That's funny because this still tracks when you use the Craiglist-interpretation of "Tina" – still, there is no alternative when you're smoking it [0]...
[0] I do not recommend smoking it, with well-over a decade "clean" (currently watching a good neighbor succumb to it, which is awful to witness)
I don't know a single white collar worker who isn't using AI for their job. Not like forced, but like "Oh damn, this bot thing can do a lot of tedious leg work for me".
To think that people won't pay $60-$80/mo to continue using it is wild to me. In a white collar environment it pays for itself in a few hours of use.
If you focus on how much value AI brings to people (mostly in time saved), the bubble hardly looks bubbly.
Did Uber die when a trip across town went from $3 to $13? No. It's giving more rides than ever, 10x more than when it was $3.
I'm sure there will be some losers, but unlike the dot-com boom, the entirety of the population already has all the tools needed to leverage AI.
$80 per month in a 5,000 people Enterprise is 400k per month which translates to ~5m per year. That is easily in the 8-16% of their total IT budget. I don’t know you, but companies don’t like to increase recurring cost. Is AI good? Yes. Is it clear how it generates enough ROI to justify a material IT cost increase? No. This is the problem now, not how much white collar workers use AI, but what is the productivity narrative of it. Because AI is so generic, there is no clear specialized use case, like when Salesforce invented the CRM for example.
The thing that's a bit different about AI from a CRM is it translates pretty directly into time savings. You only need each employee to save 1-2 hours of work per month to break even with a $80/seat subscription.
It's true that it can easily save 1-2 hours per month, but it only seems worth it if those hours saved are spent doing something else productive/profitable.
It appears that a lot of the time saved is spent either 1) doing nothing instead, 2) waiting for the AI to finish, or 3) increasing the hours spent elsewhere down/up the chain for reviewing, fixing, auditing.
The obvious benefit of anything offering to save time is that now you have time for "higher order work", but often, we use the time saved to be lazy.
After all, that's the underlying reason we used the time-saving, corner-cutting tool in the first place.
People will pay out of their pockets if an AI can free up a few of their hybrid hours a month. And already it looks like the vast majority of white collar AI spend is from individuals, not companies.
I mean, why would you show your hand by using the company AI? Use your AI and take the credit...
But the question is how much are people willing to pay for AI.
I use AI every day at work, but I only pay $20. That's the most I will ever want to pay. And, so far, it gives me everything I need.
If OpenAI or Anthropic suddenly said "Sorry, the game is up. You'll have to pay $100/month now", I would 100% look into cheaper Chinese solutions.
I suspect the AI subscription (or API) economy is whale economy. You have a small, small percentage of users that are happy to pay whatever it takes - while the vast majority will either use the free tier, and then the next group will pay for the cheapest or next cheapest subscription.
EDIT: And I'll echo what another user wrote here. The VAST majority of office users around the world don't work for tech companies flush in cash. Even adding something like $20 - $100 subscriptions to every user is a serious enough financial obligation that it needs to go through budget planning / boards. Even more so for all the government workers around the world.
Rofl, I already went for cheaper Chinese solutions. I can achieve more than I could when I was giving OpenAI 20 bucks a month, and for a fraction of the price.
I have no idea why people still use ChatGPT or Claude.
Most people don't understand how cheap models like DeepSeek and similar are. I have $5 in an account that I use for months. I can get very complex code for less than 50 cents.
Me too. Sinophobia may be the culprit, too many years watching fox or cnn is my guess?
I love chinese phones, cars, cities, people, and now AI models. There is a beautiful world out there to admire if people is willing to open their eyes.
You will spend thousands of dollars on hardware to run those at lower quality (quantized) than the benchmarks where they match March SOTA performance.
Source: I have the hardware to run those locally. I would never recommend it to anyone trying to save money. It’s so much cheaper to pay even Anthropic or OpenAI.
I spent $18k in credits last month, I have 3x 20x Anthropic accounts (for Fable to bypass limits) which runs me another $600ish p/m and then a chatgpt pro account another $200 p/m. The value I got out of it is easily 1-200x what I paid.
What are you using it for? I have to assume some kind of agentic workload. Also curious to know when (if ever) you would pivot to Chinese providers to save $$$
1 to 200 is a pretty big spread to between losing 18k dollars and making 3.6 milllion - do you have any actual numbers on the value produced from this 18k investment?
Even adding something like $20 - $100 subscriptions to every user is a serious enough financial obligation that it needs to go through budget planning / boards
I doubt that. In my tiny town in India, office workers use the 1800 INR (20$) plans. 1800 INR/month is like... 4% of the salary these guys get. And since this is mostly MS-office and windows explorer and chrome based stuff in very small, non-tech companies, it pays for itself in literally a day. If they _could_ pay less they would. After all they pirate MS office. But the cheap chinese plans dont have any distribution, whereas OpenAI seems to have effectively marketed to them via IPL ads and such. No one there even knows about deepseek. It also doesn't help deepseek and such are focused on the coding market. No multimodal, office plugins, etc.
For personal stuff, people there use Meta AI - mostly because it is directly available in whatsapp and these days they are pushing it by adding a dedicated button for it right on the home page. Regardless, they still call it "chatgpt". OpenAI brand is unmatched.
Remember that 20$ is a subsidized rate openai is currently willing to provide. Once that goes down you think these guys would be willing to pay token based billing charges ?
It is not going to go away. Rather they have doubled down and opened 399 INR/mo (4$/mo) plans that as of late have GPT 5.6 Luna.
The reason this works is that you get lesser inference time compute used for queries on these cheaper plans which makes it sustainable and this is enough for the tasks these guys do.
And some local telcos are bundling this subscription as well, so most people just get it for "free". For example, I get Google AI Pro for free with my 350 INR/mo telco plan.
These companies have spent billions of investor dollars and they will need to recoup that cost soon. And then show year over year growth on top of that.
Unless they can massively scale down training and inference cost or implement AGI I don't know what their plan is. Just provide a subsidized plan for the next 10 or 20 years? Their costs are directly proportional to the amount of tokens the LLM produces. How is a monthly subscription plan supposed to account for such costs?
They dont need to scale down anything. AGI is a red herring.
Even Deepseek at its absurd prices is a very healthy business. Regarding their return on capex multiple, their CEO said they make a six-fold profit on their compute capex with 10 month recuperation. Because of this, all of them are spending aggressively on compute. Apart from that, user acquisition and data labelling are the major costs that are preventing net profitability right now. High quality data labelling is said to not have a cost advantage in china etc as well and they pay global market prices for this. I can confirm this is true in india too the model companies I know pay global market rates for high quality data.
Their costs are directly proportional to the amount of tokens the LLM produces. How is a monthly subscription plan supposed to account for such costs?
By limiting the number of tokens you use per month? per week, per hour? And by limiting the inference time compute dedicated to each turn in each session.
need to recoup
the world economy has shown itself capable of handling decade-scale recouping easily
The main obstacle today in the inference business is the high variability in usefulness/token. This does not need to be solved, but rather only quantified. Innovation is needed to be able to reasonably bound this variance for a reasonable subset of tasks. And we are making progress on this. Naturally though, tasks on the frontier of current capabilities have very high variance. The last couple of years has followed the pattern where tasks no longer on the frontier have reduced variance, but I am not claiming this will continue to be the case generally as the frontier improves.
their return on capex multiple, their CEO said they make a six-fold profit on their compute capex with 10 month recuperation
I am not familiar with chineese model companies as much as I am with US based ones so I don't have much to say beyond that the CEO is incentivced to pump up those numbers.
By limiting the number of tokens you use per month? per week, per hour? And by limiting the inference time compute dedicated to each turn in each session.
If this was so simple I don't know why GitHub copilot went to token based billing at my company.
This does not need to be solved, but rather only quantified. Innovation is needed to be able to reasonably bound this variance for a reasonable subset of tasks
It's much better to make a business case for them after finding this bound right? Currently I can't use copilot for anything serious since I cannot predict how many credits one request is going to consume.
Your point about non frontier tasks using less tokens makes sense. As you said, let's see if it holds up
Even Deepseek at its absurd prices is a very healthy business. Regarding their return on capex multiple, their CEO said they make a six-fold profit on their compute capex with 10 month recuperation.
That's "theoretical profit" - in some imaginary world where the free subscribers would pay the top tier cost.
These companies have spent billions of investor dollars and they will need to recoup that cost soon. And then show year over year growth on top of that.
I remember the piracy era, and it used to be a joke whether anyone could actually stop it.
Eventually MS dealt with piracy by going after firms, and finally by offering something cheap enough.
Between the legal aspects, and the fact that its MS office, paying Rs 1800 a month is possible.
I’ve been trawling every source I can find to understand what the story on the ground is, and when it comes to productivity it’s a huge mixed bag. The variance in outcomes between independent coders, frontier labs, someone in SV and someone in India is mind-bending right now.
Most firms which talk about their AI plans are not seeing traction, and the AI projects are going to the same place that the ML projects used to go to die.
It’s at the individual level that I am seeing productivity gains, however that isn’t something firms are happy to hear right now.
Eventually MS dealt with piracy by going after firms, and finally by offering something cheap enough.
Sure, because they realized that selling software at any cost because software has extremely low overhead. Selling 100M Windows at 2 USD was easier and my guess is a wash at that price.
However, AI is not selling software, it's selling hardware usage which is not free and has real cost. If Rs 1800 which is 20 USD is not profitable, then they will not offer it.
First I'd like to understand what are these "small" non-tech companies in tiny towns 45k to its employees? The last labor survey had average Indian salary pegged at 21k.
it pays for itself in literally a day
Most comments in the thread say this but don't explain the how. How does it pay for itself "literally" in a day? Through productivity/time savings? How exactly?
I can only assume people are talking about productivity gains. Have their been reports of increased productivity or layoffs in tiny town, small non-tech companies in India that I am unaware of? What is the source here? Without a concrete source this is mostly anecdotal evidence and not backed by data.
Today there is novelty factor of AI. Questions or tasks which required people to really plan out now can be tested through AI quickly. To many people that seems like progress and that AI is paying itself. But most of it is busy work and isn't adding value to the companies. Given that everyone is on the AI hype train they might not be looking at the expenses closely. With time they will start poring it over and cancelling contracts en-masse.
They processed 10 cases in a day rather than 7. In some cases, new work took its place and the company makes a little bit more money, but this is a lagging effect naturally.
Have their been reports of increased productivity or layoffs in tiny town, small non-tech companies in India that I am unaware of
PWC et al don't go around doing reports on these tiny companies bro. The only information you get is if you know people in that place and they tell you yourself, which is how I got to know.
