Anthropic just borrowed 133 billion dollars from different investors in September, and 20 billion in total over 2025. Just where would they get such amount of actual money? Ask Softbank for a donation? Those guys do like to pour money into a fire pit.
> 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 median household consumption per region [1]. (If you don't use your allocation, you should earn a rebate.) Above that, market price. Same for preferred commercial uses, e.g. retail and local government.
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?
Super large DC's yes. However, there is tons of what I call stranded power. Only need 1-2MW? Well, that's $30-100M worth of compute capex, but it really isn't hard to find power / space for that.
> 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 wait.
> 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].
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).
The sources for some of the future data centers will be local and not necessarily influence the US grid. Consider also that cement production uses about 3000 TWh per year worldwide, and aluminium smelting uses about 1000 TWh per year.
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.
We can see in Ukraine that AI plays a very very small role in the war. Production of small drones and their control across jammed areas is the current problem space.
And I bet that a majority of those LLM usage in Ukraine is via local focused models, good for picking shapes on a grainy images and not much else, but fast and cheap. There is literally zero use for the gigantic general LLMs which can produce human-like output and routinely generate fake numbers, in the army setting.
That's the story the proponents of the AI bubble would have you believe, because they are sucking in all available funding to their enrichment, or because they've been huffing their own hype gas for so long that they have no brain cells of their own left.
It is, however, complete nonsense, and the next few years of failed promises on AGI will eventually bring people to their senses, if a market crash and sustained economic depression doesn't do that first. It would be funny if it wasn't going to cause suffering for millions of people, whether we succeed at AGI or not.
I _like_ AI, I find LLMs and many other aspects of useful, and I am optimistic for the long term prospects of AI. But the rush to try and get to AGI is completely out of control at this point, and the fallout from when the bubble pops will set AI, and our societies, back a long time.
Yes they will, they will prefer natural gas of those two. The main reason of emission reductions as we transition off of coal.
I really hope we go all in on nuclear though, with some natural gas and get rid of the windmills. Solar, hydro, geothermal can stay where it makes sense.
Unfortunately our skies and land are littered with windmills made out of unrecyclable polymers that are terrible for raptors.
Queue someone quoting how many cats kill birds like competing with the top predator is a good thing, or ignoring the fact that we put these in raptors' wind streams and cats don't hunt those large birds (which are usually endangered)
"Allowing" doesn't necessarily translate into "doing". Many people are seeing higher energy prices which are at least in part or wholly due to data center loads on local power grids. In Ohio, for example, we just skated by with a ruling from our Public Utilities Commission which effectively required data centers to pay for their impact on local grids [1].
Additionally, while these data centers do provide some jobs, where states are giving them grants, loans, infrastructure improvement, or otherwise they are ultimately extractive developments (like parking lots) where the wealth flows out from states like Ohio and flows in to states where the CEOs and HQ sit (California, New York, etc.).
I can tell you that people in Ohio across the political spectrum are not happy. We are losing good farm land, utilizing water, and our power costs are going up for negligible benefits at best. But hey now our state representatives can say "Meta is coming to central Ohio". Meanwhile costs are going up and we still have to ship produce in from other countries and states.
If our representatives and governors office thought about this all for about 2 seconds they would require any data center development to include 2x the number of corporate jobs over a certain income threshold or else not approve the development. If the developers balk, then fine it's not like we want them anyway.
The Trump Administration (and for that matter probably any admin) isn't doing jack shit.
Oh nice, that basically solves the issue. I've been hearing horror stories of datacenters overloading existing grids and raising prices for the average person, If datacenters generate their own power that basically solves the issue.
Out here in AZ, solar combined with battery would be perfect for datacenters.
> 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.
The most cost effective way to run a datacenter at some definition of "pedal to the metal", 24/7. This is not appropriate for solar, which is why these companies are looking into power sources that are most cost effective when they run pedal to the metal, 24/7, like nuclear.
Will this pan out? We don't know, no one knows. But this isn't "a scam" there is a plausible future where a large percentage of white collar (or dare I say it, blue collar) work will have an assistant and that assistant requires a considerable subscription (200/mo? 1000/mo?).
Interesting to see all of the leading labs in the West make this bet.
