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?
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).
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.
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.
This is how you know ads are inevitable. YouTube is probably a good indicator of how BigLabs will operate for free users.
With so many free models available the ai companies are going to struggle to convert active free users to paid.
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 think that AI is going to become just another utility people pay to stay relevant. Same as their internet, electricity or gas.
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.
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.
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.
> 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.
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.
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...
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.
Totally untrue.
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)
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.
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.
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.
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.
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 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.
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.
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.
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.
That’s quite the hot take, considering it’s literally an R&D company that got to where it is by doing 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.
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...
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.
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
AI is so important, I want to have it under my control. Even if I have to pay a penalty in terms of capabilities.
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 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.
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.
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.
If I were really forced to.
LLMs provide me about the same value as a car does.
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
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.
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.
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.
There's perhaps a metaphor or two lurking about bait and switch tactics.