Given that everyone is on the AI hype train they might not be looking at the expenses closely
Yeah maybe F500 ceos exposed to LLM coding agents are on the hype train to some extent, but these M.S office pirating small companies are not on any hype train lol.
The last labor survey had average Indian salary pegged at 21k.
Averaging across the whole of India gives you a useless number. In the richer regions, 21k is less than what a full time food delivery/quick commerce driver makes.
PWC et al don't go around doing reports on these tiny companies bro. The only information you get is if you know people in that place and they tell you yourself, which is how I got to know.
I knew the post screamed "trust me bro" vibes and now I know for sure. Government does Labor force surveys. Here's one for last year. Point me out how the participation is changing through AI paying for itself.
https://www.pib.gov.in/PressReleasePage.aspx?PRID=2246009&re...
Averaging across the whole of India gives you a useless number. In the richer regions, 21k is less than what a full time food delivery/quick commerce driver makes.
See the above survey for the figures. But then this is like someone using California to say how well America is doing.
What do you expect to glean out of PLFS though? Why would broad based labour force participation or other similar data change due to AI paying for itself? What?
If you meant to somehow glean some information out of the wages report, note that the wages survey was last done in 25. The quarterly and monthly bulletins don't contain wage data. Regardless, these are too macro to give any sense of any change at all happening in AI-accelerable knowledge work.
By "PWC report", I meant one where they went to firms and recorded using the same methodology some same notion of AI "paying for itself" or "increasing productivity".
You cannot wear out an SSD through AI inference alone. LLM weights are only read from the SSD, and read operations do not contribute to SSD wear. SSD wear is primarily caused by write and erase operations.
Have you been trashing the non-replaceabe SSD with constant AI workloads?
Your other point is valid, but you're either vastly underestimating how many reads/writes a modern SSD can take or vastly overestimating how much output a typical LLM is capable of.
Humans do die eventually though, which is an even more severe form of obsolescense. I don't think they'll be AGI either, but that argument doesn't seem super compelling.
On the other hand you have the old saying "science advances one funeral at a time", indicating death solves at least as many problems as it causes.
(the point being that old, powerful professors have more than once blocked progress in their fields for decades. And only a funeral, eventually, solves the problem ...)
anthropic and openai are a part of the ai economy, but not the only. the most valuable company in the world is nvidia, a massive player that benefits from the existance of the ow models.
what im trying to say is that we focus a bit too much on the labs, but the biggest part is in the infra that makes everything possible.
But AI use translates directly to time savings (if it works). You only need 1-2 hours of time savings per employee per week to break even on a $100/seat/month subscription.
How do people mentally price things. I'm on the fringe but I'm willing to pay $60 for a AAA 60 hour game that I'm positive I won't finish but am find if I get 6-10 hours out of it.
2. Cell phone usage was always localized. You had local telcoms and that was about it. Deutsche Telekom Germany didn't have to compete all the time with NTT Docomo Japan.
And there are probably 20 other economic differences I'm missing.
The video quotes sales of $2.5 trillion a year to payoff those investments. What workforce you divide it by is a bit up in the air, but let's use 500 millions, which is basically 100% of the workforce of US + Europe + some change. That gives you around $500 a month per employee (from hedge fund manager to flipping burgers at McDonald's).
And there will be competition. I see zero moat right now. The user specific part of the state of the model sits outside of the control of the model (it is basically my code base, or my prompts, all of which I can transfer to any competitor).
Speaking of user experience: OpenAI has some pretty cool tools, some of which even have decent UX. But I’m amazed that anyone can get real work done with the first party web UI. It is UNBELIEVABLY LAGGY. It makes GitHub seem snappy. It scrolls to a black screen and slowly populates later. It takes tens of seconds to load conversations. On the rare occasions that I’ve tried the fun “ask Pro mode a math question” I think I’ve spent 15 minutes waiting for content to populate in the UI. Oh, and all the obvious features like forking a chat either don’t exist or are somehow hidden. This is on a top-of-the-line client machine.
Maybe other users are doing something differently or are being served a different front end JavaScript blob?
The iOS app is hot garbage as well. The list of old chats will have topics randomly jumping around which makes it really annoying to try to choose the scheduled task that notified you.
This is the way it is on Wall Street more each century, and for AI to go public on that exchange it's going to have to go big or go home. Considering the amount of money that has already been spent privately.
There is no alternative pipeline to replenish those reserves, and different people have different ideas about capacity and bottlenecks relative to ambitions and what they are supposed to get for their money.
As has been mentioned in another comment, there is no alternative foundation other than optimism either. The exchange wouln't exist if it weren't originally intended to trade only shares that were all worth holding otherwise. Naturally some worth holding more than others. This is so big it will be necessary to be able to fool way more people with way more money than usual, otherwise those that prevail will have nowhere near their wildest dreams come true.
Naturally AI is never going to fly off the shelf like it could until it starts getting cheaper all the time for huge jumps in performance. Cheap home computers will need to be able to do quite a bit more than they can only do today while connected to a massive AI data center, without ever having been connected to anything outside the home at all. Otherwise AI can not ever be considered "general" any more than computing could be considered personal, until you were no longer reliant on a remote mainframe in a huge out-of-state data center somewhere tied by a thread leading through a squeaky modem over monopolized communication lines. This differential between clunky (clanky?) old data centers' overall computing power relative to the amount held freely within homes & businesses is something that looks like it could be regaining exploitability like never in decades. I know one day Woz jumped right in without needing to be a greedy businessman because there was no possible downside, and he could let just about any dedicated growth leader make as much money as they wanted off his technology. Plus Jobs was no slouch in many ways and Woz never needed to worry with a product that sells itself to begin with, putting Jobs in hog heaven where he could persuasively bring it to the next level almost whenever he wanted like you can do few other ways.
So not just dot-com related hardware & software.
That goes for memory and storage too, people should do the math on how affordable nominal amounts are supposed to be by now. If everything were still normal 16GB of DDR4 would be about $10, at least by 2027, so it hasn't merely doubled in the last year which is the most obvious part, it has skyrocketed to 10x what it would have been. And still rising not falling, so all this is going to have to be reversed for consumers to even afford what they used to be able to do.
For people who didn't want to spend a thing on AI until it's naturally way more capable and constantly getting cheaper like it should be, it's pretty frustrating when it's already been unavoidably costing "indirectly" more than they have been willing to pay if they were getting maximum benefit, when most of them have only gotten more surveillance. One day "indirectly" was just no longer true through no fault of millions of people. And growing as fast as the money stream will allow.
Up until recently AI was way more niche and fewer ordinary consumers were exposed, where now there are millions more aware of its influence and growing fast. But the more aware more ordinary people become, millions more opinions are going to come to the surface and need to be dealt with.
whale economy.
Thar she blows!
One pervasive opinion is that so far it's made by rich people for rich people, and more so than most those who want to get rich quick are joining the bandwagon. Of course what consumers see in the media is only the "band directors" who are already rich to begin with, they set the example because they now want to get richer quicker and this is the vehicle they have chosen to do that. This is not unexpected, after all they always get what they want when it's things you have to be ungodly rich to do.
This makes for company valuations based more on optimism than anything else and once that has eclipsed underlying worth then any unforeseen bottlenecks or financing stumbles become highly magnified.
For one thing there may not be enough momentum to even fully build enough data centers to fulfill some players' plans for recouping such large investments, but at the same time AI's not going to really get good until data centers are not needed at all, which is so ominous there appears to be even more force being applied to encourage people to ignore things like this. Otherwise it could get too shocking.
This disparity in upside just plain instinctively sidelines more ordinary people as it builds, so the number of people who are just going to have to wait for AI and everything associated with it to start getting cheaper all the time becomes a force in itself. It'll be easy to notice without being a finance guru.
To put it into perspective if we consider that US salaries are 3 to 5 times those of other "rich" countries then it's already a 60 to 100 dollar subscription for the next richest part of the world.
I have a devils advocate argument: AI spend doesn't need to scale with salary.
If a company was OK with paying 10k a month for an employee in the US because they considered their work more valuable. And the same company is only willing to spend 300 per month on someone in a 3rd world country, because they consider the work to be lower quality.
Wouldn't it be reasonable to consider a situation where the same company would be willing to only spend 100 per month more on the 10k employee to make them a bit faster. At the same time be willing to spend 700 per month on the 300 employee, if it levels out their capabilities and makes them nearly as valuable as the 10k employee?
--
This argument is riddled with assumptions, but I do feel like it's one possible direction that some companies might try to take.
But the question is how much are people willing to pay for AI.
That depends on how much more money they can make with AI or save with AI. It shouldn't be very hard to know that number in a well run business. Even for a one man business.
If they can get productivity increases equivalent to .1% it pays for itself. That's roughly equivalent to firing 1 in 1000 employees to finance the software.
After using it a bit I’m convinced it’s not, as a whole, a bubble.
Bubbles exist around it. Data center construction may end up overbuilding, much like the dark fiber thing during dot.com, since better chips and models will shrink power and space requirements over time.
But the core tech is the most powerful new thing I have seen since discovering the Internet, at least. Maybe more.
A bubble doesn’t mean a technology is useless. It just means it’s overvalued
In the last 30 years, we’ve had multiple bubbles in the tech industry. Every one of them was backed by real stuff that we still use today. They were still bubbles and a ton of people were hurt when they popped.
The railway bubble and the optical fiber bubble left the world with lots of railways and optical fibers. Once their original owners went bankrupt, the assets themselves were quite useful and changed the world.
The cryptocurrency bubble, however, didn't. Not really. At best you can use it to buy drugs. Some of the GPUs may have been repurposed for AI, but not enough to make a dent in AI, and the AI reesarchers who kicked off the AI bubble were having to buy their GPUs for high prices to compete with the miners, not low prices after they went bust, so we can't even say miners helped kick it off.
Many of those assets were useless - like railways from nowhere to nowhere. Only some ended up being used.
And many innocent people ended up being hurt in the process.
Bubbles are not good. They are harmful. The ability yo recoup some of the losses in the span of next 15 years does not make the bubbles something positive.
Even railways from nowhere to nowhere often ended up being useful when the ends were extended to somewhere or when somewhere grew to encompass nowhere.
For example the Berlin Ringbahn was once some billionaire's toy but now it serves very useful traffic, as Berlin grew outwards to where the Ringbahn is.
It is not "often" it is "sometimes" and it is also less than a participation trophy. It does not make the bubble not harmful or even less harmful. It just means that sometimes you can repurpose an expensively build junk.
bubbles are bad for society too; families get thrown out on the street, big players/individuals buy all the cheap/discounted assets furthering economy divisions, and depending on the size civil unrest etc....
This new technology more powerful than the internet can't answer the simplest questions straight. If this is the future then I don't want it. LLMs are pretty great at generating code and recognizing patterns and that's about it.
