Seeing that R&D costs are the lion's share, I wonder if we are at a point where the focus can shift to improving the cost of inference.
Unless we are genuinely pushing to find AGI, at which point nothing matters, LLMs in their current form don't replace knowledge workers but are an effective force multiplier. How good is enough?
For instance, I pay about $1-2 a month for DeepSeek. It's not as sophisticated as Claude, but it still doubles my productivity as a SWE.
If Fable comes out and demands 50x the price of DeepSeek in order for Anthropic to make a profit on it, how much more productive would I be compared to my personal experience + DeepSeek? 3x? 50x?
Is it cost effective for a business to hire someone without SWE experience + Fable verses hiring someone with SWE experience and DeepSeek? When does R&D hit diminishing returns?
> I wonder if we are at a point where the focus can shift to improving the cost of inference.
There's always working on improving the cost of inference, but I don't think this is an area of R&D that will slow down. The reason is:
1. A better competitor model risks eating away at how much they can charge for inference (i.e. revenue)
2. Whoever unlocks AGI will unlock even more growth
3. Even when you unlock AGI, you'll want to throw gobs of money at it to improve itself and all sorts of things.
> If Fable comes out and demands 50x the price of DeepSeek in order for Anthropic to make a profit on it, how much more productive would I be compared to my personal experience + DeepSeek? 3x? 50x?
You're pricing it wrong and looking at it wrong. First, the per token price doesn't consider that a smarter model can end up using fewer tokens overall to achieve a result. Secondly, if the difference is between failing to accomplish the task and accomplishing the task, suddenly that 50x can seem like a bargain.
> Is it cost effective for a business to hire someone without SWE experience + Fable verses hiring someone with SWE experience and DeepSeek? When does R&D hit diminishing returns?
At this time, someone without SWE experience + <name AI model> vs someone good with SWE experience and <name another AI model> is a no-brainer. The AI model is an accelerant but the "no SWE experience" will be accelerated into a wall. Now maybe that doesn't matter for prototyping and certain other things, but anything in production the lack of experience will hurt them with things they won't even know about or even know how to look for it (e.g. slow, insecure, etc).
> You're pricing it wrong and looking at it wrong. First, the per token price doesn't consider that a smarter model can end up using fewer tokens overall to achieve a result. Secondly, if the difference is between failing to accomplish the task and accomplishing the task, suddenly that 50x can seem like a bargain.
Perhaps, yeah. Napkin math; if we go off of the assertion that DeepSeek makes me twice as productive for an extra $24 per year.
The same number of tokens with Fable would be around 50k/y. Say Fable uses half as many, that's still 25k
The question is then; How does Fable make an engineer more productive than DeepSeek?
- Prompt adherence?
- Inference speed?
- Can engineers delegate more work to unsupervised autonomous workflows?
Are unsupervised workflows suitable for production applications or does the review overhead erode the productivity benefits?
Perhaps a company could hire a mid level engineer + Fable to achieve senior engineer levels of productivity for the same price as a senior + deepseek.
For clarity, inference is typically a COGS and therefore hits Gross Margin vs model training which would typically be in OpEx (where R&D lives) and would hit operating margin.
Even if you discount superhuman AI (which I would emphasize that frontier researchers do not discount and expect to see soon) think it’s still hard to have enough confidence that the ground is solid. Someone in 2024 trying to go down this route would have invested a lot of now-pointless effort into prompt engineering.
So... 50% operational costs and about $100 spent on sales for each paying customer.
If they manage to keep those customers for several years without more sales, that bit looks like a normal "high-touch" business.
They shouldn't look like a "high-touch" business, but their unitary numbers look way better than I expected. They just need to grow some 10 times to star making a profit... Maybe 100 to cover the opportunity cost of their capital.
It's just a matter of finding 5 billion people willing to pay US prices :)
I'd be cool with that. YouTube premium is one of the best value subscriptions I have. Steering people toward paying instead of ads-by-default is a net good imo
They won't try to. ChatGPT is already starting with ads, which is potentially far more profitable (as evidenced by the fact that the most profitable company of all time makes 90%+ of their revenue through ads).
>as evidenced by the fact that the most profitable company of all time makes 90%+ of their revenue through ads
the biggest reason for this is that the digital ad market is a duopoly (charitably a triopoly if you count Amazon in), if all of the LLM companies start to go into ads that's going to be a much more competitive market for ad buyers. It's not going to be so straight forward when both customers and merchants have ten different places to go.
Also not to forget that ChatGPT has zero moat, unlike social Facebook and Google.
I'm guessing that might be so in certain professions, but I would expect the employer to pay for that. For the rest of us, it seems unlikely. At least for me, I don't have a need of a device to generate text for me. And I bet most people are are in the same boat as me.
These numbers seem insufficiently detailed to really evaluate anything. They’re had $13bn in gross revenue in 2025, and they cost of that revenue was $7.5bn. Both are growing fast (we assume) and the ratio ought to stay roughly constant.
But: how are they calculating the cost of revenue? Do they have rapidly depreciating assets that are also needed to produce that revenue? (Starlink has this issue.) Will their cost per arithmetic operation for inference rise or fall? (Anthropic is paying xAI an absolutely insane amount to lease GPUs. They must be betting that they will not need to repeat that.) Is a large portion of the cost allocated to R&D actually being used to support their revenue?
I certainly believe that the cost of inference can be plenty low for them to make a profit, but a more granular breakdown would make it easier to evaluate.
I feel like I have a different $20 plan than everyone else. I have no problem hitting my 5 hour and weekly limits. Don’t get me wrong, it’s a great deal compared to API pricing, but it’s a far cry from “unlimited”.
