Hacker News Reader

Comment by arcticbull | Hacker News Reader

People tend to equate this to the railroad boom when saying that infrastructure spending will yield durable returns into the future no matter what.

When the railroad bubble popped we had railroads. Metal and sticks, and probably more importantly, rights-of-way.

If this is a bubble, and it pops, basically all the money will have been spent on Nvidia GPUs that depreciate to 0 over 4 years. All this GPU spending will need to be done again, every 4 years.

Hopefully we at least get some nuclear power plants out of this.

View on HN

Yeah, the short-lived GPU deprecation cycle does feel very relevant here.

I'm still a fan of the railroad comparisons though for a few additional reasons:

1. The environmental impact of the railroad buildout was almost incomprehensibly large (though back in the 1800s people weren't really thinking about that at all.)

2. A lot of people lost their shirts investing in railroads! There were several bubbly crashes. A huge amount of money was thrown away.

3. There was plenty of wasted effort too. It was common for competing railroads to build out rails that served the same route within miles of each other. One of them might go bust and that infrastructure would be wasted.

The boom might not last long enough for western countries to pull heads out of their collective asses and ramp up production of nuclear plants.

It takes China 5 years now, but they've been ramping up for more than 20 years.

Imagine the progress we could have made on climate change if this money had been funneled into that, instead of making some GPU manufacturers obscenely wealthy.

Throwing away the future for "AI" slop.

Yeah it’s infuriating to think about.

What percentage of data centre build costs are the GPUs vs power stations, water cooling plants, buildings, roads, network, racks, batteries, power systems, etc

A bunch of the money is being spent on data centers and their associated cooling and power systems and on the power plants and infrastructure. Those should have much longer depreciation schedules.

There's a lot more to infrastructure spending than GPUs. Companies are building data centers, cooling systems, power plants (including nuclear), laying cables under oceans, launching satellites. Bubble or not, all of this will continue to be useful for decades in the future.

Heck if nothing else all the new capacity being created today may translate to ~zero cost storage, CPU/GPU compute and networking available to startups in the future if the bubble bursts, and that itself may lead to a new software revolution. Just think of how many good ideas are held back today because deploying them at scale is too expensive.

> including nuclear

Note that these are just power purchase agreements. It's not nothing, but it's a long ways away from building nuclear.

The recycling industry will boom. From what demand you ask? We'll find out soon enough.

>> basically all the money will have been spent on Nvidia GPUs that depreciate to 0 over 4 years

I agree the depreciation schedule always seems like a real risk to the whole financial assumptions these companies/investors make, but a question I've wondered: - Will there be an unexpected opportunity when all these "useless" GPUs are put out to pasture? It just seems like saying a factory will be useless because nobody wants to buy an IBM mainframe, but an innovative company can repurpose a non-zero part of that infrastructure for another use case.

Railroads need repair too? Not sure if it's every 4 years. Also, the trains I take to/from work are super slow because there is no money to upgrade.

I think we may not upgrade every 4 years, but instead upgrade when the AI models are not meeting our needs AND we have the funding & political will to do the upgrade.

Perhaps the singularity is just a sigmoid with the top of the curve being the level of capex the economy can withstand.

For what it's worth they cost a lot less than highways to maintain. Something like the 101 in the Bay Area costs about $40,000 per lane-mile per year, or about $240,000.

Trains are closer to $50-100,000 per mile per year.

If there's no money for the work it's a prioritization decision.

I think the hardware infrastructure may be obsolete but at the moment we are still just beginning to figure out how to use AI. So the knowledge will be the important thing that’s left after the bubble. The current infrastructure will probably be as obsolete as dial up infrastructure.

This is precisely why the AI bubble is so much worse than previous bubbles: the main capital asset that the bubble is acquiring is going to depreciate before the bubble's participants can ever turn a profit. Regardless of what AI's future capabilities are going to be, it's physically impossible for any of these companies to become profitable before the GPUs that they have already purchased are either obsolete or burnt out from running under heavy load.

Reply on news.ycombinator.com