12k a year out of your paycheck for an advanced Clippy "assistant"? Sorry, this is not plausible. Oh and by blue collar work do you mean done by walking talking robots? I bet you think we'll be flying cars to work w/in 5 years too huh. Oh yeah and when is your chatbot going to solve physics and cure cancer again? You ppl have lost your minds.
Actually yeah 12k a year better be really really good, because that can get you a lot of quality human. At 12k per year and $100/hour, that gets you 120 hours of time which means you get ~20 minutes per day, on average. Or if you get down to $33 an hour it's an hour a day.
I don't think it requires robots. Although that's possible too.
I think that HoloLens has a reasonable demonstration of how to assist blue collar work about 10 years ago (AFAIK it flopped). I would bet a dollar that similar technology augmented with LLMs could be useful to blue collar work.
Well clearly Microsoft don't see how given that they canceled the project this year. Do you really think we need some video game HUD in our vision at all times? Come back to the real world it's nice here.
but it’s private money? who gives a fuck about the $/job created? if anything, it’s a good thing that Anthropic can afford to do it because they can so efficiently use capital. at least, so far…
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.
Lets put that quote in its full context, because its designed to sound much more impressive than it actually is.
> Anthropic serves more than 300,000 business customers, and our number of large accounts—customers that each represent over $100,000 in run-rate revenue—has grown nearly sevenfold in the past year.
Let me deconstruct that:
> Anthropic serves more than 300,000 business customers
Hard fact. No qualification on spend or activity, are they on trails or fully paid with contracts and minimum spend
> and our number of large accounts—customers that each represent over $100,000 in run-rate revenue
run-rate revenue is an extrapolation. (https://www.fool.com/terms/r/run-rate/) That could be buisnesses that trail anthropic for a month, spend 24K and think "fuck thats expensive" and stops spending. average that over 2 months, then times by 12, boom 100k account.
> has grown nearly sevenfold in the past year.
no starting base....
Its unconvincing, because its smoke and mirrors. Give me the numbers of paying customers, over time with revenue. Then show the opex/capex.
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.
They buy hardware, replace it as the years go on, and continue doing business.
The investment isn't just in raw compute - they have to build buildings, pay staff, and other things. For the hardware and software - they just keep pace as all the other computing companies have to.
Anthropic doesnt have $50B to invest, even over multiple years, so who is really writing the checks?
The writing is on the (tomshardware.com) wall. Don't invest in AI. Invest in power, HDDs, GPUs, HVAC equipment, etc.
https://news.ycombinator.com/item?id=45896707
Anthropic just borrowed 133 billion dollars from different investors in September, and 20 billion in total over 2025. Just where would they get such amount of actual money? Ask Softbank for a donation? Those guys do like to pour money into a fire pit.
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 median household consumption per region [1]. (If you don't use your allocation, you should earn a rebate.) Above that, market price. Same for preferred commercial uses, e.g. retail and local government.
[1] https://www.eia.gov/energyexplained/use-of-energy/electricit...
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?
> should be a tenure element to power access
What does this mean?
> what is the argument in favor of longtime residents having to pay for that CapEx?
Jobs and tax base. These benefit the middle class more than the poor, hence my redistribution proposal.
Super large DC's yes. However, there is tons of what I call stranded power. Only need 1-2MW? Well, that's $30-100M worth of compute capex, but it really isn't hard to find power / space for that.
the administration is allowing datacenters to be utilities and build their own power plants, nuclear or otherwise.
the excess can be sold to the grid.
it's really the only way forward. seems like a win/win.
> 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 wait.
> 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
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...
The sources for some of the future data centers will be local and not necessarily influence the US grid. Consider also that cement production uses about 3000 TWh per year worldwide, and aluminium smelting uses about 1000 TWh per year.
Yes but American datacenters and industries are actually useful compared to AI.
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?
If they are treated like every other customer, all of our energy bills go up the more data centers they build (whether o
This is not a theoretical concern, it is happening already.
https://www.youtube.com/watch?v=YN6BEUA4jNU
OK then I guess we should disconnect every large industrial electricity customer from the grid. That way we'll all have lower energy bills, right?
Because we live on a finite planet and unregulated capitalism won't end well.
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.
We can see in Ukraine that AI plays a very very small role in the war. Production of small drones and their control across jammed areas is the current problem space.
And I bet that a majority of those LLM usage in Ukraine is via local focused models, good for picking shapes on a grainy images and not much else, but fast and cheap. There is literally zero use for the gigantic general LLMs which can produce human-like output and routinely generate fake numbers, in the army setting.