I get about 20 minutes of work from my 5h limit with the $20 plan. It wouldn't bother me as much if codex would continue after the token bucket refills instead of waiting for me to show up and tell it to continue. I don't jump to the $100 plan because I would be in the exact same situation.
Harness matters in this. Using the Codex sub with Hermes eats tokens like nothing. Using it with Pi is much less but you don’t get the long term memory. When you were able to use the Claude subscription with Pi, I barely hit the 5hr limit. When they stopped allowing that, CC harness just chews thru tokens.
Interesting. I'm mostly using Claude, so perhaps I'm not nearing the limits, but I do use Codex (for coding and reviews occasionally) and use chatgpt for second opinion many times, including "pro" research. Never got to my limits. But again, not my main go to tool.
> Interesting. I'm mostly using Claude, so perhaps I'm not nearing the limits, but I do use Codex (for coding and reviews occasionally) and use chatgpt for second opinion many times, including "pro" research. Never got to my limits. But again, not my main go to tool.
Not to mention they will need to research how to make their models faster and cheaper to run in order to fit some margin within what people are actually willing to pay.
It's more like once you figure out how to make a really good lamp then producing lots of lamps will be profitable. But the lamps are currently suboptimal so we'll be in the red until that time.
It's more like you have a business making engines, each generation of engine has eventually turned out to be profitable over its lifespan, but each generation has an exponentially increasing R&D cost and your customers will switch from the old engines to a competitor if they don't like the newest generation.
You're stuck racing against your competitors with the distinct possibility that your R&D costs will outgrow the market demand, and you can't stop because otherwise your customers will stop investing in your dead end tech and switch.
OpenAI won't be able to cut R&D spend and collect rent on their existing models as long as the Chinese models keep up the pace of being ~6 months behind them for a fraction of the price.
And then someone will come up with lamp pro max and you’ll be out of business. You realize why R&D exists in tech companies even though it’s a cost center right?
Watch them flare out like a star… but there is lots of questions re the the return on RnD. Is it worth spending another order of magnitude for only marginal frontier gains?
People keep overlooking the fact that costs for these providers scale along with customer acquisition. Most startups don't have that linear expense. Also, training costs are accelerating to get new models out faster. One doesn't simply "get rid of R&D" costs as a comment upstream mentioned. I can't actually imagine R&D goes down anytime soon unless you're willing to play third fiddle.
Unless these frontier providers feel some type of squeeze or constraint the Chinese are well positioned to leave the US bag holders of an NVidia bound system. And if anyone has to wonder how one provider for a critical piece of infrastructure will go, well...
Even if they keep the R&D costs, more efficient inference and 0 Marketing spend also gets you there. Inference is honestly super inefficient at this point, we can do far better than GPUs, push utilisation up, build more efficient datacentres.
My takeaway from this is that it's incredibly validating as a business model. Inference is _highly_ profitable. Of course, like any company that has ever tried to grow at breakneck pace, you run at a loss until you "win."
Isn't this what all of the big companies that spend a lot on R&D and engineers promise?
And then the reality turns out not to be the case - you have to continuously spend on R&D to avoid getting your lunch eaten by someone else.
This isn't a social media network with lockin either. People can and will just switch to whatever whenever they feel like it. Maybe it becomes a defacto standard like google but if someone is much better than you, well...
> My takeaway from this is that it's incredibly validating as a business model. Inference is _highly_ profitable.
The problem is you can't just separate training costs from inference costs. If OpenAI just didn't train a new model for the next five years, sure, they'd do OK. Assuming all those dirt cheap Chinese models nipping at their heels don't make up the gap while OpenAI is resting on their laurels.
Without being a frontier model (read: continuous, incredibly expensive training), they effectively don't have much to sell. So inference and training costs are intertwined to some extent.
Yes, it is like a new era - the startups have huge direct revenue on real products instead of "users" which yet to be monetized.
And the network effect which ruled for the last 20 years seems to have relaxed its death grip just a bit (of course it is still there as having more customers using your tools and models provides more training data, etc., yet the current network effect doesn't seem to have that high exponential value like before)
I wonder how effective the marketing is (not much it seems).
I was watching a World Cup match last week and one of the TV ads during half time was something to the tune of ChatGPT being used by kids to improve their street soccer skills. This was Brazilian TV. Anyone even remotely familiar with Brazil would find this ad deeply, thoroughly out of touch. I can't think of a worse chatbot pitch than that.
6bn seems excessive but despite GPT 5.5 arguably being better than Claude I don't see a lot of adoption of Codex yet.
Some of my coworkers even use Sonnet (the default in Claude Code for the 20 USD subscription) and see no reason to change even though that model is definitely "outdated" compared to current SOTA.
Marketing might help at some workplaces, presumably that are dedicated to Microsoft, for example our network blocks Claude (and DeepSeek) and is slowly rolling out Codex team by team. They should encourage Amazon/AWS to market for them.
I'm really curious about something: how far will you go to support AI? Clearly they'll need to monetize things further, would you still use [whatever AI you are paying for] if the price was doubled? Tripled? Where would you stop and would you stop using AI altogether or would you look at competitors?
I will do nothing to “support” AI. Either it has utility or it doesn’t. I feel no loyalty or duty to help make it work if it doesn’t.
Anyway: Zero, as of right now.
I fully expect to be able to run useful LLMs on a machine I can justify buying for other reasons. I already can on the secondhand kit I own, and I don’t expect the cost-benefit analysis of local LLMs to ever really get worse.