That’s not the argument though. The argument is that conversion rates as per OpenAI are about 5% and to serve the other 95% you need to build significant infrastructure that costs hundreds of billions annually. So for every white collar worker that you know and uses it and pays for it there are 20 free loaders. And mind you, 5% conversion rate is extremely high already as a number per se. It won't magically reach 20% in a couple of years. Chances are it will fall particularly for the enterprise segment due to the advent of the Chinese models.
The reason uber didn't loose that much business is because it replaced the established businesses and left most consumers with little alternative.
That's not the case with AI - unless the frontier labs achieve AGI or some sort of super intelligence that lets them create infinite economic value (at which point, any further discussion is pointless for obvious reasons), the average worker can do most of their tasks with 5% of the api costs using an open source model from China and achieve the same results.
The reason uber didn't loose that much business is because it replaced the established businesses and left most consumers with little alternative.
The established businesses were awful. There was not a thriving and beloved taxi service in every city pre-Uber. Taxis were bad and very expensive.
Also did you forget that Lyft exists and normal taxis are still available? Taxi services had to improve their business when the ride sharing companies entered the market because they previously relied on their own grip on on-demand transportation to keep prices high and service poor.
Imagine suddenly a new taxi business was able to come into your city and provided taxi rides for 5 cents on the dollar. What would happen?
Edit: now add on top of that that even if uber tried to undercut them again by running losses as they did in the past, that new taxi business would not loose any money if the number of their rides went down because their operating costs simply scale with the number of rides - while uber was loosing money on every ride.
We're talking about corporations here. Of course they use Linux, and of course they'll provide their employees access to the LLMs that cost half the price and are 90% as good.
How is the effort switching from an operating system containing all your data and apps, which might not be compatible with the new one, comparable to just switching an api key?
How many are the white collar workers? 300 mln? even with $100 subscription each, the monthly revenue for your addressable market rises to $30 bln. If we increase the figure to half of humanity and they are 3 bn, you multiply it upto 10 and $300 bn monthly, $3.5 tn annually.
Now start cutting it down keeping in mind that not everyone is an office worker, does not need $100 plan, has access to cheaper models, could get ai through someone elses subscription or local models. I'm not sure how down it will get, but it can get down a lot and will have to be split among few players. At least for me it does not seem unrealistic that the revenue won't be enough for everyone.
The real question is not whether there will be a market for AI, it is obvious that there will be one (as it was obvious the Internet was a gigantic thing in the dotcom boom). It is rather whether the value will be in the models (and if so in which?) or the infrastructure or somewhere else.
Interesting thoughts from Dwarkesh Patel on the future of models [1]. In summary, no moat if a client can switch to a competitor from a dropdown when all models are similar-ish (so models are commodities), but to really compete with humans, AI needs to start "learning on the job", i.e. accumulating experience like an employee (as opposed to the training / inference approach where the model isn't learning after it is released). If your model does that, then you have vendor lock-in as you can't replace an experienced model with a competitor. Then the value is in the model.
As in the 90s, this could go in all sort of different directions, hard to make predictions. And remember the 90s: the value was in the browser (you had to purchase it, eg Netscape), then the browser came for free with the OS (no value in the browser), then the browser was the most important thing in the world as it was control over the default search engine, everyone thought portals was the most important thing to control, so ISPs had their own applications accessing the internet from their own portal, etc - none of those people were stupid.
I work for a large company and for any LLM I use professionally, whether programmatically or via a UI, I get to pick from a dropdown from Gemini, Claude, OpenAI and the open weight ones. We ain't locked in.
Yeah, and that is completely fine business-wise for all the businesses in that dropdown. Averaged over the whole TAM, they will all make a lot of money. The reason is that the total market becomes _bigger_ with every new model (upto a reasonable point of course, but for 5 models it is obviously true), they don't end up competing for the same slices of the pie.
The core business for each of the 5 will not be companies like yours, but rather few hundred to few thousand companies each who exclusively use their model not for technical reasons, but because they were wined and dined or due to bundling with another service. The other hundreds of thousands of customers are all bonus. This is true for most categories of SaaS not just LLM inference.
If the vendor of a core piece of software that you are already locked into tells you they are raising prices and adding AI inference features, you are going to cancel whatever AI plan you have and use theirs. Or pay the amount anyways (which is even better business-wise for them).
Eventually, semi-random differences that arise in their distribution may end up informing model capabilities as well, and _that_ is when true model-level differentiation may happen. But this is not necessary and if it comes true would just be a bonus.
Not hosting in China will be a moat. Tech (software) never really experienced it, but the rest of Western industry all have stories of getting burned by China and their policy of ignoring IP rights.
I agree that it's working good enough and that it's here to stay: mostly everybody around me uses AI too but...
Did Uber die when a trip across town went from $3 to $13? No. It's giving more rides than ever, 10x more than when it was $3.
The competition from open-weights models is already putting pressure on how much this can be billed. That's today, with stuff like OpenAI announcing 80% cuts on terra and luna. Tomorrow, in addition to the pressure from open-weights models, there's going to be pressure from chinese hardware: Tencent, Alibaba, etc. are all already working on their own AI chips.
What good would, say, a SOTA open-weights model running on chinese chips (chinese chips located in China or elsewhere btw) do to, say, Meta or Oracle's insane AI investements?
And it doesn't even have to be chinese chips: what good AMD AI chips doing inference by having weights etched on silicon will do to OpenAI and Anthropic?
All the people, especially here on HN, who were explaining six months ago that OpenAI and Anthropic models where so good for coding that coding was solved once and for all can now run better open-weights models than the proprietary ones from six months ago. For a fraction of the price.
If people, instead of paying $20 per month pay $80 per month for their AI subscriptions, what makes you believe that money is going to go to one of these companies participating in this $2 trillion debt?
And really, in which new domain have we seen the first movers being able to keep a lifetime grip on the market without getting their arse handed to them by the competition? Things moves quickly today: I'm not sure OpenAI and Anthropic are the Ford equivalent for AI.
That AI is here to stay is a given. That people may be willing to pay more is likely (even though IMO they won't pay 4x more for something only marginally better). But that the current players are going to stay on top even though the competition closed the gaps to weeks?
That’s never really how efficiency gains tend to work except in places where entire industries ceased to exist (printers, weavers, clothes washers, etc).
People may pay less than they are willing to pay as competition drives the price down. I paid $20/mo a while which I was happy with but I found I could get similar for free so pay $0 now.
This story had 49 points and 81 comments in about four hours and was on the front page. Then it seems to have disappeared from the top stories feed completely.
Google funds Anthropic
|
v
Anthropic promises to rent Google's TPUs
|
v
Google guarantees the infrastructure
needed to fulfil Anthropic's promise
|
v
Wall Street lends against Google's guarantee
|
v
Borrowed money buys Google-designed TPUs
|
v
The TPU purchases "prove" demand for Google TPUs
|
v
Anthropic's compute capacity and valuation rise
|
v
Google's investment in Anthropic rises in value
|
v
Higher valuations justify still more financing
|
+--------------------------------------+
|
v
DO IT AGAIN
That's like saying my delivery company is profitable because with current gas prices, my margin is 80%. Yea, what about all money you spent to get there? You still profitable then?
As Ed Zitron pointed out, that "quarter" of profitability is EBITDA profitability and comes with plenty of creative accounting. When they do it for a year, we can say "They have found profitability"
Why are you citing old news? Zitron wrote that months ago when Anthropic was reported to have ~30B ARR. Anthropic now has more than double that, at even higher margins.
There's no doubt at this point that Anthropic is profitable.
If they were profitable, they'd release real numbers and announce concrete plans for an IPO. They are very obviously not profitable; they just announced a $65B Series H two months ago.
Because I don't know how you can be so totally and completely wrong for so many years, and still have people lend you credibility. But I definitely do understand how you can rage farm subscription dollars from suckers for years.
Until they are public and they release full financials using standard accounting practices, I would take what they say with a grain of salt.
SpaceX also painted a pretty rosey picture of their financials, but when they filed for IPO it became clear they were doing some real cherry picking and in actuality they were losing a ton of money.
If they really were doing so well they would file publicly and be rushing to IPO.
I guarantee they are losing billions per quarter. You can say they make money on inference, but that is irrelevant. That doesn't include all of their overhead and indirect costs let alone model training.
Demand for compute is growing 10x year-over-year, faster than inference profits can fund expansion. At 80% margin the payback period is 1.25 years. Cash flow can't keep pace with that scale.
That point is irrelevant because the number comes from SemiAnalysis, not Dario's interview:
SemiAnalysis estimates that Anthropic's overall gross margin has rebounded from negative 94% in 2024 to the mid-60% range, with the gross margin of its API business exceeding 80%.
Look, that's an estimate (by a presumably biased observer). I don't really have a horse in this race, new large tech companies are fine by me, even if only to reduce the power of the current ones.
However, I find these numbers incredibly hard to believe, and most likely deceptive, given that they recently (like March) started making enterprises pay API rates, so even if they were profitable (and if they paid from Fable training from this post cost money) I would be sceptical that this will continue, given all the competition in this space.
tl;dr let's all wait for the S1 (it will presumably be soon, unless SpaceX declines get them to postpone).
But pay for who? $50 for a cheaper competitor? $20 for a chinese one? Or just buy hardware and host/co-host for privacy? And to balance the insane spending even your $60-80 example would be hardly enough from that white collar audience.
At my $day_job we are forced to use a shared virtual machine for all work activities. The Windows instance has 64gb of RAM, 8 virtual cores of old EPYC CPU and no GPU acceleration. It is also shared between 6 users.
They will not spend extra 20$ per user per month to add compute and increase productivity.
The question is not whether LLMs have value; the question is who will earn how much money off them. I don't see any scenario where the companies that put hundreds of billions into their models make that money back.
Pays for itself how? Customers don't care if a business saved time on bullshit paperwork. It needs to actually convert to increased competetiveness which leads to increased profit to pay for itself. But if everyone is doing it anyway businesses will quickly realise they're just paying an AI tax and look for ways to eliminate it. Maybe by just not doing the bullshit they got AI to do.
I avoid subscriptions like the plague and I was getting enough value from Claude that I was paying for the Max plan for a couple months. I don't know how much Anthropic needs to make to be sustainable but I think it's worth north of $50 or $100 per month for many knowledge workers.
I highly doubt anything more than a small single-digit percent of users are willing to pay $50-100 for a Claude subscription themselves. Just like everything else (if your company isn’t paying for a seat) 90% of people will just use the free tier, 7-8% will use the first paid tier, 2-3% use some kind of power user plan. It could end up being a whale-driven business model in the end.
