> A back-of-the-envelope calculation finds that covering AI capex through identifiable AI income requires revenue on the order of $2.5trn per year,
Ok, so what is the botec? They don't say, but it's probably something like "next year's $1.5T/year in capex * 50% return on invested capital".
Obviously a given year's capex doesn't need to return its investment immediately the following year, but over it's ~5-7 year depreciation period. So in making this botec, they're not just taking next year's capex, but extrapolating it to a steady state of $1.5T/year in capex.
But why would that level of spending be steady? It's exceedingly unlikely to be. Future capex is not locked in. It's contingent on revenue and revenue growth. AI revenue has been growing faster this year than even the most optimistic projections suggested. It's hardly a surprise that capex projections were dialed up. If revenue growth lagged instead, it would go the other way.
(You need only look at Google Cloud growth and margins to get an idea of how good an investment last year's AI capex was. But at the time, the arguments for it being crazy were identical to those made today, except with the numbers substituted.)
There's a separate issue, which is that you can't really think about this on a sector-wide basis. In most businesses there will be winners and losers, demanding everyone be a winner is unrealistic. E.g. right now anyone with the business model of renting out the compute is being showered in money, people with a business model of building non-frontier models are losing money. The latter group's bad strategy doesn't invalidate the former group's good business.
AI will probably end up reducing lots of firms productivity because costs to defend against cyber attack is gonna skyrocket. Even for tech firms writing more code (the only thing AI is really good at so far) doesn't scale linearly with profits
Why would it obviously give attackers the upper hand? If we accept the premise that AI helps find vulns more efficiently, shouldn’t this be at least equally helpful in securing against attacks and make that increasingly efficient?
It won't do that for free, companies (and open source maintainers) will have to pay for the tokens used finding vulnerabilities in their own software. That increases the cost of releasing the same product/feature/amount of code vs what it was before.
everything you build potentially increases the surface area of attack. If the benefit of building it is zero or negative then all you've done is increase your risk profile.
Particularly when you can't use Fable 5 to secure your software, you don't have access to Mythos 5, and you got your access to GLM-5.2 and other Chinese and/or open models capable of doing so taken away.
Good point. Within the next few years black hat open models will be able to do a lot of damage for very little cost.
It could easily turn into a Red Queen situation where companies that don't spend hard on security are going to face huge potential losses, and both sides will be scrambling constantly to get an advantage.
A lot of people are slowly realizing that code generation was not the bottleneck in their internal development process, and waving a magic wand to make that go faster doesn't actually lead to more revenue.
One of the weirdest things about the whole AI boom is the mass formation psychosis where everyone started acting like lines of code written was an important metric for software development productivity.
We've known for decades that more lines of code isn't good. Bill Gates was joking about it in the OS/2 days!
We've known for decades that the actual creation of the code is the smallest part of an actual software developer's job.
We've known for decades that "every line of code is a business decision" that has to be written with an understanding of the underlying goal, and the hardest part of training up software engineers is making them understand that.
We threw all that out because "the chatbox writes code really fast, software is a solved problem!"
This is a rewriting of history that I don't really like. Back before LLMs and before that again, it was very hard to write code. That's why so many people asked questions and figuring out how to write a piece of code was a big part of the value add of a programmer. It separated the ones who knew their stuff from the ones who didn't. Yeah sure software design was also important but it was less important than the actual writing of the code, which is why a lot of old code had bad designs. It took decades to figure that stuff out.
> Back before LLMs and before that again, it was very hard to write code. That's why so many people asked questions and figuring out how to write a piece of code was a big part of the value add
This may be true if you work in unfamiliar area or on frontier of programming. But this is not were most of programmer jobs are. That writing code itself is hard is a view of non-programmer. Programmers know that past some threshold of familiarity designing and weighting decisions becomes central problem.
Sure it may be that way now. But if you go back 20 years to for example video game programming, the hard part was figuring out how to solve the problem, not spend weeks or months trying to figure out the abstract design of the code itself. 10-20 years ago a lot of the value of programming was figuring out how to do something, which is by its nature about writing the code, I would argue.