If I ever need to pay for it, it will likely be to shift some of the capacity into the cloud for either business or pragmatic personal reasons (so I can just carry an iPad etc.)
I fully intend my expenditure to be negligible. Because once one realises that outspending others is impossible, only spending minimisation makes sense.
I foresee it potentially making sense for me to move some mature tools off a local LLM to openrouter, maybe. But probably to the same or similar models.
I don’t and won’t support AI. For a while I paid 200€ a month and would have been happy to pay up to maybe 600€. However I don’t want to participate anymore in using such an anti-human technology and industry
I don’t think there’s a single person out there that will ‘support’ AI
Maybe it’s just your phrasing but people will only pay for what works, no one is loony enough to support a trillion dollar industry out of the kindness of their heart or spirit of innovation
I make an important distinction between cloud services and local AI. My lifetime spending on cloud AI is probably less than $500, and I don't intend to spend any more. But I've already dropped $2.5k on new hardware for local inference, and could easily see myself spending more in the future. In fact, I'm regularly browsing for deals. I would also be open to paying for local models, if there was a way to make that compatible with fully open models.
AI is so important, I want to have it under my control. Even if I have to pay a penalty in terms of capabilities.
It depends on how much time it saves me and how much I make per hour generally, right?
If AI allows me to cut my time to do something in half on average or allows me to do 2x more it would be worth it to pay up to what my monthly income was before assuming my income scaled with my output.
I've spent a grand total of $25 on AI ever, so apparently my answer is $25. But I'm not a big time software dev like the rest of you.
When I bought my last GPU, running AI models locally was a consideration though not the only one, and I have it set up but haven't used it much yet. I mostly use the free tiers of ChatGPT or Google to write the occasional script for me. I guess they're going to have to inject a truly unfathomable number of ads to get their money's worth.
I have a feeling my experience is closer to an average persons' than a dev, but it doesn't seem like they'll be able to monetize just from devs even if each one is spending thousands a month.
I'm not a coder but now work way faster than the coder I pay, stuff breaks but it's tenable and it's easier to get things to completion as the harnesses get better.
Don't give up just keep trying you can truly build personally life changing things. Don't look at it purely from a how do I sell this lense, just empower yourself with these tools while the getting is good
For work, it depends, but if I have to spend more than a few hundreds bucks probably I'll start looking for alternatives (local models, Chinese providers, ecc)
PS: I'm in Italy, I guess in several parts of the world these figures are even smaller.
Stretching the analogy, something that gets you from point A to point B for a fraction of the price without the same level of comfort is totally fine for me. For some of my tasks, that means using local models. For others it might mean a frontier-last-year kind of model. That's totally acceptable most of the time. For anything else I guess it's like renting a truck to move; just get the right vehicle as needed and pay the premium.
A $50k car used 1,000 miles per month probably costs close to a thousand per month, assuming 200k miles of life. I imagine this is not unusual in the US.
Agreed. For personal use it's already easily worth $100 a month (to me personally). More probably. For work, it's entirely based on its financial impact for a given role, and for some people/companies it will be worth the cost even at $X thousand per month per seat.
This title is not how I'd actually interpret the results.
Glad to see more sane takes in the comments. All these articles on their current financials are missing the point.
OpenAI is doing pretty well.
Capital expenditure is required to deliver on 1) better models 2) better infra and 3) better products. Insane CapEx is required to do all the above + compete with Google, Meta, Microsoft, Apple, Anthropic, etc. etc. etc. who are all trying to do the same. These financials are sane, considering the scenario.
The scale of the numbers is exceptional, but the shape is pretty typical for a high-growth, scale startup with a big TAM where a winner can take most. And compute, supply constrained as it is for the foreseeable future, is absolutely a moat. I come away from this thinking OpenAI is actually in very good shape given that revenue is growing fast enough that break-even has a clear path without doing anything draconian.
You mean the lack of pro-Anthropic/OpenAI comments, who are gambling tokens at their casinos and won't admit that they are very expensive.
This is because people here are quietly realizing that they fell for the "token-maxxing" marketing drive which was complete BS for you to gamble more money on tokens as the big AI labs gave heavily subsidized token prices they cannot afford.
Jevon's paradox does not exist at those companies, but it certainly exists at the Chinese AI Labs at Deepseek, Alibaba, z.AI and Xiaomi.
>This is because people here are quietly realizing that they fell for the "token-maxxing" marketing drive which was complete BS for you to gamble more money on tokens as the big AI labs gave heavily subsidized token prices they cannot afford.
Good callout. All these "trends" in AI were definitely from the AI companies themselves in order to push the sales of more tokens. What's after agent orchestration? Whatever it is, it will involve a big spend.
I'm a simple guy and I don't understand the "sales and marketing" cost.
I don't like these products. I have several negative opinions on them. To the extent they work and there is a customer base what marketing could you /possibly/ be engaged in? Doesn't the product sort of market itself? Or another way is this a product that you can market to expand your MAUs?
It's so polarizing I can't imagine how that $5.7B is being spent.
Half of the comments on this site at any given moment are from bots or shills shilling OpenAI and Anthropic. Now include Reddit, Twitter and everywhere else with a tech audience, paying for all that "organic" marketing doesn't come cheap.
I didn't look at the financials but the subscription product is heavily discounted relative to the API pricing and that difference could well be booked as a marketing expense. They also have a string of grant and similar initiatives (like $50M each) that could be marketing. There's a lot of stuff they could assign at least partially to marketing, and it sounds like they spend money pretty freely.
I cannot consume any content anywhere without being slapped in the face with an unending stream of OpenAI ads and paid plugs. I'd guess most of that money is going directly to Google and Facebook.