To think that people won't pay $60-$80/mo to continue using it is wild to me.
I'm fine paying $200 per month. Companies seem to be fine at $2000 per month but I would balk at that. Companies are already balking at $20,000 per month.
So, we KNOW what the numbers are. If corporations are only willing to pay $2000 per month per developer for all the developers in the US, the amount of VC investment already exceeds what they can get back from 10 years of revenue.
Why $60/mo when you can pay $5? Why $5 when you can pay $1. Unless you happen to be in one of the industries OpenAI focuses on you would always just use a Chinese model. Even now, but especially in the next 1-5 years.
Did Uber die when a trip across town went from $3 to $13? No. It's giving more rides than ever, 10x more than when it was $3.
Uber has a fairly large moat: the regulatory mess they’ve waded through and the collection of drivers willing to drive for them. Sure, it’s probably easier to start a competitor now than 20 years ago, but there’s still a substantial network effect — a transportation service needs a lot of cars to provide good service, and they need a lot of users to keep those cars busy.
An AI service like chatgpt.com or claude.ai costs literally billions of dollars to develop, except that several Chinese companies are currently happy to spend that money and release the models and weights for free. And even a small company can rent the datacenter capacity they need, at least at moderate scale, and they only need enough scale to reliably fill up the batches on a handful of servers. There is no obvious network effect — a tiny Uber with a total of two cars in a city is useless, but a tiny AI provider that only services 2000 users is just fine as long as someone else provides the models weights.
And the costs to run these services keep coming down. The numbers people are reporting for unquantized DeepSeek v4 Flash 0731 on 2x RTX 6000 Pro [0] using open weight models and open source inference stacks are astounding, and one single copy of this could probably cover the entire inference usage of a decent sized business doing the kind of desk work that OpenAI and Anthropic could need to monetize to cover their expenses. (Using bespoke models has value but may reduce available usage of a single inference machine. Quantizing can make everything cheaper and is often good enough.)
This is not to say that these companies won’t make it on the strength of their enterprise offerings (never underestimate the willingness of enterprises to overpay for services they’re already using or by becoming the next Googles and finding other revenue sources beyond just inference.
[0] I admit to some skepticism that the pile of sketchy, obviously AI-generated patches to SGLang that allegedly achieve this are doing what the say they’re doing, but the point stands.
If OpenAI wants to generate revenue by charging its users for services, this is a big deal. If, on the other hand, OpenAI wants to become like Meta or TikTok, then they don’t actually need to charge these users.
But latter model will not be easy. TikTok does very well, but their revenues are not enough to justify OpenAI’s expenses. Meta is very, very good at capturing value from users’ attention, but they are not going to give this market up without a fight. In general, I’m somewhat suspicious that the total addressable market for attention is not infinitely scalable. At the end of the day, the money comes from marketing budgets, and those marketing budgets scale with the amount of stuff being sold and, much less strongly, with the efficacy of the marketing.
There’s a theory that there is money to be made having at AI-like system that will actually buy things for the end user. Amazon tried very hard to make this work (in the Alexa division) and basically gave up. I’m not finding my breath for OpenAI, or for anyone else, to pull this off any time soon.
And I know very few white collar workers who are, beyond "I guess Google Search is AI now". Each their own bubble, I guess.
AI has some value in some scenarios. Most models still fail badly enough often enough that I personally would not pay $80 for it. I would make use of it if it were free, or $10 / month, but it would have to significantly improve if I were to ask my employer to pay $150 / month for it - let alone $1500 / month.
The evidence is obvious: everyone and their dog has access to highly subsided AI agents right now. Literally every single company is pushing its employees to trial AI - some even forcing AI use. So why aren't we seeing a giant white-collar productivity boom yet?
The idea that every single white-collar worker is going to get a $800 / month AI subscription is ludicrous. The big question is, can the AI companies survive off everyone getting a $5 / month subscription, with a handful of workers getting a $500 / month one?
The problem is the value you speak doesn't seem to be materializing. I barely know anyone who sings the praises of saving time with AI, but I know plenty of people who complain about peers doing a shoddy job because they used AI to do their work, or having to clean up someone else's AI generated slop.
Why are there an endless flood of comments like this on HN and other social media insisting that because there’s such high demand, it’s IMPOSSIBLE that these tech companies are overvalued and over leveraging debt.
Reality is way more nuanced. Something can have high demand AND be on the precipice of financial catastrophe. See: Spirit Airlines.
It's a near guarantee that when things shall heat up and the stock market shall crash and people are going to say that "this time it's worse than 1929", we'll hear the brrrrrrrr sound of the money counterfeiters, very likely synchronizing their counterfeiting operations worldwide (US / EU at least) to print ever more money.
Wall Street makes money from issuing debt, trading debt, and IPOs. They need to get the last two mega IPOs off before the curtain call on this era.
But they can only partially influence the market, they can’t control it. The question is if the market will let them get these IPOs off or if the jig is up.
Here’s an exercise for anyone curious: pull up the median stock in the S&P 500. Look at its P/E. Take 15 minutes and look through the company’s financials and figure out what you think about the quality of its earnings.
Then decide if that P/E is appropriate.
The bubble isn’t in just AI, the bubble is everywhere, and AI is its largest manifestation.
And capitalism also works with crashes taking unprofitable companies down. And frauds, eventually, gets companies valuation like Enron and FTX to zero. And people go to jail.
Now I'm not saying OpenAI or Anthropic are frauds. What I'm saying is that, eventually, things revert to what is just.
The late 90s SV tech-bros behind pets.com or webvan for example faked it for 18 months to 36 months or so. At some point when the expenses outnumber the revenues, capitalism does its thing.
A number of famous business leaders have admitted they had no idea what they were doing in their memoirs. A few have even suggested that was an advantage.
Everyone running these AI firms is being forced to make educated guesses about the future of the technology and the industry. It's unlikely they're all making accurate predictions.
Subsets of HN just want to know how long before they need to withdraw their savings, or sell their position.
I would say growth is even expedited with high allowances for borrowing/leverage, but at the cost of increasing systemic risk.
Obviously the people who will suffer from this are the taxpayers, who don't get a say in the matter, so no one is going to stop and consider whether it's a bad idea to continue issuing more and more debt.
The creativity of the financial industry is quite insane. Combined with the scale-at-all-cost strategy in AI companies, this is a ticking time bomb, no doubt.
I think it is easy to forget that a revolutionary technology does not automatically make a viable business model. I think we are yet to see the real winners in this game.
I agree with you. Look at railroads, cars, computers, and other technologies that really did end up being world-changing. How many of the original companies are still around?
Sometimes I feel like OpenAI will end up like Fairchild Semiconductor: an early innovator that quickly became an historical footnote.
There's a kind of inconsistency in that he ends the video talking about overpriced stocks and how it's hard to short them but the title is about the $1.75tn in debt which seems to mostly be from professional investors, apparently:
JPMorgan Chase and Morgan Stanley, private equity and credit giants like Blue Owl Capital, BlackRock, and PIMCO, alongside international commercial banks
who probably know what they are doing and read the footnotes.
My guess is that the lending is actually ok because there's a lot of real demand for compute. $1.75 tn is about 1.4% of global GDP which doesn't seem that silly in the AI boom.
to modify sinclair: it is difficult to get a man to understand something, when his bonus depends on his not understanding it
As he explains here: https://youtu.be/NufJ7g63KSY?t=1233
TINA, There Is No Alternative
That's funny because this still tracks when you use the Craiglist-interpretation of "Tina" – still, there is no alternative when you're smoking it [0]...
[0] I do not recommend smoking it, with well-over a decade "clean" (currently watching a good neighbor succumb to it, which is awful to witness)
It might be more accurate to say that many retail investors are ignoring big tech’s debt.
That does not explain the buy recommendations for SpaceX
Oh that one's explained by bribery. Just pay some analyst to put out ridiculous "predictions" about your company
It’s worse. They are doing this in the hopes of getting future underwriting business.
The papers mentioned:
1. — "The Big Market Delusion: Valuation and Investment Implications" - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3501688
The paper behind the "big market delusion" he discusses.
2. - "Do Stock Prices Fully Reflect Information in Accruals and Cash Flows About Future Earnings?" - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2598
This is the "accrual anomaly" paper Boyle describes.
3. - "The ‘Incomplete Revelation Hypothesis’ and Financial Reporting" - https://publications.aaahq.org/accounting-horizons/article/1...
Paper explains why information can be public yet still not be fully incorporated into prices when extracting it takes effort.
4. - "Limited Arbitrage in Equity Markets" - https://www.hbs.edu/ris/Publication%20Files/Limited%20Arbitr...
Mitchell and Pulvino work on the "limits of arbitrage."
5. "How Pervasive Is Corporate Fraud?" - https://www.chicagobooth.edu/research/rustandy/social-impact...
The paper on about one-third of corporate fraud is detected
6. "There’s never been a better time to commit financial fraud" - https://www.economist.com/business/2026/07/29/theres-never-b...
7. "Firm Data on AI" - https://www.nber.org/system/files/working_papers/w34836/w348...
The Bank of England study on executive AI usage and productivity.
I don't know a single white collar worker who isn't using AI for their job. Not like forced, but like "Oh damn, this bot thing can do a lot of tedious leg work for me".
To think that people won't pay $60-$80/mo to continue using it is wild to me. In a white collar environment it pays for itself in a few hours of use.
If you focus on how much value AI brings to people (mostly in time saved), the bubble hardly looks bubbly.
Did Uber die when a trip across town went from $3 to $13? No. It's giving more rides than ever, 10x more than when it was $3.
I'm sure there will be some losers, but unlike the dot-com boom, the entirety of the population already has all the tools needed to leverage AI.
using how?
summarize email and then generate slop that gets summarized on the other end
Not to forget asking questions and getting wrong answers... Thus possibly causing net loss for company.
$80 per month in a 5,000 people Enterprise is 400k per month which translates to ~5m per year. That is easily in the 8-16% of their total IT budget. I don’t know you, but companies don’t like to increase recurring cost. Is AI good? Yes. Is it clear how it generates enough ROI to justify a material IT cost increase? No. This is the problem now, not how much white collar workers use AI, but what is the productivity narrative of it. Because AI is so generic, there is no clear specialized use case, like when Salesforce invented the CRM for example.
Huh? If that was not a joke, here is the history of CRM:
https://en.wikipedia.org/wiki/Customer_relationship_manageme...
Hmm. Today that'd be creepy.
The thing that's a bit different about AI from a CRM is it translates pretty directly into time savings. You only need each employee to save 1-2 hours of work per month to break even with a $80/seat subscription.
It's true that it can easily save 1-2 hours per month, but it only seems worth it if those hours saved are spent doing something else productive/profitable.