Edit: Also, even outside of video games, I don't remember people being obsessed about software design 10-20 years ago the way they are now. You had some design principles of course and there were obviously bad or redundant ways of doing things, but it's on a whole other level the past 5 years at least from what I remember.
I think what's happened, so far at least, is productivity has been moved around.
Anyone doing consulting has seen it, where now the "I vibe coded this last week" competition is way more intense. Basically the people with the problems now have a chance to spin up something that looks like the solution they want, they then get in trouble and need someone to sort it out.
LLMs _are_ great tools for software development but if you haven't noticed their near complete inability to reason about why what they're doing might work you haven't been trying hard enough.
So from the top it doesn't look like much has changed, whereas workers are trying to offload as much work as they can onto their claude subscription.
People in desperation to min/max work per unit dollar, will leverage AI to free themselves from as much work as possible, while still claiming credit for the work.
So as the labs start ratcheting up the price, people will pay more and more to keep their "secret" assistant.
Wouldn't that explanation predict that lower ranking employees would be more impressed with AI than higher ranking ones? Because what we actually observe [0] tends to be the opposite. Executives are the most bullish on AI, followed by Managers, with employees having the least adoption.
Those studies are of workers using company provided AI.
They call out the MIT study in the article, which found very low adoption of corporate AI initiatives.
However if you read that study, they found that almost every worker uses AI heavily, but they just use their personal AI account.
Again, people are using their own AI to benefit themselves, it makes the most sense. Using company AI shows that the AI is doing xyz, and maybe perhaps you aren't that valuable. Using your own AI makes it look like you are doing xyz, and you are the valuable one.
I guess there is an inverse effect to bubbles (things go linearly up and then abruptly down), where nothing happens for a long time until all things happen at once.
Here’s what I’ve seen: it can make skilled people more productive by acting like a “mech suit for your brain.”
For unskilled people, it creates a false sense of productivity, but if the user of the tech does not understand what they are doing they can’t apply the tool productively or judge its output or deal with the things beyond its ability.
So like all other tools before it: it works best in the hands of a skilled user.
Just like in all other fields.
There is no tool that makes an unskilled user skilled, but there are many that can amplify or extend the capability of skill.
Problem is AI weakening our brain, arguably the only thing that distinguish us from animals. As if decreasing brain size of modern humans was not enough.
Not really. Most things take more time to transform than people think, even if capabilities exist. Which is not to say that they do exist, but rather, independent of whether they exist, that this is not very surprising.
Speaking as a programmer, the net effect so far is that we work more and deliver far more results and get paid the same or less.
And I run my own firm... but I have to compete against a global market of other programmers using similar tools who are also trying to compete on price, and then there are hordes of new entrants who can vibe-code things that superficially look like they'll meet a customer's needs and are better at marketing than I am, so my area of business gets constrained to just customers who need help after the problems with the vibe-coded solution appear, the vendor who made it is long gone, and they already spent an inordinate amount of money on the vibe-coded solution since they thought it was complete and it looked pretty good.
Most of my past roles had very simple code at the important deliverable level, on bizarre and arcane platforms that would have little context for frontier LLMs. Most the creativity and complexity was in tooling for research/development/testing.
Agentic coding could help a lot for the latter, not necessarily the former.
> A back-of-the-envelope calculation finds that covering aicapex through identifiable ai income requires revenue on the order of $2.5trn per year, more than tech’s entire combined revenue today.
…
> All these complex calculations roughly tally with a much simpler one: adding up the ai revenue of the firms selling most of the ai. Anthropic pulls in perhaps $75bn, annualised; Openai makes tens of billions; Google, via its ai model Gemini, and Microsoft probably get a bit less. SpaceX may have a few billion dollars’ worth of revenue from enterprise aithis year. Meta also makes a few bucks from ai. Add this up and you land at roughly $150bn a year.
$2.5tn is about 8% of U.S. GDP. So a handful of AI firms really expect that they will become larger than the entire U.S. sectors of manufacturing, government, or healthcare?
Investors expect that they will, I guess. If you really believe the strong ai use cases (like exponential improvements, or it replacing hundreds of knowledge workers at every company" then maybe this isn't unreasonable.