I've seen physical billboards in the Portland, OR area for OpenAI, so I guess that accounts for at least part of it. Not really sure what kind of return they're getting on those but apparently they can just do whatever they want, even if they're losing money.
They need marketing because they have competition that essentially offers an identical product. Why should a consumer choose openai over anthropic or whatever else there is? The answer is not obvious.
They have a large and rapidly growing enterprise sales organization. If you want to sell to enterprises you need account executives, solutions engineers, forward deployed engineers, etc.
>It's so polarizing I can't imagine how that $5.7B is being spent.
In every way imaginable and then more, looks like beyond the imagination :)
>I don't like these products. I have several negative opinions on them.
You're not alone, and the crowd seems to be building at the same time enthusiasts are proliferating too.
So much widespread negativity I would guess that's about what it's expected to cost to fully overcome resistance and objections. Which must be bigger than we think, they sure have more information than us.
I've seen lots of ads saying I should use chatgpt to plan a workout or give me recipes. Thats apparently the killer app for 95% of the population at this point.
During the internet bubble collapse in the 00s quite some companies went bankrupt. But that's actually a good thing. It doesn't stop progress. It creates new opportunities and new baselines. Same will happen here. AI will not be less or gone or reduced to useless. It will become better , bigger and faster.
Everyone's financial literacy seems to evaporate when discussing AI companies. They assume that companies need to be profitable or they're a bubble waiting to burst.
The whole point of the company is that they are investing a huge amount of money upfront in order to make models that are better and better, and thus have a higher productivity multiplier.
They are very profitable on inference, they just know that the race to AGI requires a huge amount of investment, compute, getting the best researchers, etc.
I think that the issue most people have is that the degree to which they would need to be profitable in order to pay back their debt is not realistic. It is unlikely that they would be able to get that large a portion of US GDP and if they did then there will likely be riots in the streets.
Look, for coding and a lot of other things, AI is awesome.
But the here's the killer. I have a dinky 16gb VRAM card, and that's kind of the sweet spot for the level of AI I actually want. I don't want it doing too much, I'd rather create slowly than have it one shot something that I have to then pore over later.
Feels like a company investing kazillions in, i don't know, air-conditioning or building wi-fi. Yes, it's going to be around, and also no one's gonna need THAT MUCH.
If anything this is MORE evidence that the infinite money printer will be coming online any second now! Yep aaaaany second now... OH THERE IT- awww one of you guys wasn't praying hard enough.
Anyone remember how immensely incorrect most of HN commenters were on Uber's eventual profitability? For years we heard endless admonishment of Uber being an unsound business model. They made $10b in profit last year, $150b company at 18 P/E ratio. I would take the average HN opinion of business profitability with a grain of salt.
Seeing that R&D costs are the lion's share, I wonder if we are at a point where the focus can shift to improving the cost of inference.
Unless we are genuinely pushing to find AGI, at which point nothing matters, LLMs in their current form don't replace knowledge workers but are an effective force multiplier. How good is enough?
For instance, I pay about $1-2 a month for DeepSeek. It's not as sophisticated as Claude, but it still doubles my productivity as a SWE.
If Fable comes out and demands 50x the price of DeepSeek in order for Anthropic to make a profit on it, how much more productive would I be compared to my personal experience + DeepSeek? 3x? 50x?
Is it cost effective for a business to hire someone without SWE experience + Fable verses hiring someone with SWE experience and DeepSeek? When does R&D hit diminishing returns?
> I wonder if we are at a point where the focus can shift to improving the cost of inference.
There's always working on improving the cost of inference, but I don't think this is an area of R&D that will slow down. The reason is:
1. A better competitor model risks eating away at how much they can charge for inference (i.e. revenue) 2. Whoever unlocks AGI will unlock even more growth 3. Even when you unlock AGI, you'll want to throw gobs of money at it to improve itself and all sorts of things.
> If Fable comes out and demands 50x the price of DeepSeek in order for Anthropic to make a profit on it, how much more productive would I be compared to my personal experience + DeepSeek? 3x? 50x?
You're pricing it wrong and looking at it wrong. First, the per token price doesn't consider that a smarter model can end up using fewer tokens overall to achieve a result. Secondly, if the difference is between failing to accomplish the task and accomplishing the task, suddenly that 50x can seem like a bargain.
> Is it cost effective for a business to hire someone without SWE experience + Fable verses hiring someone with SWE experience and DeepSeek? When does R&D hit diminishing returns?
At this time, someone without SWE experience + <name AI model> vs someone good with SWE experience and <name another AI model> is a no-brainer. The AI model is an accelerant but the "no SWE experience" will be accelerated into a wall. Now maybe that doesn't matter for prototyping and certain other things, but anything in production the lack of experience will hurt them with things they won't even know about or even know how to look for it (e.g. slow, insecure, etc).
> You're pricing it wrong and looking at it wrong. First, the per token price doesn't consider that a smarter model can end up using fewer tokens overall to achieve a result. Secondly, if the difference is between failing to accomplish the task and accomplishing the task, suddenly that 50x can seem like a bargain.
Perhaps, yeah. Napkin math; if we go off of the assertion that DeepSeek makes me twice as productive for an extra $24 per year.
The same number of tokens with Fable would be around 50k/y. Say Fable uses half as many, that's still 25k
The question is then; How does Fable make an engineer more productive than DeepSeek?
- Prompt adherence?
- Inference speed?
- Can engineers delegate more work to unsupervised autonomous workflows?
Are unsupervised workflows suitable for production applications or does the review overhead erode the productivity benefits?