That's the story the proponents of the AI bubble would have you believe, because they are sucking in all available funding to their enrichment, or because they've been huffing their own hype gas for so long that they have no brain cells of their own left.
It is, however, complete nonsense, and the next few years of failed promises on AGI will eventually bring people to their senses, if a market crash and sustained economic depression doesn't do that first. It would be funny if it wasn't going to cause suffering for millions of people, whether we succeed at AGI or not.
I _like_ AI, I find LLMs and many other aspects of useful, and I am optimistic for the long term prospects of AI. But the rush to try and get to AGI is completely out of control at this point, and the fallout from when the bubble pops will set AI, and our societies, back a long time.
Building gigantic data centers doesn't help in that respect. The data centers are there to do inference at scale, not cutting edge research.
Colocated power might be more efficient, depending on how it's done. It avoids transmission losses and allows the grid to be used for other purposes.
That's not a way forward, that's standing still.
I'm all for more efficient usage, and it's in AI companies best interest to do so to minimize costs.
...but it's a growing industry, it will need more power.
No, it's not standing still. It's setting the priorities straight.
Completing the energy transition is an enormous undertaking. Building huge data centers is a distraction, not a way forward.
Even better, don’t just “allow”, but force them to create power plants equal in size to their usage. Must be renewable.
I don't care what they use. Nuclear preferred but whatever works for the area.
Natural gas is the main reason our emissions have gone down as it replaces coal.
Also I don't think forcing is necessary. These datacenters want to, why impose more regulations.
Your grandchildren will care if they use coal or natural gas.
Aim higher. Do better.
Yes they will, they will prefer natural gas of those two. The main reason of emission reductions as we transition off of coal.
I really hope we go all in on nuclear though, with some natural gas and get rid of the windmills. Solar, hydro, geothermal can stay where it makes sense.
Unfortunately our skies and land are littered with windmills made out of unrecyclable polymers that are terrible for raptors.
Queue someone quoting how many cats kill birds like competing with the top predator is a good thing, or ignoring the fact that we put these in raptors' wind streams and cats don't hunt those large birds (which are usually endangered)
Our kids deserve nuclear.
> Unfortunately our skies and land are littered with windmills made out of unrecyclable polymers that are terrible for raptors.
Like our lands and oceans are not littered with drilling platforms and fracking stations already
"Allowing" doesn't necessarily translate into "doing". Many people are seeing higher energy prices which are at least in part or wholly due to data center loads on local power grids. In Ohio, for example, we just skated by with a ruling from our Public Utilities Commission which effectively required data centers to pay for their impact on local grids [1].
Additionally, while these data centers do provide some jobs, where states are giving them grants, loans, infrastructure improvement, or otherwise they are ultimately extractive developments (like parking lots) where the wealth flows out from states like Ohio and flows in to states where the CEOs and HQ sit (California, New York, etc.).
I can tell you that people in Ohio across the political spectrum are not happy. We are losing good farm land, utilizing water, and our power costs are going up for negligible benefits at best. But hey now our state representatives can say "Meta is coming to central Ohio". Meanwhile costs are going up and we still have to ship produce in from other countries and states.
If our representatives and governors office thought about this all for about 2 seconds they would require any data center development to include 2x the number of corporate jobs over a certain income threshold or else not approve the development. If the developers balk, then fine it's not like we want them anyway.
The Trump Administration (and for that matter probably any admin) isn't doing jack shit.
[1] https://www.news5cleveland.com/news/state/ohio-regulators-tu...
Oh nice, that basically solves the issue. I've been hearing horror stories of datacenters overloading existing grids and raising prices for the average person, If datacenters generate their own power that basically solves the issue.
Out here in AZ, solar combined with battery would be perfect for datacenters.
> If datacenters generate their own power that basically solves the issue.
If they do so cleanly otherwise they'll dump their externalities on the rest of us.
How many times has that happened so far?
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.
The real win/win would be to require them to build enough solar to power the operation.
> to build enough solar to power the operation
The most cost effective way to run a datacenter at some definition of "pedal to the metal", 24/7. This is not appropriate for solar, which is why these companies are looking into power sources that are most cost effective when they run pedal to the metal, 24/7, like nuclear.
And then the entire energy grid collapses with the AI bubble.