1) people like me exist: I don't live in the USA, yet it is possible for me to buy things which are made in the USA
2) fiat money is only created or deleted by laws the government controls (plus a tiny quantity from forgery and damage to coins and notes), so everyone losing their jobs doesn't make the money disappear, just who has it available to spend
3) previous waves of automation have created new business opportunities; while this has not been too good for the people who lost jobs to the automation, it has generally boosted the overall economies this happened in, so it is absolutely possible to wipe out tech employees without it seriously harming consumers collectively
If you look at recent trends in consumer spending, a lot more of it is from the upper classes and less from the middle. Selling services to the rich can displace the loss of white collar workers. (in theory). as a bonus, ideally you create some more really wealthy people to spend on things during this shift.
I'm yet to see what part of AI really wealthy people would like to spend their money on. It's not like we can get AI Yachts and AI Rockets. What would the premium product for the upper class be?
I guess the bet is that they will become embedded in basically 8% of the tasks in every sector. which I guess could happen long term? thinking about the whole ass economy in this way always make my mind boggle so I don't really feel equipped to talk intelligently about this.
I suspect that they believe they can integrate with a substantial portion of not just the US economy, but also most of the European and formal Latin America economies. And I also suspect that the number is quite high because there are so many AI companies planning to snag like 80% market share for AI.
Why would capex need to be covered by income during a time of investment and infrastructure buildup? Did income from the Apollo program cover its expenses?
Public investment was responsible for Apollo, so there was no expectation of profitability. Private investment is largely responsible for the AI buildout, and private investment isn't doing that for charity
If I give you 5 trillion dollars I want to get that back plus a percentage increase on top for my troubles, right? Government programs might run without the expectation of paying back taxpayers but investors are less generous.
Also they might not be as smart as they have made out to be. This time it seems they might be spending more to capture market than the market is worth.
Still, there are plenty of people who take their cut of the money on the way. So at least those people will be happy.
Because most of the capital assets depreciate over a period of 5 years or so so you have to make the expenditures back during that time or you can't book any revenue.
> Exponential View, a consultancy, counts $175bn of generative-AI revenue, on an annualised basis, in June. In a recent paper Anton Korinek of Anthropic and Patrick McKelvey of the Bank of Canada estimate total “AI services” revenue. Adapting their methodology, we reckon this was $220bn (again annualised) in the first quarter of this year. Ramp’s data imply that 2-3% of business spending now goes on AI, pointing to $170bn a year.
So between $170bn-220bn in annualized AI revenue today. Maybe it doesn't cover trillion-dollar bets but this is a very substantial number.
This is also at the end of the tokenmaxxing era, so I'm not sure I would draw a straight line out from here. A LOT of companies are clamping down on token costs right now.
Even more companies are just starting AI/ML initiatives. The tokenmaxxing thing only happened at a handful of tech companies, most companies don't have unlimited money to do this kind of thing.
Once the Pathfinders find out where the real value is there is much more room for growth from companies who have be mindful of budget.
You really see the first mover disadvantages in US AI labs. And how second movers (open Chinese labs) can disrupt them. First movers tend to get overcapitalized and way ahead of their skis. You saw a smaller version of this in the vector database market (remember that?) where companies like Pinecone were getting billion+ valuations for capabilities which now seem like a commodity. There's vector DB companies now with much less debt / VC obligations that don't need to clear insane revenue levels to justify investment.
> And how second movers (open Chinese labs) can disrupt them
Yes, I will concede this point. But how are the Chinese labs going to thrive with giving the weights away for free? Will they suffer the same fate open source database companies did at the hands of AWS?
If the model is commoditized then the AI race becomes competition over building infrastructure to support it. China is very good at building infrastructure.
They don't need to run a business, they are a Chinese state enterprise. The tax dollars of Chinese workers covers the cost. China is a communist country with mock free enterprise, it is not the team blue version of the US.
The thing is, AI is a productivity boost but staff drive all the productivity. The catch 22 is AI is expensive so in order to provide it to everyone they need to reduce staffing. The result is remaining folks are able to produce more but the organization as a whole is not exceeding pre-layoff output. And now things are being dropped on the floor and falling in between the cracks -- which impedes efficiency and velocity. Happening at my company now.