Perhaps a company could hire a mid level engineer + Fable to achieve senior engineer levels of productivity for the same price as a senior + deepseek.
For clarity, inference is typically a COGS and therefore hits Gross Margin vs model training which would typically be in OpEx (where R&D lives) and would hit operating margin.
Even if you discount superhuman AI (which I would emphasize that frontier researchers do not discount and expect to see soon) think it’s still hard to have enough confidence that the ground is solid. Someone in 2024 trying to go down this route would have invested a lot of now-pointless effort into prompt engineering.
So... 50% operational costs and about $100 spent on sales for each paying customer.
If they manage to keep those customers for several years without more sales, that bit looks like a normal "high-touch" business.
They shouldn't look like a "high-touch" business, but their unitary numbers look way better than I expected. They just need to grow some 10 times to star making a profit... Maybe 100 to cover the opportunity cost of their capital.
It's just a matter of finding 5 billion people willing to pay US prices :)
But it is still better than I expected.
The win for something like OpenAI isn't getting a ton of customers to pay $10-100/mo.
It's getting businesses to pay $2k/mo or more per professional employee, like a lot of Anthropic customers.
Anthropic is ahead of them there, but that is how they win.
> just a matter of finding 5 billion people willing to pay US prices
This is how you know ads are inevitable. YouTube is probably a good indicator of how BigLabs will operate for free users.
I'd be cool with that. YouTube premium is one of the best value subscriptions I have. Steering people toward paying instead of ads-by-default is a net good imo
I think we can all understand the ways in which embedded advertisement in LLMs will be fundamentally different than view-based advertisement.
Sam didn't lie, they are in fact a non profit.
You are an LLM dataset's worst nightmare. XD
”The company reports over 900 million weekly active users of ChatGPT, though only about 50 million of those are paid subscribers.”
With so many free models available the ai companies are going to struggle to convert active free users to paid.
They won't try to. ChatGPT is already starting with ads, which is potentially far more profitable (as evidenced by the fact that the most profitable company of all time makes 90%+ of their revenue through ads).
>as evidenced by the fact that the most profitable company of all time makes 90%+ of their revenue through ads
the biggest reason for this is that the digital ad market is a duopoly (charitably a triopoly if you count Amazon in), if all of the LLM companies start to go into ads that's going to be a much more competitive market for ad buyers. It's not going to be so straight forward when both customers and merchants have ten different places to go.
Also not to forget that ChatGPT has zero moat, unlike social Facebook and Google.
None of the free models offer anything even remotely close to the output you can get on a relatively inexpensive model.
I think that AI is going to become just another utility people pay to stay relevant. Same as their internet, electricity or gas.
> another utility people pay to stay relevant
I'm guessing that might be so in certain professions, but I would expect the employer to pay for that. For the rest of us, it seems unlikely. At least for me, I don't have a need of a device to generate text for me. And I bet most people are are in the same boat as me.
Sonnet 4.6 is free and works really well for coding for me from just pasting `tree` and `cat` output directly in the chat window on claude.ai
Will they do it at utility / commodity prices though, or the inflated costs we see now?
These numbers seem insufficiently detailed to really evaluate anything. They’re had $13bn in gross revenue in 2025, and they cost of that revenue was $7.5bn. Both are growing fast (we assume) and the ratio ought to stay roughly constant.
But: how are they calculating the cost of revenue? Do they have rapidly depreciating assets that are also needed to produce that revenue? (Starlink has this issue.) Will their cost per arithmetic operation for inference rise or fall? (Anthropic is paying xAI an absolutely insane amount to lease GPUs. They must be betting that they will not need to repeat that.) Is a large portion of the cost allocated to R&D actually being used to support their revenue?
I certainly believe that the cost of inference can be plenty low for them to make a profit, but a more granular breakdown would make it easier to evaluate.
Discussed yesterday: https://news.ycombinator.com/item?id=48550465
Is this surprising to anyone? I thought that was a given. I'm getting de-facto unlimited use of a model more expensive than Opus 4.8 for $20 a month.
I feel like I have a different $20 plan than everyone else. I have no problem hitting my 5 hour and weekly limits. Don’t get me wrong, it’s a great deal compared to API pricing, but it’s a far cry from “unlimited”.
I get about 20 minutes of work from my 5h limit with the $20 plan. It wouldn't bother me as much if codex would continue after the token bucket refills instead of waiting for me to show up and tell it to continue. I don't jump to the $100 plan because I would be in the exact same situation.
Harness matters in this. Using the Codex sub with Hermes eats tokens like nothing. Using it with Pi is much less but you don’t get the long term memory. When you were able to use the Claude subscription with Pi, I barely hit the 5hr limit. When they stopped allowing that, CC harness just chews thru tokens.
Interesting. I'm mostly using Claude, so perhaps I'm not nearing the limits, but I do use Codex (for coding and reviews occasionally) and use chatgpt for second opinion many times, including "pro" research. Never got to my limits. But again, not my main go to tool.
> de-facto unlimited (...) for $20 a month
Would love to hear some details on that one...
Or was that a typo and you meant the $200/mo plan instead maybe? That one I could believe, assuming no or frugal subagent use that is.
Copying my other reply.
> Interesting. I'm mostly using Claude, so perhaps I'm not nearing the limits, but I do use Codex (for coding and reviews occasionally) and use chatgpt for second opinion many times, including "pro" research. Never got to my limits. But again, not my main go to tool.
If these numbers are right, it's actually not that bad. Cut r&d costs and they are mostly profitable.
Yes if you ignore all the reasons why they’re horribly unprofitable, they’re profitable.