It appears that a lot of the time saved is spent either 1) doing nothing instead, 2) waiting for the AI to finish, or 3) increasing the hours spent elsewhere down/up the chain for reviewing, fixing, auditing.
The obvious benefit of anything offering to save time is that now you have time for "higher order work", but often, we use the time saved to be lazy.
After all, that's the underlying reason we used the time-saving, corner-cutting tool in the first place.
People will pay out of their pockets if an AI can free up a few of their hybrid hours a month. And already it looks like the vast majority of white collar AI spend is from individuals, not companies.
I mean, why would you show your hand by using the company AI? Use your AI and take the credit...
It's probably crystal clear for the people making those decisions.
There is also a very clear specialized use case, which is translation.
Currently, a $20/month openai subscription gives roughly $400 in codex api credits. That doesn’t count what you can get from the web chatbot.
So it better be that api tokens are seriously overpriced. There is value at $20/month. I am not so sure if it’s $400/month or more.
But the question is how much are people willing to pay for AI.
I use AI every day at work, but I only pay $20. That's the most I will ever want to pay. And, so far, it gives me everything I need.
If OpenAI or Anthropic suddenly said "Sorry, the game is up. You'll have to pay $100/month now", I would 100% look into cheaper Chinese solutions.
I suspect the AI subscription (or API) economy is whale economy. You have a small, small percentage of users that are happy to pay whatever it takes - while the vast majority will either use the free tier, and then the next group will pay for the cheapest or next cheapest subscription.
EDIT: And I'll echo what another user wrote here. The VAST majority of office users around the world don't work for tech companies flush in cash. Even adding something like $20 - $100 subscriptions to every user is a serious enough financial obligation that it needs to go through budget planning / boards. Even more so for all the government workers around the world.
Rofl, I already went for cheaper Chinese solutions. I can achieve more than I could when I was giving OpenAI 20 bucks a month, and for a fraction of the price.
I have no idea why people still use ChatGPT or Claude.
Most people don't understand how cheap models like DeepSeek and similar are. I have $5 in an account that I use for months. I can get very complex code for less than 50 cents.
Me too. Sinophobia may be the culprit, too many years watching fox or cnn is my guess?
I love chinese phones, cars, cities, people, and now AI models. There is a beautiful world out there to admire if people is willing to open their eyes.
You can now, as of this week, run locally models with the same performance of a SOTA model of March this year:
https://news.ycombinator.com/item?id=49214008 https://news.ycombinator.com/item?id=49229621
The FAFO day of Anthropic and OpenAI arrived.
You will spend thousands of dollars on hardware to run those at lower quality (quantized) than the benchmarks where they match March SOTA performance.
Source: I have the hardware to run those locally. I would never recommend it to anyone trying to save money. It’s so much cheaper to pay even Anthropic or OpenAI.
Not many people have $10k of specific hardware lying around
I spent $18k in credits last month, I have 3x 20x Anthropic accounts (for Fable to bypass limits) which runs me another $600ish p/m and then a chatgpt pro account another $200 p/m. The value I got out of it is easily 1-200x what I paid.
What are you using it for? I have to assume some kind of agentic workload. Also curious to know when (if ever) you would pivot to Chinese providers to save $$$
Examples like this come up as sub comments very often, but they don’t address the point. Most people will not go beyond $20, forget a 18k.
I use AI daily and pay zero dollars. I get my work done and I stay sharp. A huge mass will always choose the cheapest option.
Absolutely! The opposite end of the preference scale from someone willing to spend 18k a month.
Most people won’t go beyond free. A lot of people will spend $20 and some will go above $50.
You make ~2 million dollars a month?
1 to 200 is a pretty big spread to between losing 18k dollars and making 3.6 milllion - do you have any actual numbers on the value produced from this 18k investment?
Honestly with that kind of usage how do you even know what you’re delivering?
I doubt that. In my tiny town in India, office workers use the 1800 INR (20$) plans. 1800 INR/month is like... 4% of the salary these guys get. And since this is mostly MS-office and windows explorer and chrome based stuff in very small, non-tech companies, it pays for itself in literally a day. If they _could_ pay less they would. After all they pirate MS office. But the cheap chinese plans dont have any distribution, whereas OpenAI seems to have effectively marketed to them via IPL ads and such. No one there even knows about deepseek. It also doesn't help deepseek and such are focused on the coding market. No multimodal, office plugins, etc.
For personal stuff, people there use Meta AI - mostly because it is directly available in whatsapp and these days they are pushing it by adding a dedicated button for it right on the home page. Regardless, they still call it "chatgpt". OpenAI brand is unmatched.
Remember that 20$ is a subsidized rate openai is currently willing to provide. Once that goes down you think these guys would be willing to pay token based billing charges ?
It is not going to go away. Rather they have doubled down and opened 399 INR/mo (4$/mo) plans that as of late have GPT 5.6 Luna.
The reason this works is that you get lesser inference time compute used for queries on these cheaper plans which makes it sustainable and this is enough for the tasks these guys do.
And some local telcos are bundling this subscription as well, so most people just get it for "free". For example, I get Google AI Pro for free with my 350 INR/mo telco plan.
These companies have spent billions of investor dollars and they will need to recoup that cost soon. And then show year over year growth on top of that.
Unless they can massively scale down training and inference cost or implement AGI I don't know what their plan is. Just provide a subsidized plan for the next 10 or 20 years? Their costs are directly proportional to the amount of tokens the LLM produces. How is a monthly subscription plan supposed to account for such costs?
They dont need to scale down anything. AGI is a red herring.
Even Deepseek at its absurd prices is a very healthy business. Regarding their return on capex multiple, their CEO said they make a six-fold profit on their compute capex with 10 month recuperation. Because of this, all of them are spending aggressively on compute. Apart from that, user acquisition and data labelling are the major costs that are preventing net profitability right now. High quality data labelling is said to not have a cost advantage in china etc as well and they pay global market prices for this. I can confirm this is true in india too the model companies I know pay global market rates for high quality data.
By limiting the number of tokens you use per month? per week, per hour? And by limiting the inference time compute dedicated to each turn in each session.
the world economy has shown itself capable of handling decade-scale recouping easily
The main obstacle today in the inference business is the high variability in usefulness/token. This does not need to be solved, but rather only quantified. Innovation is needed to be able to reasonably bound this variance for a reasonable subset of tasks. And we are making progress on this. Naturally though, tasks on the frontier of current capabilities have very high variance. The last couple of years has followed the pattern where tasks no longer on the frontier have reduced variance, but I am not claiming this will continue to be the case generally as the frontier improves.
I am not familiar with chineese model companies as much as I am with US based ones so I don't have much to say beyond that the CEO is incentivced to pump up those numbers.
If this was so simple I don't know why GitHub copilot went to token based billing at my company.
It's much better to make a business case for them after finding this bound right? Currently I can't use copilot for anything serious since I cannot predict how many credits one request is going to consume.
Your point about non frontier tasks using less tokens makes sense. As you said, let's see if it holds up
That's "theoretical profit" - in some imaginary world where the free subscribers would pay the top tier cost.
https://techcrunch.com/2025/03/01/deepseek-claims-theoretica...
Or else?
Their investors will throw the CEO under the bus and hire a new one that will enshittify it enough to make them money.
subsidized versus API pricing sure, but do we have concrete proof of how much it is subsidized versus the breakeven inference costs of those tokens?
I remember the piracy era, and it used to be a joke whether anyone could actually stop it.
Eventually MS dealt with piracy by going after firms, and finally by offering something cheap enough.
Between the legal aspects, and the fact that its MS office, paying Rs 1800 a month is possible.
I’ve been trawling every source I can find to understand what the story on the ground is, and when it comes to productivity it’s a huge mixed bag. The variance in outcomes between independent coders, frontier labs, someone in SV and someone in India is mind-bending right now.
Most firms which talk about their AI plans are not seeing traction, and the AI projects are going to the same place that the ML projects used to go to die.
It’s at the individual level that I am seeing productivity gains, however that isn’t something firms are happy to hear right now.
Sure, because they realized that selling software at any cost because software has extremely low overhead. Selling 100M Windows at 2 USD was easier and my guess is a wash at that price.
However, AI is not selling software, it's selling hardware usage which is not free and has real cost. If Rs 1800 which is 20 USD is not profitable, then they will not offer it.
First I'd like to understand what are these "small" non-tech companies in tiny towns 45k to its employees? The last labor survey had average Indian salary pegged at 21k.
Most comments in the thread say this but don't explain the how. How does it pay for itself "literally" in a day? Through productivity/time savings? How exactly?
I can only assume people are talking about productivity gains. Have their been reports of increased productivity or layoffs in tiny town, small non-tech companies in India that I am unaware of? What is the source here? Without a concrete source this is mostly anecdotal evidence and not backed by data.
Today there is novelty factor of AI. Questions or tasks which required people to really plan out now can be tested through AI quickly. To many people that seems like progress and that AI is paying itself. But most of it is busy work and isn't adding value to the companies. Given that everyone is on the AI hype train they might not be looking at the expenses closely. With time they will start poring it over and cancelling contracts en-masse.
They processed 10 cases in a day rather than 7. In some cases, new work took its place and the company makes a little bit more money, but this is a lagging effect naturally.
PWC et al don't go around doing reports on these tiny companies bro. The only information you get is if you know people in that place and they tell you yourself, which is how I got to know.
Yeah maybe F500 ceos exposed to LLM coding agents are on the hype train to some extent, but these M.S office pirating small companies are not on any hype train lol.
Averaging across the whole of India gives you a useless number. In the richer regions, 21k is less than what a full time food delivery/quick commerce driver makes.
I knew the post screamed "trust me bro" vibes and now I know for sure. Government does Labor force surveys. Here's one for last year. Point me out how the participation is changing through AI paying for itself. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2246009&re...
See the above survey for the figures. But then this is like someone using California to say how well America is doing.
I am very aware of MOSPI. We collaborate in compiling the data for them! BTW there is an MCP server now: https://github.com/nso-india/esankhyiki-mcp.
What do you expect to glean out of PLFS though? Why would broad based labour force participation or other similar data change due to AI paying for itself? What?
If you meant to somehow glean some information out of the wages report, note that the wages survey was last done in 25. The quarterly and monthly bulletins don't contain wage data. Regardless, these are too macro to give any sense of any change at all happening in AI-accelerable knowledge work.
By "PWC report", I meant one where they went to firms and recorded using the same methodology some same notion of AI "paying for itself" or "increasing productivity".
Won't the Chinese providers have to raise their prices as well due to the economics of serving inference at scale?
Buy an M5 max for $4k and you have a portable Deepseek 0731 for life.
But it won't last for life. It will probably last about 4 years so that equates to about 100/mo.