AI inference needs to get cheaper or there will always be this "terminal velocity." I suspect AI labs' incentives to reduce costs is only where there is overlap to free up hardware/utilization (to then provide inference to more paying customers.) I highly doubt they will want to make things cheaper for users -- they have debts to pay.
Open weight models are the way and forward, it is the only way an organization can truly control costs by self hosting or buying cheaper inference. Relying on closed weight models is a business risk. Kimi models are only marginally worse than opus but significantly better than the bleeding edge of yesterday's sonnet.
I kind of think the AI boom is the worst for (non-inference-providing) companies. If all companies are using the same "frontier" LLMs, and if they are competitive, what gives one an edge over the other? I think, just the people. Which was the same as before, but now with the additional AI spend that they can't cut, or they become less competitive.
> If all companies are using the same "frontier" LLMs, and if they are competitive
The answer is that all companies will not be using the same frontier LLMs competitively
This will be yet another technology that will benefit from economies of scale. Big businesses and those in PE portfolios will lead the charge, adopt AI, increase efficiency, and gain market share at the expense of mom and pop shops.
This will be good for our 401ks because we're all primarily invested in big business
It makes sense that if software turns to a commodity by AI, and models are also commoditized, then distribution is all that matters. Anthropic is therefore attacking other's distribution in order to protect its business. Perhaps they should focus on their own ways of distribution.
I noticed that were a lot of traditional, non-tech companies interviewing for AI engineers in the Feb/March timeframe that have halted hiring in those roles entirely. It seems that if those roles didn't close by mid-April that they didn't close at all. This seems to match the timeframe in which cost suddenly became prominent in the AI zeitgeist.
I don't know _anyone_ outside of SV who has successfully replaced even a single employee completely with the current models. Maybe someone has pulled this off in call centers, but the POCs have all failed.
I'm very much pro-AI, but I just think we're on a false summit. As the article points out, there is absolutely no way to recoup the investment costs unless the models allow companies to start displacing human workers by the millions _and_ recapture a significant fraction of the displaced workers total comp. If either of those aren't true, then the bubble is going to pop... soon.
I mean, all one has to do is review https://isaiprofitable.com/ to see just how bad it is. None of this is a profitable venture. It's just a money pit for all players involved, all the while draining our ability to buy electronics and drink water. This has to end at some point.
It's kind of funny. I was looking at the top of the page thinking 1.5T investment for so far 770B revenue isn't bad at all, that could break even after some time. But then you look the listed companies, there costs vs revenue.
Turns out 516B of the revenues are from Nvidia, which is not an AI company, but just selling shovels. Nvidias revenues are effectively costs to actual AI companies.
> A back-of-the-envelope calculation finds that covering AI capex through identifiable AI income requires revenue on the order of $2.5trn per year,
Ok, so what is the botec? They don't say, but it's probably something like "next year's $1.5T/year in capex * 50% return on invested capital".
Obviously a given year's capex doesn't need to return its investment immediately the following year, but over it's ~5-7 year depreciation period. So in making this botec, they're not just taking next year's capex, but extrapolating it to a steady state of $1.5T/year in capex.
But why would that level of spending be steady? It's exceedingly unlikely to be. Future capex is not locked in. It's contingent on revenue and revenue growth. AI revenue has been growing faster this year than even the most optimistic projections suggested. It's hardly a surprise that capex projections were dialed up. If revenue growth lagged instead, it would go the other way.
(You need only look at Google Cloud growth and margins to get an idea of how good an investment last year's AI capex was. But at the time, the arguments for it being crazy were identical to those made today, except with the numbers substituted.)
There's a separate issue, which is that you can't really think about this on a sector-wide basis. In most businesses there will be winners and losers, demanding everyone be a winner is unrealistic. E.g. right now anyone with the business model of renting out the compute is being showered in money, people with a business model of building non-frontier models are losing money. The latter group's bad strategy doesn't invalidate the former group's good business.
“ According to Mr Yotzov’s study, nine in ten executives report no impact of ai on their firm’s productivity over the past three years.”
Brutal stuff.