R&D costs are hurting profit side and while you can cut that one just becomes irrelevant overnight in this space if you do, hence the problem.
Not to mention they will need to research how to make their models faster and cheaper to run in order to fit some margin within what people are actually willing to pay.
> R&D costs are hurting profit
That’s quite the hot take, considering it’s literally an R&D company that got to where it is by doing R&D.
Isn’t the post above saying the same thing after the part where you cut it off…?
So you’re saying if you cut all the cost centers a company would only have profit centers? If you ignore all the losses you’ll only have profits?
It's more like once you figure out how to make a really good lamp then producing lots of lamps will be profitable. But the lamps are currently suboptimal so we'll be in the red until that time.
It's more like you have a business making engines, each generation of engine has eventually turned out to be profitable over its lifespan, but each generation has an exponentially increasing R&D cost and your customers will switch from the old engines to a competitor if they don't like the newest generation.
You're stuck racing against your competitors with the distinct possibility that your R&D costs will outgrow the market demand, and you can't stop because otherwise your customers will stop investing in your dead end tech and switch.
OpenAI won't be able to cut R&D spend and collect rent on their existing models as long as the Chinese models keep up the pace of being ~6 months behind them for a fraction of the price.
And if you wait 12 months, someone will be giving away lamps for free that work just as well.
And then someone will come up with lamp pro max and you’ll be out of business. You realize why R&D exists in tech companies even though it’s a cost center right?
This is private equity 101 no?
OpenAI can easily cut R&D costs by replacing engineers with Claude Code
I am having difficulty parsing this sentence ... :-)
Cut down on the one thing they need to keep themselves relevant in this space?
Watch them flare out like a star… but there is lots of questions re the the return on RnD. Is it worth spending another order of magnitude for only marginal frontier gains?
While you cant discount 100% R&D they are close, agreed
I bet any FAANG spend is mostly R&D.
If it's not materials, not energy or taxes, not manufacturing, not licensing or rental fees, then I can only think of R&D.
People keep overlooking the fact that costs for these providers scale along with customer acquisition. Most startups don't have that linear expense. Also, training costs are accelerating to get new models out faster. One doesn't simply "get rid of R&D" costs as a comment upstream mentioned. I can't actually imagine R&D goes down anytime soon unless you're willing to play third fiddle.
Unless these frontier providers feel some type of squeeze or constraint the Chinese are well positioned to leave the US bag holders of an NVidia bound system. And if anyone has to wonder how one provider for a critical piece of infrastructure will go, well...
Actually reduce R&D to ZERO and they are still losing money.
If they cut down on R&D they will be no better than the open source models you can run at cost yourself.
Even if they keep the R&D costs, more efficient inference and 0 Marketing spend also gets you there. Inference is honestly super inefficient at this point, we can do far better than GPUs, push utilisation up, build more efficient datacentres.
Numbers are probably not right as classifying everything aa r&d is going to the temptation
My takeaway from this is that it's incredibly validating as a business model. Inference is _highly_ profitable. Of course, like any company that has ever tried to grow at breakneck pace, you run at a loss until you "win."
Isn't this what all of the big companies that spend a lot on R&D and engineers promise?
And then the reality turns out not to be the case - you have to continuously spend on R&D to avoid getting your lunch eaten by someone else.
This isn't a social media network with lockin either. People can and will just switch to whatever whenever they feel like it. Maybe it becomes a defacto standard like google but if someone is much better than you, well...
> My takeaway from this is that it's incredibly validating as a business model. Inference is _highly_ profitable.
The problem is you can't just separate training costs from inference costs. If OpenAI just didn't train a new model for the next five years, sure, they'd do OK. Assuming all those dirt cheap Chinese models nipping at their heels don't make up the gap while OpenAI is resting on their laurels.
Without being a frontier model (read: continuous, incredibly expensive training), they effectively don't have much to sell. So inference and training costs are intertwined to some extent.
>> Inference is _highly_ profitable.
Totally untrue.
Yes, it is like a new era - the startups have huge direct revenue on real products instead of "users" which yet to be monetized.
And the network effect which ruled for the last 20 years seems to have relaxed its death grip just a bit (of course it is still there as having more customers using your tools and models provides more training data, etc., yet the current network effect doesn't seem to have that high exponential value like before)
I wonder how effective the marketing is (not much it seems).
I was watching a World Cup match last week and one of the TV ads during half time was something to the tune of ChatGPT being used by kids to improve their street soccer skills. This was Brazilian TV. Anyone even remotely familiar with Brazil would find this ad deeply, thoroughly out of touch. I can't think of a worse chatbot pitch than that.
Almost 6 bln in sales in marketing? It looks an enormous amount given that they used to have the best models and used to give-aways tokens.
6bn seems excessive but despite GPT 5.5 arguably being better than Claude I don't see a lot of adoption of Codex yet.
Some of my coworkers even use Sonnet (the default in Claude Code for the 20 USD subscription) and see no reason to change even though that model is definitely "outdated" compared to current SOTA.
Marketing might help at some workplaces, presumably that are dedicated to Microsoft, for example our network blocks Claude (and DeepSeek) and is slowly rolling out Codex team by team. They should encourage Amazon/AWS to market for them.
most people are behind the curve
Who needed leaks to know that?
I'm really curious about something: how far will you go to support AI? Clearly they'll need to monetize things further, would you still use [whatever AI you are paying for] if the price was doubled? Tripled? Where would you stop and would you stop using AI altogether or would you look at competitors?
I will do nothing to “support” AI. Either it has utility or it doesn’t. I feel no loyalty or duty to help make it work if it doesn’t.
Anyway: Zero, as of right now.
