Live data from Hacker News

Anthropic invests $50B in US AI infrastructure

anthropic.com

61–70 of 129 posts

Re: Anthropic invests $50B in US AI infrastructure

#61

Earlier quoted context omitted.

> forcing AI companies to use power more efficiently How? Also, why? Why are datacentres the use to tamp down on versus other industrial and commercial uses? This reminds me of California rationing residential water use so alfalfa farmers can flood their fields.

All good questions, I don't claim to have all the answers. What I'm saying is that using gigawatts of power for "AI" in this day and age is madness. I do like the market insulation idea you propose in another comment (I would link to it, but apparently HN doesn't allow that).

> What I'm saying is that using gigawatts of power for "AI" in this day and age is madness

Why? American datacentres--of all types--use about 250 TWh per year, with another 500 TWh additional capacity expected by 2030 [1]. American paper manufacturing used about that much energy in 2018 [2].

[1] https://www.iea.org/reports/energy-and-ai/energy-demand-from...

[2] https://www.eia.gov/energyexplained/use-of-energy/industry.p... 2,491tn BTUs ~ 730 TWh

Re: Anthropic invests $50B in US AI infrastructure

#62
post #31
post #15

Earlier quoted context omitted.

Growth curves mean nothing if you're selling $0.90 dollars. You have to show a growth curve when price > cost. It's not even clear that value > cost. I absolutely love Anthropic; but I am worried about the fiscal wall they will hit that will ratchet up my opex as they will need to steeply raise prices.

So the critical question here really is whether they are selling API access to their models for less than the unit cost it takes to serve them.

I don’t think it is only that consideration

While Gross margins numbers are estimates vary widely, 40-60% numbers some analysts throw around seems realistic.

In an equity only company that is good enough metric , but all the major players have long since now transitioned to also raising debt.

The debt would need to be serviced even if fresh training investments stopped fully .

The cost of debt servicing would depend on the interest rates and the economy etc inaddition to the risk of the debt itself.

Quite possible that model companies would need to jack prices even with good gross margins to handle their debt load.

Re: Anthropic invests $50B in US AI infrastructure

#63

>The project will create approximately 800 permanent jobs For $50 billion? I think there's a serious problem here.

AI spending is more on hardware than on jobs. That shouldn't be too great a surprise.

$62 million per job does seem a bit more hardware-heavy than reasonable, though...

Re: Anthropic invests $50B in US AI infrastructure

#64

Earlier quoted context omitted.

Google is working with Kairos Power for nuclear reactors at their datacenters. Microsoft struck a deal with Three Mile Island to get a reactor up. And Amazon is working with Talen Energy. Might be others I missed.

So the public is expected to deal with higher prices after every data center build? The work can't be done in advance?

> the public is expected to deal with higher prices after every data center build?

No. But the states that let companies put trillions of dollars of datacentre and power hardware in their borders will probably reap benefits from it for decades to come regardless of how AI pans out.

Re: Anthropic invests $50B in US AI infrastructure

#65

We still need to fix the problem with powering these Datacenters....

> We still need to fix the problem with powering these Datacenters.... Not really. We need to insulate consumers from the market that is solving and will solve that problem. That's a financial engineering and policy problem. America is good at the first. We're bad at the second. That implies state and local initiatives should take the lead. My proposal: one market for essential residential consumption, defined as the…

There should be a tenure element to power access. The reductio is a well-funded adversary should not be able to buy all the power in a region just because they are able to pay more.

Our utilities are generally regulated, and have some mandate to provide power to residents. If a datacenter creates a sudden dislocation in the demand for power that causes massive unplanned CapEx by the utility, what is the argument in favor of longtime residents having to pay for that CapEx?

Re: Anthropic invests $50B in US AI infrastructure

#66
post #39
post #31

Earlier quoted context omitted.

So the critical question here really is whether they are selling API access to their models for less than the unit cost it takes to serve them.

But inference is cheap! If they stop doing everything and become Inference Inc., they'll be profitable.

Until China drops another open weight model you can run yourself at cost price.

Re: Anthropic invests $50B in US AI infrastructure

#67
post #30

Earlier quoted context omitted.

> it's really the only way forward. seems like a win/win. There is another way forward, which is not building these data centers, forcing AI companies to use power more efficiently, and use the excess energy production capacity towards the energy transition in order to avoid the worst consequences of climate change. It's not going to happen, at least not right now, but it's clearly what we ought to do. ChatGPT can wa…

Why should they be treated differently than any other customer?

(I don't know the answer to this,) is it common for a single customer to enters a market with an ask for 10GW to start?

Re: Anthropic invests $50B in US AI infrastructure

#68

Earlier quoted context omitted.

All good questions, I don't claim to have all the answers. What I'm saying is that using gigawatts of power for "AI" in this day and age is madness. I do like the market insulation idea you propose in another comment (I would link to it, but apparently HN doesn't allow that).

> What I'm saying is that using gigawatts of power for "AI" in this day and age is madness Why? American datacentres--of all types--use about 250 TWh per year, with another 500 TWh additional capacity expected by 2030 [1]. American paper manufacturing used about that much energy in 2018 [2]. [1] https://www.iea.org/reports/energy-and-ai/energy-demand-from... [2] https://www.eia.gov/energyexplained/use-of-energy/indus…

If I read the data right (1) the US currently produces roughly 4,000 TWh of electricity every year. 500 TWh is a significant portion of that! The US will need a lot of additional capacity for things like electric cars and heat pumps. Most of the effort should be going towards that, not huge data centers attending to unproven demand (how many people will pay the real price for ChatGPT once the VC subsidy ends remains to be proven).

1: https://www.ceicdata.com/en/indicator/united-states/electric...

Re: Anthropic invests $50B in US AI infrastructure

#69
post #19

Earlier quoted context omitted.

How impulse-buy has business AI licensing got? Is it really a surprise later, the cost?

Less impulse, more "oh we expected that we'd get more return". We have a project at the moment thats all based around sharepoint. They have ingested many tens of thousands of documents, and are expecting that MS copilot studio will be able to a) RAG and B) produce meaningful answers with a 4 line prompt.

Does it take more than a quarter or two to figure that out?

Re: Anthropic invests $50B in US AI infrastructure

#70

Earlier quoted context omitted.

> it's really the only way forward. seems like a win/win. There is another way forward, which is not building these data centers, forcing AI companies to use power more efficiently, and use the excess energy production capacity towards the energy transition in order to avoid the worst consequences of climate change. It's not going to happen, at least not right now, but it's clearly what we ought to do. ChatGPT can wa…

I think you may be missing the whole national security part of the AI race. This isn't just about asking a computer what recipe you can cook tonight with the items in your fridge. In many ways it is similar to the race to build a nuclear bomb. We may individually not like that, but we might individually be best served to live in the nation that got there first.

[dead]
Post reply on HN