Why Everybody Is Losing Money On AI
wheresyoured.at
wheresyoured.at
> Cursor sends 100% of their revenue to Anthropic, who then takes that money and puts it into building out Claude Code, a competitor to Cursor. Cursor is Anthropic's largest customer. Cursor is deeply unprofitable, and was that way even before Anthropic chose to add "Service Tiers," jacking up the prices for enterprise apps like Cursor.
Cost to run a million tokens through GPT-3 Da-Vinci in 2022: $60
Cost to run a million tokens through GPT-5 today: $1.25
This is particularly important if you constantly replay the previous conversation and grow it with each subsequent prompt.
GPT-5 offers a 90% discount on these cached tokens! That's a huge saving for this kind of pattern.
An unknown to me: are the less powerful models cheaper to serve, proportional to how much less capable they are than frontier models? One possible explanation for why e.g. OpenAI was eager to retire GPT 4 is that those older models are still money losers.
The strongest evidence is that the models I can run on my own laptop got massively better over the last three years, despite me keeping the same M2 64GB machine without upgrading it.
Compare original LLaMA from 2023 to gpt-oss-20b from this year - same hardware, huge difference.
The next clue is the continuing drop in API prices - at least prior to the reasoning rush of the last few months.
One more clue: o3. OpenAI's o3 had a 80% price drop a few months ago which I believe was due to them finding further efficiencies in serving that model at the same quality.
My hunch is that there are still efficiencies to be wrung out here. I think we'll be able to tell if that's not holding if API prices stop falling over time.
It's much cheaper to timeshare that hardware with other users than to buy and run it yourself.
I suppose there's a distinction between new less capable models, I can see why those would be more efficient. But maybe the older frontier models are less efficient to serve?
Plus I believe they have to keep each model in GPU memory to serve it, which means that any GPU serving an older model is unavailable to serve the newer ones.
[0] https://venturebeat.com/ai/openai-announces-80-price-drop-fo...
Generally I'm skeptical of the idea that any of the major providers are selling inference at a loss. Obviously they're losing money when you include the cost of research and training, but every indication I've seen is that they're not keen to sell $1 for 80 cents.
If you want a hint at the real costs of inference look to the companies that sell access to hosted open source models. They don't have any research costs to cover so their priority is to serve as inexpensively as possible while still turning a profit.
Or take a good open weight model and price out what it would cost to serve at scale. Here's someone who tried that recently: https://martinalderson.com/posts/are-openai-and-anthropic-re...
Simply because strategically if there's high long term potential, it initially makes sense to put more money in than you get out of.
Not saying that AI is this, but if you determined that you have a golden goose that laid out 10 trillion USD worth of eggs when it got 10 years old, how much would you pay for it in the auction, and what would you have to show for it for the initial 9 years?
Now what if the golden goose scaled to 10 trillion each year linearly? First years people sound of mind would overpay for what it makes.
I personally think that there's many levels of innovation still to come in terms of robotics, APIs/Frameworks/Coding languages/Infra specifically for LLMs to provide easier and more foolproof ways to code and do things otherwise. I think it's far from played out. I personally think that a lot of potential is still untapped.
But the people that didn't definitely had a harder time adjusting when it got increasingly annoying to live without a smartphone.
It's ultimately a choice you can make, but it definitely also comes with consequences - especially if your dayjob is software - as this is an industry that loves to discriminate against people that aren't aboard the hype train and don't have "10 yrs of experience in d̵o̵c̵k̵e̵r̵ LLMs"
There is a large market for Java, C++ and COBOL engineers to this day, despite all startups on here are talking about React and Rust. There will still be a large need for actual engineers that use their meat brain and are not paid by line committed for the foreseeable future. Not everyone is writing junior-tier boilerplate that benefits from LLMs.
However, the writing is on the wall and it's likely going to become one of the bullet points you'll be expected to have significant experience in when changing jobs. And I doubt that's gonna take 10 yrs
And no one was forcing phones on you first 5-10 years of it existing. There was advertisement and competition like with any other product. You was certainly not getting free phones and there was no real top down push to have them like we see with ai.
Think that's rare? Nope. It's coming everywhere. Most companies are at the stage of trying to monitor AI usage and encourage it, but eventually it'll stop being optional.