AI will probably end up reducing lots of firms productivity because costs to defend against cyber attack is gonna skyrocket. Even for tech firms writing more code (the only thing AI is really good at so far) doesn't scale linearly with profits
Why would it obviously give attackers the upper hand? If we accept the premise that AI helps find vulns more efficiently, shouldn’t this be at least equally helpful in securing against attacks and make that increasingly efficient?
It won't do that for free, companies (and open source maintainers) will have to pay for the tokens used finding vulnerabilities in their own software. That increases the cost of releasing the same product/feature/amount of code vs what it was before.
everything you build potentially increases the surface area of attack. If the benefit of building it is zero or negative then all you've done is increase your risk profile.
Maybe people believe that the process complexity is not symetric. I don't blame them. Building an house is more complex than destroying it.
It's still an additional cost that doesn't drive revenue.
Particularly when you can't use Fable 5 to secure your software, you don't have access to Mythos 5, and you got your access to GLM-5.2 and other Chinese and/or open models capable of doing so taken away.
Good point. Within the next few years black hat open models will be able to do a lot of damage for very little cost.
It could easily turn into a Red Queen situation where companies that don't spend hard on security are going to face huge potential losses, and both sides will be scrambling constantly to get an advantage.
A lot of people are slowly realizing that code generation was not the bottleneck in their internal development process, and waving a magic wand to make that go faster doesn't actually lead to more revenue.
One of the weirdest things about the whole AI boom is the mass formation psychosis where everyone started acting like lines of code written was an important metric for software development productivity.
We've known for decades that more lines of code isn't good. Bill Gates was joking about it in the OS/2 days!
We've known for decades that the actual creation of the code is the smallest part of an actual software developer's job.
We've known for decades that "every line of code is a business decision" that has to be written with an understanding of the underlying goal, and the hardest part of training up software engineers is making them understand that.
We threw all that out because "the chatbox writes code really fast, software is a solved problem!"
This is a rewriting of history that I don't really like. Back before LLMs and before that again, it was very hard to write code. That's why so many people asked questions and figuring out how to write a piece of code was a big part of the value add of a programmer. It separated the ones who knew their stuff from the ones who didn't. Yeah sure software design was also important but it was less important than the actual writing of the code, which is why a lot of old code had bad designs. It took decades to figure that stuff out.
> Back before LLMs and before that again, it was very hard to write code. That's why so many people asked questions and figuring out how to write a piece of code was a big part of the value add
This may be true if you work in unfamiliar area or on frontier of programming. But this is not were most of programmer jobs are. That writing code itself is hard is a view of non-programmer. Programmers know that past some threshold of familiarity designing and weighting decisions becomes central problem.
Sure it may be that way now. But if you go back 20 years to for example video game programming, the hard part was figuring out how to solve the problem, not spend weeks or months trying to figure out the abstract design of the code itself. 10-20 years ago a lot of the value of programming was figuring out how to do something, which is by its nature about writing the code, I would argue.
Edit: Also, even outside of video games, I don't remember people being obsessed about software design 10-20 years ago the way they are now. You had some design principles of course and there were obviously bad or redundant ways of doing things, but it's on a whole other level the past 5 years at least from what I remember.
Next you'll be telling me that using number of PRs deployed in a week as a KPI is a bad idea.
I think what's happened, so far at least, is productivity has been moved around.
Anyone doing consulting has seen it, where now the "I vibe coded this last week" competition is way more intense. Basically the people with the problems now have a chance to spin up something that looks like the solution they want, they then get in trouble and need someone to sort it out.
LLMs _are_ great tools for software development but if you haven't noticed their near complete inability to reason about why what they're doing might work you haven't been trying hard enough.
Most people will pocket AI gains for themselves.
So from the top it doesn't look like much has changed, whereas workers are trying to offload as much work as they can onto their claude subscription.
People in desperation to min/max work per unit dollar, will leverage AI to free themselves from as much work as possible, while still claiming credit for the work.
So as the labs start ratcheting up the price, people will pay more and more to keep their "secret" assistant.
Wouldn't that explanation predict that lower ranking employees would be more impressed with AI than higher ranking ones? Because what we actually observe [0] tends to be the opposite. Executives are the most bullish on AI, followed by Managers, with employees having the least adoption.
