Hacker News Reader

Comment by dgellow | Hacker News Reader

AI inference and training GPUs require way, way more power than anything manufactured before. Like, between 2020 and 2026 we went from ~400W for an A100 to ~1400W for an NVIDIA B300. And you cannot make that happen without changing the datacenter itself. More noise, way more complicated cooling setup, new power sources, etc. Not all datacenters are created equal.

The financing for datacenters is also very shady and its own can of worms

View on HN

This is very important, filling racks with GPU's instead of CPU's is vastly more power hungry (and hence will generate a lot more heat). Vera Rubin is 1800-2300W, mind boggling, these data centers even have new power architectures coming.

And I know what people will say, Vera Rubin is vastly more efficient, but that only matters if you have saturated your needs, it's meaningless if you need that capacity, and I'm sure it will run pretty much at a 100% 24/7.

> The financing for datacenters is also very shady and its own can of worms

Could you expand on that?

I don't think it's necessarily shady, but let's say that we're assisting at bets and capex spending we've never seen in history.

Companies like Meta, Alphabet, etc that were drowning in cash flow and buying back their stock are now back at issuing shares, creating a huge number of SPVs, issuing debt, etc.

There's a gigantic spending spirale with no signs of ending. Nvidia is financing companies to buy its own hardware.

Mind you, this is nothing new in some businesses. Airbus and Boeing have been lending money to companies to buy their own airplanes forever.

But its just all too much, with very little money coming. To understand how massive is the spending: even if every adult on the planet spent 250$ on AI per year this would still not cover anywhere near what's being invested.

> even if every adult on the planet spent 250$ on AI per year this would still not cover anywhere near what's being invested.

Your math does not make sense. That's something like $1.5 trillion every year. That's no where near what is being spent in AI hardware buildouts on an annual basis.

Last year estimated spend was $380 billion and this year it's $680 billion [0] just for the big five. Given global projections through 2030 being as high as 7 trillion [1], OPs comment seems at least the right order of magnitude.

[0]: https://futurumgroup.com/insights/ai-capex-2026-the-690b-inf...

[1]: https://www.mckinsey.com/industries/technology-media-and-tel...

Gartner estimates that 2026 will have about $2.5 trillion in AI investments, with 2027 reaching over $3 trillion.

https://www.gartner.com/en/newsroom/press-releases/2026-05-1...

Those are totals, not annual spending.

Thank you for the correction.

1. The estimates, for data center build outs in 2026 alone, in US alone, and even limiting just to the hyperscalers + oracle is in the 600 to 800 billions range.

And I'm ignoring the likes of Nvidia building their owns and smaller (but still huge) outside the biggest players. Or the comparably smaller, yet still multi billions spent in the rest of the world.

2. That excludes salaries, research, taxes, cost of running the datacenters themselves, etc.

In fact several estimates are pointing out that global annual spending on AI better be in the range of 3T$ range (that's almost a 1'000 $ / year per every adult in the world) or this is all going bust.

The math that does not make sense is to expect such a gargantuan spending to have any meaningful monetary return.

I don't necessarily know what the parent was referring to, but I can see two this:

Megacorps often get significant tax breaks to build AI in a community.

Megacorps advertise that datacenters will create jobs but don't mention that most of those jobs will disappear once construction is finished.

Reply on news.ycombinator.com