Company trying to get rid if innefective people would make sense, but if you measure it by ai usage all that happens is that your ai usage will go up - and price of using it along with it.
There are always irrational managers enjoying power trips. But the norms normally dont become that irrational
Imagine you had a job doing ordinary database backed web app written in Java, and you found you had a coworker who wrote all their code in Notepad. They also refused to run linters or open code reviews, viewing it all as modern nonsense. Would you find that acceptable? Would you be surprised if a few months later that guy got let go for performance reasons? No.
Developers are given a lot degree of freedom, but in return are expected to use that freedom responsibly to deliver as much value as they can for the company. People refusing to use powerful tools had better have a watertight explanation for why. Mere negative vibes aren't good enough.
Most people miss that they have almost a billion free users that are waiting to be monetized. Google makes 400B a year and it's crazy to think OpenAI can't achieve some percentage of that. Why would you slow down and let Google catch up for the sake of short term profitability.
> In fact, even if you remove the cost of training models from OpenAI's 2024 revenues (provided by The Information), OpenAI would still have lost $2.2 billion fucking dollars.
The issue however is, can an AI company actually go "yep. We're done this is a good as it gets!"?
I don't believe they can do that, so removing training cost is kind of a moot point.
Yes, every new technology has always stayed exorbitantly priced in perpetuity.
We're standing here with a kind of survivorship bias because of all the technologies we use daily that did cost reduce and make it. Plenty did not. We just forget about them.
Except the authors own provided data says it cost them $2B in inference costs to generate $4B in revenue. Yes training costs push it negative, but this is like tech growth 101, debt now to grow faster leads to larger potential upsides in the future.
But training will have to stay forever, right? Otherwise the LLM will be stuck with outdated information...
The people writing all of these "AI is unprofitable" pieces are doing financial journalism similar to analyzing the dot-com bubble by looking at pets.com's burn rate. The infra overspend was real as well as the bankruptcies, but it existentially foolish for a business to ignore the behavioral shift that was taking place.
I have to constant remind myself to stop arguing and evangelizing about AI. There is a growing crowd who insists that AI is a waste of money and that AI cannot do things I'm already doing on a daily basis.
Every minute spent explaining to AI skeptics is a minute not spent actually capitalizing on the asymmetry. They don't want to hear it anyway and I have little incentive to convince anyone otherwise.
The companies bleeding money to serve AI below cost prices won't last, but thats all more the reason use them now while they're cheap.
Obviously the goal of these companies is to generate as much profit as possible as soon as possible. They will turn the tables eventually. The asymmetry will go in the opposite direction, maybe to the extend that one takes advantage of the current asymmetry.
Those were companies that had multiple expensive IT restructurings one after another, each making them more innefective and then either run out of cash or barely made it.
It worked well for companies that were choosing smart.
LLMs have been around for two years. it took decades before the PC really took hold.
2) Yes, they still have some time to fit the market better, but that doesn't change what they'll need to do to fit the market better.
But virtually everybody has been using LLMs already. How long would it have taken for the PC if everybody had had the opportunity to use one for more than a year?
Whereas the PC was clearly economically viable.
e.g. at time of writing a post about MentraOS has 11 points in 1 hour compared to this article's 51 in 53 minutes, but this is ranked 58th to Mentra's 6
Dang, can we chat about collaborative filtering bubbles please?
That’s why I use https://news.ycombinator.com/active to find the interesting topics instead of having to rely on the strict moderation algorithm
The bulk of the cost was model training and research(~4B). They are forced to train new models and improve existing one because of the market and competition.
You can't just strip out those costs. You have to train new models or the information in the model will be out of date.
The only scenario where the training cost won't decrease is in case the limit of the scaling law is not yet reached, or they discover new approaches. But from what we have seen this year, it doesn't seem to be the case.
PS: Plus, we are still not speaking about the elephant in the room: ads. Today's revenues are basically from API usage and subscription, but we all know that at some point ads will come in, and the revenues will increase.
There is clearly some kind of market for this technology. It will eventually be profitable either through some technology breakthrough that allows creating/processing tokens cheaper or by finding a cost structure consumers can live with.
The cat is already out the bag. This technology isn't going away.
This means that they operate existing models with very healthy 50% profit margin, that’s excellent unit economics actually. Losing money by investing more into R&D that you make is not the same as burning it by selling a dollar for 90 cents.