[0] https://businesschief.com/news/why-are-executives-using-ai-m...
Those studies are of workers using company provided AI.
They call out the MIT study in the article, which found very low adoption of corporate AI initiatives.
However if you read that study, they found that almost every worker uses AI heavily, but they just use their personal AI account.
Again, people are using their own AI to benefit themselves, it makes the most sense. Using company AI shows that the AI is doing xyz, and maybe perhaps you aren't that valuable. Using your own AI makes it look like you are doing xyz, and you are the valuable one.
I guess there is an inverse effect to bubbles (things go linearly up and then abruptly down), where nothing happens for a long time until all things happen at once.
Here’s what I’ve seen: it can make skilled people more productive by acting like a “mech suit for your brain.”
For unskilled people, it creates a false sense of productivity, but if the user of the tech does not understand what they are doing they can’t apply the tool productively or judge its output or deal with the things beyond its ability.
So like all other tools before it: it works best in the hands of a skilled user.
Just like in all other fields.
There is no tool that makes an unskilled user skilled, but there are many that can amplify or extend the capability of skill.
> it can make skilled people more productive by acting like a “mech suit for your brain.”
If mech suits existed and people actually lived in them they would either get super fat or super skinny. In either case they'd lose all muscle mass.
All technologies are a double edged sword like this, simultaneously strengthening and weakening.
https://newlearningonline.com/literacies/chapter-1/socrates-...
Humans are tool makers and this is what happens when you make tools.
Problem is AI weakening our brain, arguably the only thing that distinguish us from animals. As if decreasing brain size of modern humans was not enough.
I am surprised they don't record a decrease, maybe this year could be better?
Not really. Most things take more time to transform than people think, even if capabilities exist. Which is not to say that they do exist, but rather, independent of whether they exist, that this is not very surprising.
If programmers play it right that'll be us working less and delivering the same results. Who said everything has to benefit the capital class?
Speaking as a programmer, the net effect so far is that we work more and deliver far more results and get paid the same or less.
And I run my own firm... but I have to compete against a global market of other programmers using similar tools who are also trying to compete on price, and then there are hordes of new entrants who can vibe-code things that superficially look like they'll meet a customer's needs and are better at marketing than I am, so my area of business gets constrained to just customers who need help after the problems with the vibe-coded solution appear, the vendor who made it is long gone, and they already spent an inordinate amount of money on the vibe-coded solution since they thought it was complete and it looked pretty good.
Most of my past roles had very simple code at the important deliverable level, on bizarre and arcane platforms that would have little context for frontier LLMs. Most the creativity and complexity was in tooling for research/development/testing.
Agentic coding could help a lot for the latter, not necessarily the former.
Programmer work product (typically) benefits the capital class, directly via product engineering or indirectly via platform engineering.
If the aspiration of a 10x or more programmer is enabled through AI then the capital class win, in the typical case. Software eats the world.
What's the play for software engineers that lands them in a better place but doesn't leave money on the table for the capitalists?
> A back-of-the-envelope calculation finds that covering aicapex through identifiable ai income requires revenue on the order of $2.5trn per year, more than tech’s entire combined revenue today.
…
> All these complex calculations roughly tally with a much simpler one: adding up the ai revenue of the firms selling most of the ai. Anthropic pulls in perhaps $75bn, annualised; Openai makes tens of billions; Google, via its ai model Gemini, and Microsoft probably get a bit less. SpaceX may have a few billion dollars’ worth of revenue from enterprise aithis year. Meta also makes a few bucks from ai. Add this up and you land at roughly $150bn a year.
$2.5tn is about 8% of U.S. GDP. So a handful of AI firms really expect that they will become larger than the entire U.S. sectors of manufacturing, government, or healthcare?
Investors expect that they will, I guess. If you really believe the strong ai use cases (like exponential improvements, or it replacing hundreds of knowledge workers at every company" then maybe this isn't unreasonable.
If many people lose their jobs, thus consumers having less and less money, where would that money come from?
