Depends on your definition of profitability, They are not recovering R&D and training costs, but they (and MS) are recouping inference costs from user subscription and API revenue with a healthy operating margin.
Today they will not survive if they stop investing in R&D, but they do have to slow down at some point. It looks like they and other big players are betting on a moat they hope to build with the $100B DCs and ASICs that open weight models or others cannot compete with.
This will be either because training will be too expensive (few entities have the budget for $10B+ on training and no need to monetize it) and even those kind of models where available may be impossible to run inference with off the shelf GPUs, i.e. these models can only run on ASICS, which only large players will have access to[1].
In this scenario corporations will have to pay them the money for the best models, when that happens OpenAI can slow down R&D and become profitable with capex considered.
[1] This is natural progression in a compute bottle-necked sector, we saw a similar evolution from CPU to ASICS and GPU in the crypto few years ago. It is slightly distorted comparison due to the switch from PoW to PoS and intentional design for GPU for some coins, even then you needed DC scale operations in a cheap power location to be profitable.
They will have an endless wave of commoditization chasing behind them. NVIDIA will continue to market chips to anyone who will buy... Well anyone who is allowed to buy, considering the recent export restrictions. On that note, if OpenAI is in bed with the US government with this to some degree, I would expect tariffs, expert restrictions, and all of that to continue to conveniently align with their business objectives.
If the frontier models generate huge revenue from big government and intelligence and corporate contracts, then I can see a dynamo kicking off with the business model. The missing link is probably that there need to be continual breakthroughs that massively increase the power of AI rather than it tapering off with diminishing returns for bigger training/inference capital outlay. Obviously, openAI is leveraging against that view as well.
Maybe the most important part is that all of these huge names are involved in the project to some degree. Well, they're all cross-linked in the entire AI enterprise, really, like OpenAI Microsoft, so once all the players give preference to each other, it sort of creates a moat in and of itself, unless foreign sovereign wealth funds start spinning up massive stargate initiatives as well.
We'll see. Europe has been behind the ball in tech developments like this historically, and China, although this might be a bit of a stretch to claim, does seem to be held back by their need for control and censorship when it comes to what these models can do. They want them to be focused tools that help society, but the American companies want much more, and they want power in their own hands and power in their user's hands. So much like the first round where American big tech took over the world, maybe it's prime to happen again as the AI industry continues to scale.
Why would China censoring Tiananmen Square/whatever out of their LLMs be anymore harmful to the training process when the US controlled LLMs also censor certain topics, eg "how do I make meth?" or "how do I make a nuclear bomb?".
Because China censors very common words and phrases such as "harmonized", "shameless", "lifelong", "river crabbed", "me too". This is because Chinese citizens uses puns and common phrases initially to get around censors.
They are absolutely different flavors. OpenAI is not being told by the government to censor violence, sex or racism - they're being told that by their executives.
News flash: household-name businesses aren't going to repeat slurs if the media will use it to defame them. Nevermind the fact that people will (rightfully) hold you legally accountable and demand your testimony when ChatGPT starts offering unsupervised chemistry lessons - the threat of bad PR is all that is required to censor their models.
There's no agenda removing porn from ChatGPT any more than there's an agenda removing porn from the App Store or YouTube. It's about shrewd identity politics, not prudish shadow government conspiracies against you seeing sex and being bigoted.
Because when a small group of elites with permament term and no elections decides what is allowed and what isn't... and has full control of silencing what's not allowed and any meta discussion about the silencing itself... is different from when an elected government decides it, and then anyone is free to raise a stink on whatever is their version of twitter today without worrying about being disappeared tomorrow
It's not an elected government if you're talking about the US. These policies are also all decided by "elites with permanent term and no elections" you realize right?
They want their LLMs explicitly approved to align with the values of the regime. Not necessarily a bad thing, or at least that avenue wasn't my point. It does get in the way of going fast and breaking things though, and on the other side there is an outright accelerationist pseudo-cult.
Ignoring the moral dimension for a second, I do wonder if it is harder to implement a rather cohesive, but far-reaching censorship in the chinese style, or the more outrage-driven type of "censorship" required of American companies. In the West we have the left pre-occupied with -isms and -phobias, and the right with blasphemy and perceived attacks on their politics.
With the hard shift to the right and Trump coming into office, especially the last bit will be interesting. There is a pretty substantial tension between factual reporting and not offending right-wing ideology: Should a model consider "both sides" about topics with with clear and broad scientific consensus if it might offend Trumpists? (Two examples that come to mind was the recent "The Nazis were actually left wing" and "There are only two genders".)
> they (and MS) are recouping inference costs from user subscription and API revenue with a healthy operating margin.
I tried to Google for more information. I tried this search: <<is openai inference profitable?>>
I didn't find any reliable sources about OpenAI. All sources that I could find state this is not true -- inference costs are far higher than subscription fees.
I hate to ask this on HN... but, can you provide a source? Or tell us how do you know?