M1 is finishing it's 6th year of life and going strong ...
Have you been trashing the non-replaceabe SSD with constant AI workloads? Do you have any idea how much electricity it uses over it those 4 years?
You cannot wear out an SSD through AI inference alone. LLM weights are only read from the SSD, and read operations do not contribute to SSD wear. SSD wear is primarily caused by write and erase operations.
Your other point is valid, but you're either vastly underestimating how many reads/writes a modern SSD can take or vastly overestimating how much output a typical LLM is capable of.
models are stuck in time. eventually the world moves forward and it is trained on too much obsolete data.
training cannot end for LLMs intrinsically. it's not some fixed cost. it's an ongoing one.
This is why LLMs are never going to be AGI. Humans don’t become obsolete just because they age.
Humans do die eventually though, which is an even more severe form of obsolescense. I don't think they'll be AGI either, but that argument doesn't seem super compelling.
On the other hand you have the old saying "science advances one funeral at a time", indicating death solves at least as many problems as it causes.
(the point being that old, powerful professors have more than once blocked progress in their fields for decades. And only a funeral, eventually, solves the problem ...)
Yes. People keep thinking that Chinese models are free or near free for some reason.
In reality, they're raising prices too. I was looking at Kimi K3 prices on OpenRouter and it's nearly the same as Anthropic and OpenAI.
anthropic and openai are a part of the ai economy, but not the only. the most valuable company in the world is nvidia, a massive player that benefits from the existance of the ow models.
what im trying to say is that we focus a bit too much on the labs, but the biggest part is in the infra that makes everything possible.
But AI use translates directly to time savings (if it works). You only need 1-2 hours of time savings per employee per week to break even on a $100/seat/month subscription.
That's not how people mentally price things though, i.e. "streaming is so much cheaper than the movie theaters!", etc
How do people mentally price things. I'm on the fringe but I'm willing to pay $60 for a AAA 60 hour game that I'm positive I won't finish but am find if I get 6-10 hours out of it.
In the US. With how much they've invested in AI they need the entire planet to pay, and pay lots.
What's the capex expenditure of Magnificent 7 just this year? Close to $1tn?
They need it to be roughly as popular as first world cell phone usage for a 5-7yr ROI.
1. Cell phone usage was never free.
2. Cell phone usage was always localized. You had local telcoms and that was about it. Deutsche Telekom Germany didn't have to compete all the time with NTT Docomo Japan.
And there are probably 20 other economic differences I'm missing.
Everyone uses cell phones. Not everyone has a job where paying this much for these tools would be worthwhile for their employer.
Yeah, they may be over built for sure, I'm just saying the demand is there.
Demand for the internet was there but it still took 10+ years and a big economic crash to get there.
The video quotes sales of $2.5 trillion a year to payoff those investments. What workforce you divide it by is a bit up in the air, but let's use 500 millions, which is basically 100% of the workforce of US + Europe + some change. That gives you around $500 a month per employee (from hedge fund manager to flipping burgers at McDonald's).
And there will be competition. I see zero moat right now. The user specific part of the state of the model sits outside of the control of the model (it is basically my code base, or my prompts, all of which I can transfer to any competitor).
Speaking of user experience: OpenAI has some pretty cool tools, some of which even have decent UX. But I’m amazed that anyone can get real work done with the first party web UI. It is UNBELIEVABLY LAGGY. It makes GitHub seem snappy. It scrolls to a black screen and slowly populates later. It takes tens of seconds to load conversations. On the rare occasions that I’ve tried the fun “ask Pro mode a math question” I think I’ve spent 15 minutes waiting for content to populate in the UI. Oh, and all the obvious features like forking a chat either don’t exist or are somehow hidden. This is on a top-of-the-line client machine.
Maybe other users are doing something differently or are being served a different front end JavaScript blob?
The iOS app is hot garbage as well. The list of old chats will have topics randomly jumping around which makes it really annoying to try to choose the scheduled task that notified you.
I pay the 200 max for Fable daily driving and it’s worth every penny
This is the way it is on Wall Street more each century, and for AI to go public on that exchange it's going to have to go big or go home. Considering the amount of money that has already been spent privately.
There is no alternative pipeline to replenish those reserves, and different people have different ideas about capacity and bottlenecks relative to ambitions and what they are supposed to get for their money.
As has been mentioned in another comment, there is no alternative foundation other than optimism either. The exchange wouln't exist if it weren't originally intended to trade only shares that were all worth holding otherwise. Naturally some worth holding more than others. This is so big it will be necessary to be able to fool way more people with way more money than usual, otherwise those that prevail will have nowhere near their wildest dreams come true.
Naturally AI is never going to fly off the shelf like it could until it starts getting cheaper all the time for huge jumps in performance. Cheap home computers will need to be able to do quite a bit more than they can only do today while connected to a massive AI data center, without ever having been connected to anything outside the home at all. Otherwise AI can not ever be considered "general" any more than computing could be considered personal, until you were no longer reliant on a remote mainframe in a huge out-of-state data center somewhere tied by a thread leading through a squeaky modem over monopolized communication lines. This differential between clunky (clanky?) old data centers' overall computing power relative to the amount held freely within homes & businesses is something that looks like it could be regaining exploitability like never in decades. I know one day Woz jumped right in without needing to be a greedy businessman because there was no possible downside, and he could let just about any dedicated growth leader make as much money as they wanted off his technology. Plus Jobs was no slouch in many ways and Woz never needed to worry with a product that sells itself to begin with, putting Jobs in hog heaven where he could persuasively bring it to the next level almost whenever he wanted like you can do few other ways.
So not just dot-com related hardware & software.
That goes for memory and storage too, people should do the math on how affordable nominal amounts are supposed to be by now. If everything were still normal 16GB of DDR4 would be about $10, at least by 2027, so it hasn't merely doubled in the last year which is the most obvious part, it has skyrocketed to 10x what it would have been. And still rising not falling, so all this is going to have to be reversed for consumers to even afford what they used to be able to do.
For people who didn't want to spend a thing on AI until it's naturally way more capable and constantly getting cheaper like it should be, it's pretty frustrating when it's already been unavoidably costing "indirectly" more than they have been willing to pay if they were getting maximum benefit, when most of them have only gotten more surveillance. One day "indirectly" was just no longer true through no fault of millions of people. And growing as fast as the money stream will allow.
Up until recently AI was way more niche and fewer ordinary consumers were exposed, where now there are millions more aware of its influence and growing fast. But the more aware more ordinary people become, millions more opinions are going to come to the surface and need to be dealt with.
Thar she blows!
One pervasive opinion is that so far it's made by rich people for rich people, and more so than most those who want to get rich quick are joining the bandwagon. Of course what consumers see in the media is only the "band directors" who are already rich to begin with, they set the example because they now want to get richer quicker and this is the vehicle they have chosen to do that. This is not unexpected, after all they always get what they want when it's things you have to be ungodly rich to do.
This makes for company valuations based more on optimism than anything else and once that has eclipsed underlying worth then any unforeseen bottlenecks or financing stumbles become highly magnified.
For one thing there may not be enough momentum to even fully build enough data centers to fulfill some players' plans for recouping such large investments, but at the same time AI's not going to really get good until data centers are not needed at all, which is so ominous there appears to be even more force being applied to encourage people to ignore things like this. Otherwise it could get too shocking.
This disparity in upside just plain instinctively sidelines more ordinary people as it builds, so the number of people who are just going to have to wait for AI and everything associated with it to start getting cheaper all the time becomes a force in itself. It'll be easy to notice without being a finance guru.
Until then gamble at your own risk :\
To put it into perspective if we consider that US salaries are 3 to 5 times those of other "rich" countries then it's already a 60 to 100 dollar subscription for the next richest part of the world.
I have a devils advocate argument: AI spend doesn't need to scale with salary.
If a company was OK with paying 10k a month for an employee in the US because they considered their work more valuable. And the same company is only willing to spend 300 per month on someone in a 3rd world country, because they consider the work to be lower quality.
Wouldn't it be reasonable to consider a situation where the same company would be willing to only spend 100 per month more on the 10k employee to make them a bit faster. At the same time be willing to spend 700 per month on the 300 employee, if it levels out their capabilities and makes them nearly as valuable as the 10k employee?
-- This argument is riddled with assumptions, but I do feel like it's one possible direction that some companies might try to take.
That depends on how much more money they can make with AI or save with AI. It shouldn't be very hard to know that number in a well run business. Even for a one man business.
If they can get productivity increases equivalent to .1% it pays for itself. That's roughly equivalent to firing 1 in 1000 employees to finance the software.
bingo. it helped me sort through 2 CSVs earlier today.
grunt work, really, and nothing some python skills or excel-fu couldn't solve just as quickly.
it also got something wrong, and I had to put it into a different AI to understand why. not really to bad though, a 2 minute, sanity check.
I'd pay no more than, say $3 for the use, maybe 5. If they want to start clocking us $200 a day, or $4000 a license or whatever -- nah.
After using it a bit I’m convinced it’s not, as a whole, a bubble.
Bubbles exist around it. Data center construction may end up overbuilding, much like the dark fiber thing during dot.com, since better chips and models will shrink power and space requirements over time.
But the core tech is the most powerful new thing I have seen since discovering the Internet, at least. Maybe more.
A bubble doesn’t mean a technology is useless. It just means it’s overvalued
In the last 30 years, we’ve had multiple bubbles in the tech industry. Every one of them was backed by real stuff that we still use today. They were still bubbles and a ton of people were hurt when they popped.
The railway bubble and the optical fiber bubble left the world with lots of railways and optical fibers. Once their original owners went bankrupt, the assets themselves were quite useful and changed the world.
The cryptocurrency bubble, however, didn't. Not really. At best you can use it to buy drugs. Some of the GPUs may have been repurposed for AI, but not enough to make a dent in AI, and the AI reesarchers who kicked off the AI bubble were having to buy their GPUs for high prices to compete with the miners, not low prices after they went bust, so we can't even say miners helped kick it off.
Many of those assets were useless - like railways from nowhere to nowhere. Only some ended up being used.
And many innocent people ended up being hurt in the process.
Bubbles are not good. They are harmful. The ability yo recoup some of the losses in the span of next 15 years does not make the bubbles something positive.
Even railways from nowhere to nowhere often ended up being useful when the ends were extended to somewhere or when somewhere grew to encompass nowhere.
For example the Berlin Ringbahn was once some billionaire's toy but now it serves very useful traffic, as Berlin grew outwards to where the Ringbahn is.
It is not "often" it is "sometimes" and it is also less than a participation trophy. It does not make the bubble not harmful or even less harmful. It just means that sometimes you can repurpose an expensively build junk.
bubbles are bad for society too; families get thrown out on the street, big players/individuals buy all the cheap/discounted assets furthering economy divisions, and depending on the size civil unrest etc....