I fully expect to be able to run useful LLMs on a machine I can justify buying for other reasons. I already can on the secondhand kit I own, and I don’t expect the cost-benefit analysis of local LLMs to ever really get worse.
If I ever need to pay for it, it will likely be to shift some of the capacity into the cloud for either business or pragmatic personal reasons (so I can just carry an iPad etc.)
I fully intend my expenditure to be negligible. Because once one realises that outspending others is impossible, only spending minimisation makes sense.
I foresee it potentially making sense for me to move some mature tools off a local LLM to openrouter, maybe. But probably to the same or similar models.
I don’t and won’t support AI. For a while I paid 200€ a month and would have been happy to pay up to maybe 600€. However I don’t want to participate anymore in using such an anti-human technology and industry
I don’t think there’s a single person out there that will ‘support’ AI
Maybe it’s just your phrasing but people will only pay for what works, no one is loony enough to support a trillion dollar industry out of the kindness of their heart or spirit of innovation
I make an important distinction between cloud services and local AI. My lifetime spending on cloud AI is probably less than $500, and I don't intend to spend any more. But I've already dropped $2.5k on new hardware for local inference, and could easily see myself spending more in the future. In fact, I'm regularly browsing for deals. I would also be open to paying for local models, if there was a way to make that compatible with fully open models.
AI is so important, I want to have it under my control. Even if I have to pay a penalty in terms of capabilities.
It depends on how much time it saves me and how much I make per hour generally, right?
If AI allows me to cut my time to do something in half on average or allows me to do 2x more it would be worth it to pay up to what my monthly income was before assuming my income scaled with my output.
I pay for good tools that I use.
I spend 30 - 60 bucks a year with Horizon Labs.
I spend 25 bucks a month on Cursor. Cursor replaced an OpenAI sub.
Both support hobby projects. If either cost increased I would spend some time testing local alternatives and probably drop them.
Horizon Labs especially, I know that they have been matched by open models and are mostly a convenience at this point.
I've spent a grand total of $25 on AI ever, so apparently my answer is $25. But I'm not a big time software dev like the rest of you.
When I bought my last GPU, running AI models locally was a consideration though not the only one, and I have it set up but haven't used it much yet. I mostly use the free tiers of ChatGPT or Google to write the occasional script for me. I guess they're going to have to inject a truly unfathomable number of ads to get their money's worth.
I have a feeling my experience is closer to an average persons' than a dev, but it doesn't seem like they'll be able to monetize just from devs even if each one is spending thousands a month.
I'm not a coder but now work way faster than the coder I pay, stuff breaks but it's tenable and it's easier to get things to completion as the harnesses get better.
Don't give up just keep trying you can truly build personally life changing things. Don't look at it purely from a how do I sell this lense, just empower yourself with these tools while the getting is good
I have made life changing things with it, just not anything so life changing I'd consider paying more than $25. Stingy bastard, I am.
For personal use not more than 30$/month.
For work, it depends, but if I have to spend more than a few hundreds bucks probably I'll start looking for alternatives (local models, Chinese providers, ecc)
PS: I'm in Italy, I guess in several parts of the world these figures are even smaller.
Max 60 bucks a month. More than that and I'd just move to local qwen 35b or some other cheaper model on openrouter.
I’d easily pay multiple hundreds. Possibly a thousand a month.
If I were really forced to.
LLMs provide me about the same value as a car does.
I’d pay thousands a month, if I had no cheaper choices, my productivity is now limited by the intelligence of AI, I’m basically a PM now.
What in the world are you working on?
I’m honestly just thinking about day to day utility in my personal life.
Paying a thousand a month for a car is also very stupid.
Stretching the analogy, something that gets you from point A to point B for a fraction of the price without the same level of comfort is totally fine for me. For some of my tasks, that means using local models. For others it might mean a frontier-last-year kind of model. That's totally acceptable most of the time. For anything else I guess it's like renting a truck to move; just get the right vehicle as needed and pay the premium.
A $50k car used 1,000 miles per month probably costs close to a thousand per month, assuming 200k miles of life. I imagine this is not unusual in the US.
Agreed. For personal use it's already easily worth $100 a month (to me personally). More probably. For work, it's entirely based on its financial impact for a given role, and for some people/companies it will be worth the cost even at $X thousand per month per seat.
That's crazy. Can you provide some examples?
I would probably still pay if the cost doubled, but I would also look at competitors, offline solutions, etc
We have benchmarks on our domain and it does there are models that are 2x to 10x cheaper for a small drop in percentage points in accuracy
Never paid a cent, never will pay a cent. I have my principles.
It may put me at a disadvantage when it comes to quickly slop something together? But so far the free-to-use chat bots do as well for my needs.
Zero. It provides no value to me.
This title is not how I'd actually interpret the results.
Glad to see more sane takes in the comments. All these articles on their current financials are missing the point.
OpenAI is doing pretty well.
Capital expenditure is required to deliver on 1) better models 2) better infra and 3) better products. Insane CapEx is required to do all the above + compete with Google, Meta, Microsoft, Apple, Anthropic, etc. etc. etc. who are all trying to do the same. These financials are sane, considering the scenario.
The scale of the numbers is exceptional, but the shape is pretty typical for a high-growth, scale startup with a big TAM where a winner can take most. And compute, supply constrained as it is for the foreseeable future, is absolutely a moat. I come away from this thinking OpenAI is actually in very good shape given that revenue is growing fast enough that break-even has a clear path without doing anything draconian.
Pardon my French, but yeah, no shit?
AI companies are black holes for money the way delivery companies are (or were, considering the money people are willing to pay these days).