Most all saas accounts require lengthy and generous free trials and boy AI compute throws a bit of a hand grenade into that PLG strategy. It’s just new economics that the industry will have to figure out.
You can't just eliminate all the costs except inference
Still, companies like OpenAI and Twitter are doing just that. Thus losing money.
Will AI evolve to be again as regular software or will the business model of tech AI become closer to what traditional non-tech companies are?
How the WalMart of AI will look like?
Does SaaS with very high prices and very thin margins even work as a scalable business model?
Standard market share logic.
By simply stopping the training of new models, profitability can be achieved on the same day.
With the existing models, we have already substantial use cases, and there are numerous unexplored improvements beyond the LLM, tailored specifically to the use case.
But then they stop being up-to-date with... the world, right?
Though not really representative of what users of said models may experience financially, at this point the question should be raised: if AI compute is 7x more expensive than developer salaries, what's the point? I thought the whole idea was to save money on human resources...
It can then be used indefinitely for the cost of inference, which is cheap and will continue getting cheaper.
Yes, a pullback will kill some weaker companies, but not the ones with sufficient true fans. Plus, we’re talking about a wide-ranging technological revolution with unknown long term limits and economics, you don’t just give up because you’re afraid to spend some money.
I don’t want to pay Anthropic, because I don’t trust them, but I will absolutely pay cursor, because I trust them, and I doubt I’m alone. My cursor usage goes to GPT-5, too, so it’s definitely not 100% Anthropic, even if I’m the only idiot using GPT5 on Cursor
It’s fun to innovate. Making money is a happy byproduct of value creation. Isn’t the price of success always paid in advance, anyway? Why would winning AI tech companies pack it up and stop crushing it over the long term just because they’re afraid to lose someone else’s money in the short term? Wouldn’t capitulation guarantee losses moreso than continued effort?
Executive egos and market hype
Those who invest in money-losing AI believe that it will be the next Google and that profits will come.
Or alternatively, they hope to sell to the greater fool.
Models trained in 2025 don’t ship until 2026/7. That means the $3bn in 2025 training costs show up as expense now, while the revenue comes later. Treating that as a straight loss is just confused.
OAI’s projected $5bn 2025 loss is mostly training spend. If you don’t separate that out with future revenues, you’re misreading the business.
And yes, inference gross margins are positive. No idea why the author pretends they aren’t.
The majority of these companies know they are burning money, but more than that knew they would be losing money at this point and beyond. That is the play, the thesis is: AI will dominate nearly everything in the near future, the play is to own a piece of that. Investors are willing to risk their investment for a chance of getting a piece of the pie.
Posts that flail around yelling companies 'losing money', without addressing the central premise are just wasting words.
In short, do you think AI is not going to dominate nearly everything? Great, talk about that. If you do believe is, then talk about something other than the completely reasonable and expected state of investors and companies fighting for a piece of the pie.
As a somewhat related tangent, people seem to not understand the likely cost trajectory of model training/inference costs:
* Models will reach a 'good enough' point where further training will be mostly focused on adding recent data. (For specific market segments, not saying that we'll have a universal model anytime soon, but we'll soon have one that is 'good enough' at c++, might already be there).
* Model architecture and infrastructure will improve and adapt. I work for a company that was among the first use deep learning to control real-time kinetic processes in production scenarios, our first production hardware was a nvidia Jetson, we had a 200ms time budget for inference, and our first model took over 2000! We released our product, running under 200ms, *using the same hardware* the only difference was improvements in the cuDNN library and some other drive updates and some domain specific improves on our YOLO implementation. Long story short, yes inference costs are huge, but they are also massively disruptable.
* Hardware will adapt. Nvidia cash machine will continue, right now nvidia hardware is optimized for balance between training and inference, where TPUs, the newer ones are more tilted towards inference. I would be surprized if other hardware companies don't force Nvidia to give the more inference based solution and 2-3x cost savings at time point in the next 5 years. And for all I know, perhaps a hardware startup will disrupt Nvidia, it would be one of the most lucrative hardware plays on the planet.
Focusing inference cost is a deadend to understanding the trajectory of AI, understanding the *capability* of AI is the answer to understanding it's place in the future.