I know. I'm saying that overall I don't believe it's that overvalued.
Specific things may be overvalued. I suspect data center real estate is a bubble, for instance.
This new technology more powerful than the internet can't answer the simplest questions straight. If this is the future then I don't want it. LLMs are pretty great at generating code and recognizing patterns and that's about it.
That’s not the argument though. The argument is that conversion rates as per OpenAI are about 5% and to serve the other 95% you need to build significant infrastructure that costs hundreds of billions annually. So for every white collar worker that you know and uses it and pays for it there are 20 free loaders. And mind you, 5% conversion rate is extremely high already as a number per se. It won't magically reach 20% in a couple of years. Chances are it will fall particularly for the enterprise segment due to the advent of the Chinese models.
The reason uber didn't loose that much business is because it replaced the established businesses and left most consumers with little alternative.
That's not the case with AI - unless the frontier labs achieve AGI or some sort of super intelligence that lets them create infinite economic value (at which point, any further discussion is pointless for obvious reasons), the average worker can do most of their tasks with 5% of the api costs using an open source model from China and achieve the same results.
The established businesses were awful. There was not a thriving and beloved taxi service in every city pre-Uber. Taxis were bad and very expensive.
Also did you forget that Lyft exists and normal taxis are still available? Taxi services had to improve their business when the ride sharing companies entered the market because they previously relied on their own grip on on-demand transportation to keep prices high and service poor.
Imagine suddenly a new taxi business was able to come into your city and provided taxi rides for 5 cents on the dollar. What would happen?
Edit: now add on top of that that even if uber tried to undercut them again by running losses as they did in the past, that new taxi business would not loose any money if the number of their rides went down because their operating costs simply scale with the number of rides - while uber was loosing money on every ride.
I literally lived through this situation. It was not 5 cents on the dollar. The taxis continued to exist then and continue to exist now.
Are you also going to tell me that everyone will be switching to their favorite Linux desktop distribution over Mac & Win because it's "free"?
We're talking about corporations here. Of course they use Linux, and of course they'll provide their employees access to the LLMs that cost half the price and are 90% as good.
I can't tell if you're being serious or sarcastic.
In case it's the former, nobody in the corporate world is using Linux desktops.
And let me know when your parents start using the Chinese LLMs on their own.
This is the wrong comparison. An LLM provider isn't like a desktop OS; it's like a server OS. And everybody is using Linux for their servers.
That already happened. To be fair, they're Chinese.
How is the effort switching from an operating system containing all your data and apps, which might not be compatible with the new one, comparable to just switching an api key?
A bit of napkin math.
How many are the white collar workers? 300 mln? even with $100 subscription each, the monthly revenue for your addressable market rises to $30 bln. If we increase the figure to half of humanity and they are 3 bn, you multiply it upto 10 and $300 bn monthly, $3.5 tn annually.
Now start cutting it down keeping in mind that not everyone is an office worker, does not need $100 plan, has access to cheaper models, could get ai through someone elses subscription or local models. I'm not sure how down it will get, but it can get down a lot and will have to be split among few players. At least for me it does not seem unrealistic that the revenue won't be enough for everyone.
The real question is not whether there will be a market for AI, it is obvious that there will be one (as it was obvious the Internet was a gigantic thing in the dotcom boom). It is rather whether the value will be in the models (and if so in which?) or the infrastructure or somewhere else.
Interesting thoughts from Dwarkesh Patel on the future of models [1]. In summary, no moat if a client can switch to a competitor from a dropdown when all models are similar-ish (so models are commodities), but to really compete with humans, AI needs to start "learning on the job", i.e. accumulating experience like an employee (as opposed to the training / inference approach where the model isn't learning after it is released). If your model does that, then you have vendor lock-in as you can't replace an experienced model with a competitor. Then the value is in the model.
As in the 90s, this could go in all sort of different directions, hard to make predictions. And remember the 90s: the value was in the browser (you had to purchase it, eg Netscape), then the browser came for free with the OS (no value in the browser), then the browser was the most important thing in the world as it was control over the default search engine, everyone thought portals was the most important thing to control, so ISPs had their own applications accessing the internet from their own portal, etc - none of those people were stupid.
[1] https://www.youtube.com/watch?v=iewm45atodE
The moat is in sales, the model and its quality is largely irrelevant. Gross margins will be high enough for moats to not matter that much.
I work for a large company and for any LLM I use professionally, whether programmatically or via a UI, I get to pick from a dropdown from Gemini, Claude, OpenAI and the open weight ones. We ain't locked in.
Yeah, and that is completely fine business-wise for all the businesses in that dropdown. Averaged over the whole TAM, they will all make a lot of money. The reason is that the total market becomes _bigger_ with every new model (upto a reasonable point of course, but for 5 models it is obviously true), they don't end up competing for the same slices of the pie.
The core business for each of the 5 will not be companies like yours, but rather few hundred to few thousand companies each who exclusively use their model not for technical reasons, but because they were wined and dined or due to bundling with another service. The other hundreds of thousands of customers are all bonus. This is true for most categories of SaaS not just LLM inference.
If the vendor of a core piece of software that you are already locked into tells you they are raising prices and adding AI inference features, you are going to cancel whatever AI plan you have and use theirs. Or pay the amount anyways (which is even better business-wise for them).
Eventually, semi-random differences that arise in their distribution may end up informing model capabilities as well, and _that_ is when true model-level differentiation may happen. But this is not necessary and if it comes true would just be a bonus.
Not hosting in China will be a moat. Tech (software) never really experienced it, but the rest of Western industry all have stories of getting burned by China and their policy of ignoring IP rights.
I agree that it's working good enough and that it's here to stay: mostly everybody around me uses AI too but...
The competition from open-weights models is already putting pressure on how much this can be billed. That's today, with stuff like OpenAI announcing 80% cuts on terra and luna. Tomorrow, in addition to the pressure from open-weights models, there's going to be pressure from chinese hardware: Tencent, Alibaba, etc. are all already working on their own AI chips.
What good would, say, a SOTA open-weights model running on chinese chips (chinese chips located in China or elsewhere btw) do to, say, Meta or Oracle's insane AI investements?
And it doesn't even have to be chinese chips: what good AMD AI chips doing inference by having weights etched on silicon will do to OpenAI and Anthropic?
All the people, especially here on HN, who were explaining six months ago that OpenAI and Anthropic models where so good for coding that coding was solved once and for all can now run better open-weights models than the proprietary ones from six months ago. For a fraction of the price.
If people, instead of paying $20 per month pay $80 per month for their AI subscriptions, what makes you believe that money is going to go to one of these companies participating in this $2 trillion debt?
And really, in which new domain have we seen the first movers being able to keep a lifetime grip on the market without getting their arse handed to them by the competition? Things moves quickly today: I'm not sure OpenAI and Anthropic are the Ford equivalent for AI.
For all I know OpenAI is "La Mancelle":
https://fr.wikipedia.org/wiki/La_Mancelle
And Anthropic is Panhard.
That AI is here to stay is a given. That people may be willing to pay more is likely (even though IMO they won't pay 4x more for something only marginally better). But that the current players are going to stay on top even though the competition closed the gaps to weeks?
I don't buy it.
If you need fewer workers due to the increased productivity, then the number of white collar workers (the TAM) won't be as large as anticipated.
That’s never really how efficiency gains tend to work except in places where entire industries ceased to exist (printers, weavers, clothes washers, etc).
How sure are we that isn't the case here?
People may pay less than they are willing to pay as competition drives the price down. I paid $20/mo a while which I was happy with but I found I could get similar for free so pay $0 now.
Mods, can you explain what happened here?
This story had 49 points and 81 comments in about four hours and was on the front page. Then it seems to have disappeared from the top stories feed completely.
Anthropic has 80%+ margins on inference.
Google has 30%+ margins on compute.
Both parties have discovered a literal money printer. The payback period is <2 years.
At those unit economics, anyone not borrowing aggressively here to create more money printers is a moron.
That's like saying my delivery company is profitable because with current gas prices, my margin is 80%. Yea, what about all money you spent to get there? You still profitable then?
Yes, actually. Anthropic has gross margins of 40%+, hit $75B ARR last month, and (as of this quarter) is profitable.
As Ed Zitron pointed out, that "quarter" of profitability is EBITDA profitability and comes with plenty of creative accounting. When they do it for a year, we can say "They have found profitability"
Why are you citing old news? Zitron wrote that months ago when Anthropic was reported to have ~30B ARR. Anthropic now has more than double that, at even higher margins.
There's no doubt at this point that Anthropic is profitable.
Because again, we haven't seen further reports. As always, we are debating financials on a company that doesn't have to disclose them regularly.
If they were profitable, they'd release real numbers and announce concrete plans for an IPO. They are very obviously not profitable; they just announced a $65B Series H two months ago.
They're almost literally setting money on fire.
This[1][2] Ed Zitron, or another one?
Because I don't know how you can be so totally and completely wrong for so many years, and still have people lend you credibility. But I definitely do understand how you can rage farm subscription dollars from suckers for years.
[1]https://www.wheresyoured.at/bubble-trouble/
[2]https://www.wheresyoured.at/to-serve-altman/
Yea what do guys like Damodaran know about finance and valuation.
This board is only good to understand the consensus dipshit opinion.
You have to assume that they aren’t being too creative with their accounting, which Anthropic and Open AI most certainly are.
Until they are public and they release full financials using standard accounting practices, I would take what they say with a grain of salt.
SpaceX also painted a pretty rosey picture of their financials, but when they filed for IPO it became clear they were doing some real cherry picking and in actuality they were losing a ton of money.
If they really were doing so well they would file publicly and be rushing to IPO.
I guarantee they are losing billions per quarter. You can say they make money on inference, but that is irrelevant. That doesn't include all of their overhead and indirect costs let alone model training.
Why do they both need to keep borrowing so much money if the margins are so fat?
Demand for compute is growing 10x year-over-year, faster than inference profits can fund expansion. At 80% margin the payback period is 1.25 years. Cash flow can't keep pace with that scale.
If you read the interviews where Dario said this closely, you'll find that he's talking in hypothethicals.
We will have to wait for the S-1 to get audited figures, until then, only insiders have any idea (and only the accountants can be sure).
That point is irrelevant because the number comes from SemiAnalysis, not Dario's interview:
https://www.tradingkey.com/analysis/stocks/us-stocks/2620181...
Given Anthropic is charging $50 per million output tokens on Fable, those high margins are very believable.
Look, that's an estimate (by a presumably biased observer). I don't really have a horse in this race, new large tech companies are fine by me, even if only to reduce the power of the current ones.