Depends how many people, but don't forget that:
1) people like me exist: I don't live in the USA, yet it is possible for me to buy things which are made in the USA
2) fiat money is only created or deleted by laws the government controls (plus a tiny quantity from forgery and damage to coins and notes), so everyone losing their jobs doesn't make the money disappear, just who has it available to spend
3) previous waves of automation have created new business opportunities; while this has not been too good for the people who lost jobs to the automation, it has generally boosted the overall economies this happened in, so it is absolutely possible to wipe out tech employees without it seriously harming consumers collectively
If you look at recent trends in consumer spending, a lot more of it is from the upper classes and less from the middle. Selling services to the rich can displace the loss of white collar workers. (in theory). as a bonus, ideally you create some more really wealthy people to spend on things during this shift.
I'm yet to see what part of AI really wealthy people would like to spend their money on. It's not like we can get AI Yachts and AI Rockets. What would the premium product for the upper class be?
I guess the bet is that they will become embedded in basically 8% of the tasks in every sector. which I guess could happen long term? thinking about the whole ass economy in this way always make my mind boggle so I don't really feel equipped to talk intelligently about this.
I suspect that they believe they can integrate with a substantial portion of not just the US economy, but also most of the European and formal Latin America economies. And I also suspect that the number is quite high because there are so many AI companies planning to snag like 80% market share for AI.
The share price is based on them eliminating entire sectors yes.
I thought SpaceXAI was making like $25 billion a year renting out hardware now?
Why would capex need to be covered by income during a time of investment and infrastructure buildup? Did income from the Apollo program cover its expenses?
Public investment was responsible for Apollo, so there was no expectation of profitability. Private investment is largely responsible for the AI buildout, and private investment isn't doing that for charity
If I give you 5 trillion dollars I want to get that back plus a percentage increase on top for my troubles, right? Government programs might run without the expectation of paying back taxpayers but investors are less generous.
Investors can be very generous if they see opportunities for market capture of the next big thing
Also they might not be as smart as they have made out to be. This time it seems they might be spending more to capture market than the market is worth.
Still, there are plenty of people who take their cut of the money on the way. So at least those people will be happy.
Because most of the capital assets depreciate over a period of 5 years or so so you have to make the expenditures back during that time or you can't book any revenue.
The Apollo program wasn’t for profit and if it was it never would have yielded a return on investment.
> Exponential View, a consultancy, counts $175bn of generative-AI revenue, on an annualised basis, in June. In a recent paper Anton Korinek of Anthropic and Patrick McKelvey of the Bank of Canada estimate total “AI services” revenue. Adapting their methodology, we reckon this was $220bn (again annualised) in the first quarter of this year. Ramp’s data imply that 2-3% of business spending now goes on AI, pointing to $170bn a year.
So between $170bn-220bn in annualized AI revenue today. Maybe it doesn't cover trillion-dollar bets but this is a very substantial number.
And the total capex spend to date is roughly $1tr so revenues are ~20% of capex. It doesn't seem ridiculously out of whack.
They don't have to do the future spending unless revenues are coming in appropriately.
(source on the capex to date https://valueaddvc.com/ai-spending)
This is also at the end of the tokenmaxxing era, so I'm not sure I would draw a straight line out from here. A LOT of companies are clamping down on token costs right now.
I wouldn't bet on token consumption going down any time soon. In the past year or so, the increase has been exponential [1].
[1] https://en.macromicro.me/charts/148532/world-openrouter-toke...
Even more companies are just starting AI/ML initiatives. The tokenmaxxing thing only happened at a handful of tech companies, most companies don't have unlimited money to do this kind of thing.
Once the Pathfinders find out where the real value is there is much more room for growth from companies who have be mindful of budget.
You really see the first mover disadvantages in US AI labs. And how second movers (open Chinese labs) can disrupt them. First movers tend to get overcapitalized and way ahead of their skis. You saw a smaller version of this in the vector database market (remember that?) where companies like Pinecone were getting billion+ valuations for capabilities which now seem like a commodity. There's vector DB companies now with much less debt / VC obligations that don't need to clear insane revenue levels to justify investment.
It is not first mover disadvantage. It is "took too much initial hype and debt while ignoring reality and long term" long term disadvantage.