I don't have any qualified source and this metric would be likely be quite confidential even internally.
It is just an educated guess factoring costs of running similar/comparable models to 4o or 4o-mini per token, and how azure commitments work with OpenAI models[2], also knowing that Plus subscriptions are probably more profitable[1] than API calls.
It would be hard for even OpenAI to know with any certainty because they are not paying for Azure credits like a normal company. The costs are deeply intertwined with Azure and would be hard to split given the nature of the MS relationship[3]
----
[1] This is from experience of running LibreChat using 4o versus ChatGPT Plus for ~200 users, subscriptions should quite profitable than raw API by a order of 3 to 4x, of course different types of users and adoption levels will be there my sample while not small is not likely representative of their typical user base.
[2] MS has less incentive to subsidize than say OpenAI themselves
[3] Azure is quite profitable in the aggregate, while possibly subsidizing OpenAI APIs, any such subsidy has not shown up meaningfully in Microsoft financial reports.
> but they (and MS) are recouping inference costs from user subscription and API revenue with a healthy operating margin.
As far as I am aware the only information from within OpenAI one way or another is from their financial documents circulated to investors:
> The fund-raising material also signaled that OpenAI would need to continue raising money over the next year because its expenses grew in tandem with the number of people using its products.
Subscriptions are the lions share of their revenue (73%). It's possible they are making money on the average Plus or Enterprise subscription but given the above claim they definitely aren't making enough to cover the cost of inference for free users.
So I do question if OpenAI is able to make a profit, even if you remove training and R&D. The $20 plan may be more profitable, but now it will need to cover the R&D and training, plus whatever they lose on Pro.
Not necessarily. DeepSeek will probably only threaten the API usage of OpenAI, which could also be banned in the US if it's too sucessful. API usage is not a main revenue for OpenAI (it is for Anthropic last time I checked). The main competitor for R1 is o1, which isn't gnerally available yet.
The one your laptop can run does not rival what OpenAI offers for money. Still, the issue is not whether third party can run it, it's just the OpenAI seems not putting API as their main product.
Not quite. In 2 years their revenue has ~20x from 200M ARR to 3.7B ARR. The inference costs I believe pay for themselves (in fact are quite profitable). So what they're putting on their investor's credit cards are the costs of employees & model training. Given it's projected to be a multi-trillion dollar industry and they're seen as a market leader, investors are more than happy to throw in interest free cash flow now in exchange for variable future interest in the form of stocks.
That's not quite the same thing at all as your credit card's revenue stream as you have a ~18%+ monthly interest rate on that revenue stream. If you recall AMZN (& all startups really) have this mode early in their business where they're over-spending on R&D to grow more quickly than their free cash flow otherwise allows to stay ahead of competition and dominate the market. Indeed if investors agree and your business is actually strong, this is a strong play because you're leveraging some future value into today's growth.
Platform economics "works" in theory only upto a point. Its super inefficient if you zoom out and look not at system level but ecosystem level. It hasn't lasted long enough to hit failure cases. Just wait a few years.
As to openai, given deepseek and the fact lot of use cases dont even need real time inference its not obvious this story will end well.
I also can't see it ending well for OpenAI. This seems like it's going to be a commodity market with a race to the bottom on pricing. I read that NVIDIA has a roughly 1000% (10x) profit margin on H100's, which means that someone like Google making their own TPUs has a massive cost advantage.
Moore's law seems to be against them too... hardware getting more powerful, small models getting more powerful... Not at all obvious that companies will need to rely on cloud models vs running locally (licencing models from whoever wants that market). Also, a lot of corporate use probably isn't that time critical, and can afford to run slower and cheaper.
Of course the US government could choose to wreck free-market economics by mandating powerful models to be run in "secure" cloud environments, but unless other countries did same that might put US at competitive price disadvantage.
They do get a lot of customers buying their stuff, but on top of that, a company with unique IP and mindshare can get investors to open their wallet easily enough; I keep thinking of AMD that was not or barely profitable for like 15 years in a row.
Depends on your definition of profitability, They are not recovering R&D and training costs, but they (and MS) are recouping inference costs from user subscription and API revenue with a healthy operating margin.
Today they will not survive if they stop investing in R&D, but they do have to slow down at some point. It looks like they and other big players are betting on a moat they hope to build with the $100B DCs and ASICs that open weight models or others cannot compete with.
This will be either because training will be too expensive (few entities have the budget for $10B+ on training and no need to monetize it) and even those kind of models where available may be impossible to run inference with off the shelf GPUs, i.e. these models can only run on ASICS, which only large players will have access to[1].
In this scenario corporations will have to pay them the money for the best models, when that happens OpenAI can slow down R&D and become profitable with capex considered.
[1] This is natural progression in a compute bottle-necked sector, we saw a similar evolution from CPU to ASICS and GPU in the crypto few years ago. It is slightly distorted comparison due to the switch from PoW to PoS and intentional design for GPU for some coins, even then you needed DC scale operations in a cheap power location to be profitable.