Most of them will disappear alongside the money people have bet on them.
Suspicious lack of pro-AI comments here
their PR department is probably still trying to figure out what narrative the bots should follow for this one.
You mean the lack of pro-Anthropic/OpenAI comments, who are gambling tokens at their casinos and won't admit that they are very expensive.
This is because people here are quietly realizing that they fell for the "token-maxxing" marketing drive which was complete BS for you to gamble more money on tokens as the big AI labs gave heavily subsidized token prices they cannot afford.
Jevon's paradox does not exist at those companies, but it certainly exists at the Chinese AI Labs at Deepseek, Alibaba, z.AI and Xiaomi.
>This is because people here are quietly realizing that they fell for the "token-maxxing" marketing drive which was complete BS for you to gamble more money on tokens as the big AI labs gave heavily subsidized token prices they cannot afford.
Good callout. All these "trends" in AI were definitely from the AI companies themselves in order to push the sales of more tokens. What's after agent orchestration? Whatever it is, it will involve a big spend.
I'm a simple guy and I don't understand the "sales and marketing" cost.
I don't like these products. I have several negative opinions on them. To the extent they work and there is a customer base what marketing could you /possibly/ be engaged in? Doesn't the product sort of market itself? Or another way is this a product that you can market to expand your MAUs?
It's so polarizing I can't imagine how that $5.7B is being spent.
Half of the comments on this site at any given moment are from bots or shills shilling OpenAI and Anthropic. Now include Reddit, Twitter and everywhere else with a tech audience, paying for all that "organic" marketing doesn't come cheap.
I didn't look at the financials but the subscription product is heavily discounted relative to the API pricing and that difference could well be booked as a marketing expense. They also have a string of grant and similar initiatives (like $50M each) that could be marketing. There's a lot of stuff they could assign at least partially to marketing, and it sounds like they spend money pretty freely.
I cannot consume any content anywhere without being slapped in the face with an unending stream of OpenAI ads and paid plugs. I'd guess most of that money is going directly to Google and Facebook.
I've seen physical billboards in the Portland, OR area for OpenAI, so I guess that accounts for at least part of it. Not really sure what kind of return they're getting on those but apparently they can just do whatever they want, even if they're losing money.
They need marketing because they have competition that essentially offers an identical product. Why should a consumer choose openai over anthropic or whatever else there is? The answer is not obvious.
OpenAI will make fully autonomous killing machines while Anthropic wont.
Land mines are my favorite fully autonomous killing machine. They've also been around for a while.
The worst ones looked like brightly coloured children's toys.
There's perhaps a metaphor or two lurking about bait and switch tactics.
They are paying influencers to pretend they use LLMs, and discredit Chinese models: https://www.wired.com/story/super-pac-backed-by-openai-and-p...
They have a large and rapidly growing enterprise sales organization. If you want to sell to enterprises you need account executives, solutions engineers, forward deployed engineers, etc.
>It's so polarizing I can't imagine how that $5.7B is being spent.
In every way imaginable and then more, looks like beyond the imagination :)
>I don't like these products. I have several negative opinions on them.
You're not alone, and the crowd seems to be building at the same time enthusiasts are proliferating too.
So much widespread negativity I would guess that's about what it's expected to cost to fully overcome resistance and objections. Which must be bigger than we think, they sure have more information than us.
It costs money to get influencers to set up kool-aid stands on their platforms.
I've seen lots of ads saying I should use chatgpt to plan a workout or give me recipes. Thats apparently the killer app for 95% of the population at this point.
Don’t forget changing the background of a picture. This alone can triple the GDP.
Practically printing money!!!
That aligns pretty well with a past job. Those two areas were very popular user interests. Third one was cosmetics like skincare routine.
During the internet bubble collapse in the 00s quite some companies went bankrupt. But that's actually a good thing. It doesn't stop progress. It creates new opportunities and new baselines. Same will happen here. AI will not be less or gone or reduced to useless. It will become better , bigger and faster.
I want to see the person who thought they were losing only hundreds of millions
Beginning to see why he needed seven trillion dollars.
Everyone's financial literacy seems to evaporate when discussing AI companies. They assume that companies need to be profitable or they're a bubble waiting to burst.
The whole point of the company is that they are investing a huge amount of money upfront in order to make models that are better and better, and thus have a higher productivity multiplier.
They are very profitable on inference, they just know that the race to AGI requires a huge amount of investment, compute, getting the best researchers, etc.
I think that the issue most people have is that the degree to which they would need to be profitable in order to pay back their debt is not realistic. It is unlikely that they would be able to get that large a portion of US GDP and if they did then there will likely be riots in the streets.
I'm not surprised
Ha, not a problem.
Look, for coding and a lot of other things, AI is awesome.
But the here's the killer. I have a dinky 16gb VRAM card, and that's kind of the sweet spot for the level of AI I actually want. I don't want it doing too much, I'd rather create slowly than have it one shot something that I have to then pore over later.
Feels like a company investing kazillions in, i don't know, air-conditioning or building wi-fi. Yes, it's going to be around, and also no one's gonna need THAT MUCH.
Leaked: OpenAI is a rapidly scaling startup, has economics similar to other startups
If anything this is MORE evidence that the infinite money printer will be coming online any second now! Yep aaaaany second now... OH THERE IT- awww one of you guys wasn't praying hard enough.
Anyone remember how immensely incorrect most of HN commenters were on Uber's eventual profitability? For years we heard endless admonishment of Uber being an unsound business model. They made $10b in profit last year, $150b company at 18 P/E ratio. I would take the average HN opinion of business profitability with a grain of salt.