> And how second movers (open Chinese labs) can disrupt them
Yes, I will concede this point. But how are the Chinese labs going to thrive with giving the weights away for free? Will they suffer the same fate open source database companies did at the hands of AWS?
For the companies it's not clear.
For the state, like BVD, the goal is to be a loss-leader for Chinese infrastructure + manufacturing.
If the model is commoditized then the AI race becomes competition over building infrastructure to support it. China is very good at building infrastructure.
They don't need to run a business, they are a Chinese state enterprise. The tax dollars of Chinese workers covers the cost. China is a communist country with mock free enterprise, it is not the team blue version of the US.
The thing is, AI is a productivity boost but staff drive all the productivity. The catch 22 is AI is expensive so in order to provide it to everyone they need to reduce staffing. The result is remaining folks are able to produce more but the organization as a whole is not exceeding pre-layoff output. And now things are being dropped on the floor and falling in between the cracks -- which impedes efficiency and velocity. Happening at my company now.
AI inference needs to get cheaper or there will always be this "terminal velocity." I suspect AI labs' incentives to reduce costs is only where there is overlap to free up hardware/utilization (to then provide inference to more paying customers.) I highly doubt they will want to make things cheaper for users -- they have debts to pay.
Open weight models are the way and forward, it is the only way an organization can truly control costs by self hosting or buying cheaper inference. Relying on closed weight models is a business risk. Kimi models are only marginally worse than opus but significantly better than the bleeding edge of yesterday's sonnet.
https://archive.is/VktAK
I kind of think the AI boom is the worst for (non-inference-providing) companies. If all companies are using the same "frontier" LLMs, and if they are competitive, what gives one an edge over the other? I think, just the people. Which was the same as before, but now with the additional AI spend that they can't cut, or they become less competitive.
Domain expertise in whatever niche they are in. Like Palantir is not competing against dozens of similar startups in the shady military data space.
In introductory economics courses, they teach a simple result showing that in a competitive market, profits tend toward zero.
> If all companies are using the same "frontier" LLMs, and if they are competitive
The answer is that all companies will not be using the same frontier LLMs competitively
This will be yet another technology that will benefit from economies of scale. Big businesses and those in PE portfolios will lead the charge, adopt AI, increase efficiency, and gain market share at the expense of mom and pop shops.
This will be good for our 401ks because we're all primarily invested in big business
It makes sense that if software turns to a commodity by AI, and models are also commoditized, then distribution is all that matters. Anthropic is therefore attacking other's distribution in order to protect its business. Perhaps they should focus on their own ways of distribution.
Attributing to per-worker might be the wrong thing. I.e. how many kWh power does a person buy direct vs. use indirectly through other trade.
In other words usage per regular worker doesn't matter. Revenue overall does.
A couple anecdotes:
I noticed that were a lot of traditional, non-tech companies interviewing for AI engineers in the Feb/March timeframe that have halted hiring in those roles entirely. It seems that if those roles didn't close by mid-April that they didn't close at all. This seems to match the timeframe in which cost suddenly became prominent in the AI zeitgeist.
I don't know _anyone_ outside of SV who has successfully replaced even a single employee completely with the current models. Maybe someone has pulled this off in call centers, but the POCs have all failed.
I'm very much pro-AI, but I just think we're on a false summit. As the article points out, there is absolutely no way to recoup the investment costs unless the models allow companies to start displacing human workers by the millions _and_ recapture a significant fraction of the displaced workers total comp. If either of those aren't true, then the bubble is going to pop... soon.
I mean, all one has to do is review https://isaiprofitable.com/ to see just how bad it is. None of this is a profitable venture. It's just a money pit for all players involved, all the while draining our ability to buy electronics and drink water. This has to end at some point.
It's kind of funny. I was looking at the top of the page thinking 1.5T investment for so far 770B revenue isn't bad at all, that could break even after some time. But then you look the listed companies, there costs vs revenue.
Turns out 516B of the revenues are from Nvidia, which is not an AI company, but just selling shovels. Nvidias revenues are effectively costs to actual AI companies.
Hopefully soon. Not for personal or altruistic reasons, I just find great pleasure in seeing cults burn(which is what AI has become).