However, I find these numbers incredibly hard to believe, and most likely deceptive, given that they recently (like March) started making enterprises pay API rates, so even if they were profitable (and if they paid from Fable training from this post cost money) I would be sceptical that this will continue, given all the competition in this space.
tl;dr let's all wait for the S1 (it will presumably be soon, unless SpaceX declines get them to postpone).
But pay for who? $50 for a cheaper competitor? $20 for a chinese one? Or just buy hardware and host/co-host for privacy? And to balance the insane spending even your $60-80 example would be hardly enough from that white collar audience.
Your base premise is flawed. I know plenty of white collar workers who do not use AI for their job.
At my $day_job we are forced to use a shared virtual machine for all work activities. The Windows instance has 64gb of RAM, 8 virtual cores of old EPYC CPU and no GPU acceleration. It is also shared between 6 users.
They will not spend extra 20$ per user per month to add compute and increase productivity.
The question is not whether LLMs have value; the question is who will earn how much money off them. I don't see any scenario where the companies that put hundreds of billions into their models make that money back.
Pays for itself how? Customers don't care if a business saved time on bullshit paperwork. It needs to actually convert to increased competetiveness which leads to increased profit to pay for itself. But if everyone is doing it anyway businesses will quickly realise they're just paying an AI tax and look for ways to eliminate it. Maybe by just not doing the bullshit they got AI to do.
I avoid subscriptions like the plague and I was getting enough value from Claude that I was paying for the Max plan for a couple months. I don't know how much Anthropic needs to make to be sustainable but I think it's worth north of $50 or $100 per month for many knowledge workers.
I highly doubt anything more than a small single-digit percent of users are willing to pay $50-100 for a Claude subscription themselves. Just like everything else (if your company isn’t paying for a seat) 90% of people will just use the free tier, 7-8% will use the first paid tier, 2-3% use some kind of power user plan. It could end up being a whale-driven business model in the end.
I don’t think end users will purchases subscriptions out of their own pocket for work en masse.
If the value is there, companies will.
I'm fine paying $200 per month. Companies seem to be fine at $2000 per month but I would balk at that. Companies are already balking at $20,000 per month.
So, we KNOW what the numbers are. If corporations are only willing to pay $2000 per month per developer for all the developers in the US, the amount of VC investment already exceeds what they can get back from 10 years of revenue.
Why $60/mo when you can pay $5? Why $5 when you can pay $1. Unless you happen to be in one of the industries OpenAI focuses on you would always just use a Chinese model. Even now, but especially in the next 1-5 years.
Uber has a fairly large moat: the regulatory mess they’ve waded through and the collection of drivers willing to drive for them. Sure, it’s probably easier to start a competitor now than 20 years ago, but there’s still a substantial network effect — a transportation service needs a lot of cars to provide good service, and they need a lot of users to keep those cars busy.
An AI service like chatgpt.com or claude.ai costs literally billions of dollars to develop, except that several Chinese companies are currently happy to spend that money and release the models and weights for free. And even a small company can rent the datacenter capacity they need, at least at moderate scale, and they only need enough scale to reliably fill up the batches on a handful of servers. There is no obvious network effect — a tiny Uber with a total of two cars in a city is useless, but a tiny AI provider that only services 2000 users is just fine as long as someone else provides the models weights.
And the costs to run these services keep coming down. The numbers people are reporting for unquantized DeepSeek v4 Flash 0731 on 2x RTX 6000 Pro [0] using open weight models and open source inference stacks are astounding, and one single copy of this could probably cover the entire inference usage of a decent sized business doing the kind of desk work that OpenAI and Anthropic could need to monetize to cover their expenses. (Using bespoke models has value but may reduce available usage of a single inference machine. Quantizing can make everything cheaper and is often good enough.)
This is not to say that these companies won’t make it on the strength of their enterprise offerings (never underestimate the willingness of enterprises to overpay for services they’re already using or by becoming the next Googles and finding other revenue sources beyond just inference.
[0] I admit to some skepticism that the pile of sketchy, obviously AI-generated patches to SGLang that allegedly achieve this are doing what the say they’re doing, but the point stands.
There were so many comments[1][2][3] at the time saying that Uber has no moat, there’s no switching cost, the only value is the $3 rides, etc, etc.
I suspect it’ll play out the same.
[1] https://news.ycombinator.com/item?id=8987827
[2] https://news.ycombinator.com/item?id=6137542
[3] https://news.ycombinator.com/item?id=9984619
Agreed, ChatGPT has more than 900 million weekly active users. They're not losing the top spot anytime soon.
how many paying ones?
If OpenAI wants to generate revenue by charging its users for services, this is a big deal. If, on the other hand, OpenAI wants to become like Meta or TikTok, then they don’t actually need to charge these users.
But latter model will not be easy. TikTok does very well, but their revenues are not enough to justify OpenAI’s expenses. Meta is very, very good at capturing value from users’ attention, but they are not going to give this market up without a fight. In general, I’m somewhat suspicious that the total addressable market for attention is not infinitely scalable. At the end of the day, the money comes from marketing budgets, and those marketing budgets scale with the amount of stuff being sold and, much less strongly, with the efficacy of the marketing.
There’s a theory that there is money to be made having at AI-like system that will actually buy things for the end user. Amazon tried very hard to make this work (in the Alexa division) and basically gave up. I’m not finding my breath for OpenAI, or for anyone else, to pull this off any time soon.
And I know very few white collar workers who are, beyond "I guess Google Search is AI now". Each their own bubble, I guess.
AI has some value in some scenarios. Most models still fail badly enough often enough that I personally would not pay $80 for it. I would make use of it if it were free, or $10 / month, but it would have to significantly improve if I were to ask my employer to pay $150 / month for it - let alone $1500 / month.
The evidence is obvious: everyone and their dog has access to highly subsided AI agents right now. Literally every single company is pushing its employees to trial AI - some even forcing AI use. So why aren't we seeing a giant white-collar productivity boom yet?
The idea that every single white-collar worker is going to get a $800 / month AI subscription is ludicrous. The big question is, can the AI companies survive off everyone getting a $5 / month subscription, with a handful of workers getting a $500 / month one?
The problem is the value you speak doesn't seem to be materializing. I barely know anyone who sings the praises of saving time with AI, but I know plenty of people who complain about peers doing a shoddy job because they used AI to do their work, or having to clean up someone else's AI generated slop.
Uber is the wrong example here, as it has destroyed $27B of value since inception.
Or maybe that's exactly the right example...
I'm willing to pay >100$ a month to not have to use AI at all. It's trash and will never create anything of value.
Would Uber have died if a competitor kept offering $3 rides? Yes.
The question isn't whether AI is worth it, but whether it becomes a commodity.
Why are there an endless flood of comments like this on HN and other social media insisting that because there’s such high demand, it’s IMPOSSIBLE that these tech companies are overvalued and over leveraging debt.
Reality is way more nuanced. Something can have high demand AND be on the precipice of financial catastrophe. See: Spirit Airlines.
It doesn't cost you $80 a month, it costs you your creativity first and then your reputation
Is Sam Altman even allowed to rent a car with his level of debt?
he just rings up his microsoft daddy
As long as the money printer is running...all Quiet on the Western Front
It's a near guarantee that when things shall heat up and the stock market shall crash and people are going to say that "this time it's worse than 1929", we'll hear the brrrrrrrr sound of the money counterfeiters, very likely synchronizing their counterfeiting operations worldwide (US / EU at least) to print ever more money.
Is this an AI generated video? The guy didn't blink an eye or moved the head more than one inch in 30 minutes!
Hah instantly knew it was a Patrick Boyle video. He's a well-known expert in rap music that also does finance on the side.
True. He mentions AI vids and rap here https://youtu.be/Ak4on5uTaTg?t=161
No. There's a blink at 11:04.
I think he just cuts a lot of takes together. It's his style.
We now have several voices saying the same:
"Aswath Damodaran: Big Tech Has No Idea How AI Pays Off" - https://news.ycombinator.com/item?id=49229981
Wall Street makes money from issuing debt, trading debt, and IPOs. They need to get the last two mega IPOs off before the curtain call on this era.
But they can only partially influence the market, they can’t control it. The question is if the market will let them get these IPOs off or if the jig is up.
Here’s an exercise for anyone curious: pull up the median stock in the S&P 500. Look at its P/E. Take 15 minutes and look through the company’s financials and figure out what you think about the quality of its earnings.
Then decide if that P/E is appropriate.
The bubble isn’t in just AI, the bubble is everywhere, and AI is its largest manifestation.
its almost as if we over-financialized our economy...
Capitalism works with debt for growth. Strange HN doesn't understand this
And capitalism also works with crashes taking unprofitable companies down. And frauds, eventually, gets companies valuation like Enron and FTX to zero. And people go to jail.
Now I'm not saying OpenAI or Anthropic are frauds. What I'm saying is that, eventually, things revert to what is just.
The late 90s SV tech-bros behind pets.com or webvan for example faked it for 18 months to 36 months or so. At some point when the expenses outnumber the revenues, capitalism does its thing.
Yes and since there is no fraud there is no reason for this panic they know what they are doing
hahahahaahahahhahahhaaa
Delulu much? My God some people
A number of famous business leaders have admitted they had no idea what they were doing in their memoirs. A few have even suggested that was an advantage.
Everyone running these AI firms is being forced to make educated guesses about the future of the technology and the industry. It's unlikely they're all making accurate predictions.
Subsets of HN do understand this.
Subsets of HN just want to know how long before they need to withdraw their savings, or sell their position.
I would say growth is even expedited with high allowances for borrowing/leverage, but at the cost of increasing systemic risk.
Obviously the people who will suffer from this are the taxpayers, who don't get a say in the matter, so no one is going to stop and consider whether it's a bad idea to continue issuing more and more debt.
The creativity of the financial industry is quite insane. Combined with the scale-at-all-cost strategy in AI companies, this is a ticking time bomb, no doubt.
I think it is easy to forget that a revolutionary technology does not automatically make a viable business model. I think we are yet to see the real winners in this game.
I agree with you. Look at railroads, cars, computers, and other technologies that really did end up being world-changing. How many of the original companies are still around?
Sometimes I feel like OpenAI will end up like Fairchild Semiconductor: an early innovator that quickly became an historical footnote.
There's a kind of inconsistency in that he ends the video talking about overpriced stocks and how it's hard to short them but the title is about the $1.75tn in debt which seems to mostly be from professional investors, apparently:
who probably know what they are doing and read the footnotes.
My guess is that the lending is actually ok because there's a lot of real demand for compute. $1.75 tn is about 1.4% of global GDP which doesn't seem that silly in the AI boom.