Are there any more details on this investment? Do they have the hardware or even the power to ready support such growth?
"invests" "$50B"
>The project will create approximately 800 permanent jobs
For $50 billion?
I think there's a serious problem here.
> For $50 billion?
It’s Capex. Most home construction produces less than a single permanent job on average.
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...
This is clearly aggressive.
Will this pan out? We don't know, no one knows. But this isn't "a scam" there is a plausible future where a large percentage of white collar (or dare I say it, blue collar) work will have an assistant and that assistant requires a considerable subscription (200/mo? 1000/mo?).
Interesting to see all of the leading labs in the West make this bet.
12k a year out of your paycheck for an advanced Clippy "assistant"? Sorry, this is not plausible. Oh and by blue collar work do you mean done by walking talking robots? I bet you think we'll be flying cars to work w/in 5 years too huh. Oh yeah and when is your chatbot going to solve physics and cure cancer again? You ppl have lost your minds.
Actually yeah 12k a year better be really really good, because that can get you a lot of quality human. At 12k per year and $100/hour, that gets you 120 hours of time which means you get ~20 minutes per day, on average. Or if you get down to $33 an hour it's an hour a day.
If it can format images in Microsoft Word without breaking the document, or fix Microsoft Excel formula issues everyone will be buying.
"12k a year out of your paycheck for an advanced Clippy "assistant"? Sorry, this is not plausible"
It's certainly possible. Will it actually happen? IDK.
I don't think it requires robots. Although that's possible too.
I think that HoloLens has a reasonable demonstration of how to assist blue collar work about 10 years ago (AFAIK it flopped). I would bet a dollar that similar technology augmented with LLMs could be useful to blue collar work.
Well clearly Microsoft don't see how given that they canceled the project this year. Do you really think we need some video game HUD in our vision at all times? Come back to the real world it's nice here.
> Today, we are announcing a $50 billion investment in American computing infrastructure
> The project will create approximately 800 permanent jobs
Approximately $62.5 million per permanent job created!
but it’s private money? who gives a fuck about the $/job created? if anything, it’s a good thing that Anthropic can afford to do it because they can so efficiently use capital. at least, so far…
$170k per buisness customer. That's not including existing debt and opex.
Good luck paying that back, especially as AI is basically commodity now.
Have you seen their 2025 growth curve (and projections)?
https://www.reddit.com/r/EconomyCharts/comments/1lwdwd6/anth...
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.
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.
Are you saying that in next 10 years they'll make more money that there's an atoms in the universe?
have you seen their cost curves and projections as well?
The area under that curve is quite a bit less than 50B.
that is like 3 data point my man and you think you can project them just straight up forever and ever. this is bubble thinking.
Over a decade?
Also,
> customers that each represent over $100,000 in run-rate revenue—has grown nearly sevenfold in the past year.
isn’t unconvincing.
Lets put that quote in its full context, because its designed to sound much more impressive than it actually is.
> Anthropic serves more than 300,000 business customers, and our number of large accounts—customers that each represent over $100,000 in run-rate revenue—has grown nearly sevenfold in the past year.
Let me deconstruct that:
> Anthropic serves more than 300,000 business customers
Hard fact. No qualification on spend or activity, are they on trails or fully paid with contracts and minimum spend
> and our number of large accounts—customers that each represent over $100,000 in run-rate revenue
run-rate revenue is an extrapolation. (https://www.fool.com/terms/r/run-rate/) That could be buisnesses that trail anthropic for a month, spend 24K and think "fuck thats expensive" and stops spending. average that over 2 months, then times by 12, boom 100k account.
> has grown nearly sevenfold in the past year.
no starting base....
Its unconvincing, because its smoke and mirrors. Give me the numbers of paying customers, over time with revenue. Then show the opex/capex.
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?
None of this infra is worth anything more than five years from now.
How do you think computing companies work?
They buy hardware, replace it as the years go on, and continue doing business.
The investment isn't just in raw compute - they have to build buildings, pay staff, and other things. For the hardware and software - they just keep pace as all the other computing companies have to.
Also, 800 permanent employees is $62.5M/employee.
With the 2400 temporary employees it's $14M/job, which is very very capital intensive.
If trends continue, all investment in the economy will be directed by about 6 people at a big AI company, and what will money mean at that point?
Now let's wait until US invests 50B in Anthropic!
Look at that! where did this money come from?