'll see how the current administrations's base responds; if they stick it out, the market gets a reprieve.

We should have been building more nuclear. We’re not going to upgrade civilization with windmills and putting on sweaters. Think of how much power we’ll need for millions of robots?

Cheap ubiquitous distributed power systems will change the world in many weird ways. Watch small modular nuclear offer home installation for ~$reasonable and getting cheaper every year.

Fast forward 20 years from the advent of essentially infinite energy results in WWIII and a new “Great Detente” but only after all the assholes have wreaked all the havoc they can.

There are dark days ahead but ultimately a brighter future. Sucks to live through that transition phase though.