Forgot your password?
Close
typodupeerror
Businesses

OpenAI Losses Increased Nearly 8X In 2025, With Spending Hitting $34 Billion (wheresyoured.at) 92

Posted by BeauHD from the behind-the-scenes dept.
An anonymous reader quotes a report from independent journalist Ed Zitron: Today, I can exclusively report, based on audited financial documents viewed by this publication that have been independently verified by the Financial Times, that OpenAI lost around $38.5 billion in 2025, as well as other crucial details about the financial condition of the company. [...] At the end of the year, OpenAI had just over $50 billion in assets, with almost half of that in cash. [...] The financial condition of OpenAI is deeply concerning. $38.53 billion in losses are astronomical, and far higher than most believed it would be. Losses also appear to be mounting year-over-year at a dramatic rate, and I'm not sure how this company finds a way toward any kind of sustainability or profitability. As discussed, I have not editorialized much today. I believe the best thing I can do for the general public is to deliver this news as plainly as possible. Ars Technica's Kyle Orland offers a more editorial take, writing: All told, OpenAI's day-to-day "loss from operations" increased from $8.78 billion in 2024 to $20.92 billion in 2025, a concerning direction for a company that is telling investors it hopes to be profitable by 2030. But measured as a percentage of revenues, the company's operating losses slightly improved year to year, from 237 percent in 2024 to 160 percent in 2025.

Operating numbers aside, OpenAI's headline "net loss" number of just over $5 billion in 2024 ballooned to nearly $39 billion in 2025. But the 2025 number includes a significant accounting charge related to investor valuations that shifted amid the company's 2025 conversion to a for-profit structure. The Financial Times cites "a person familiar with the matter" in reporting that this non-recurring charge was approximately $30 billion and that OpenAI's 2025 net loss amounted to a more reasonable-looking $8 billion without it.
This discussion has been archived. No new comments can be posted.

OpenAI Losses Increased Nearly 8X In 2025, With Spending Hitting $34 Billion More

OpenAI Losses Increased Nearly 8X In 2025, With Spending Hitting $34 Billion

Comments Filter:

  • The cost of force (Score:5, Interesting)

    by karmawarrior ( 311177 ) on Wednesday June 17, 2026 @11:19AM (#66197024) Journal

    So you have something that nobody is asking for (not in the way genAI is anyway), and you decide, of all things, rather than making a case for it, you force people to use it, in the hope they get addicted, think they can't do without it, and continue using it after you start pricing it at profitable levels.

    This is the business model. Why are they not making a case for it? Why are they, instead, pretending it's something it isn't? Because nobody would take it seriously if they did the latter. The only way they can get people to use it is via force, and that means persuading idiot CEOs with a FOMO issue, while pricing it well below cost.

    The question isn't "Will they make a profit", that's not something you or I should care about. Who gives a crap if a bunch of vulture capitalists get busted? The question is "How much damage will they do with this particular con job".

    The end game of this, remember, is to get companies dependent upon genAI companies. To make them unable to function any more without handing over control of their systems to the genAI people. And idiot FOMO CEOs who have gotten the dopamine rush from using genAI tools are making sure their companies will be run that way, despite the obvious dangers.

    So the answer to "How much damage" is, so far, a crazy amount. So far. There are now many, many, companies that have lost control and knowledge of how their own businesses run. And it's getting worse.

    • Re: (Score:2, Insightful)

      by thegarbz ( 1787294 )

      So you have something that nobody is asking for (not in the way genAI is anyway), and you decide, of all things, rather than making a case for it, you force people to use it, in the hope they get addicted, think they can't do without it, and continue using it after you start pricing it at profitable levels.

      You make it sound dumb but the reality is this model has worked for many products and services before. It's why a lot of things start free. The entire free trial is based on it.

      Heck the ubiquitous "Post-it" note was created with precisely this strategy after the team developing adhesives internal to 3M were asked to stop research on their "failed" adhesive. It ticks all the boxes you just asked "why" about. They made a case, it was rejected. They pretended it is something different. No one took their produc

      • Re: (Score:3)

        by 0123456 ( 636235 )

        We're not talking about going from free to $10 a month. We're talking about business fees going from $1,000 a month to $10,000+ a month or more.

        That completely changes the business model and eliminates the benefits of using AI over humans in many roles.

        It's hard to see how companies can ever be convinced to pay the true cost of the AI services they use without the AI service waiting a few years for those companies to sack everyone who used to do the work and then suddenly ramping up the costs 10x or more wh

        • It looks a lot like they will require the VMware style Broadcom pricing increase on steroids. Seeing how well the broadcom thing is going down, I can't image how hard to swallow the AI price increase will be.

        • Re: (Score:2)

          by Zak3056 ( 69287 )

          My personal favorite example of this is OpenAI's stated plan to have $1T per year in infrastructure spending. If you do the math, you will have to replace approximately 1/3rd of the entire productive US workforce and charge their former employers about $30k a year per displaced employee to break even. On the infrastructure. OPEX not included.

          The math doesn't math.

        • We're not talking about going from free to $10 a month. We're talking about business fees going from $1,000 a month to $10,000+ a month or more.

          That completely changes the business model and eliminates the benefits of using AI over humans in many roles.

          Many things in history turned out to not have a functioning business model when all costs were included. You just don't remember them because they died out. That doesn't change the foundation of the business strategy. The AI strategy was banking on scaleup not just of their customer base, but the product itself.

          For all the shit we heap on AI, in a year it's come an insanely long way, maybe not long enough, but that doesn't change the viability of their expansion strategy. In AI especially economies of scale

    • Re: (Score:2)

      by klui ( 457783 )

      Look what's happening to Facebook when they went all in on "AI."

      https://newsletter.pragmaticen... [pragmaticengineer.com]

    • Re: (Score:2)

      by lsllll ( 830002 )

      Whether it's a con job or not is irrelevant to me. While I've found ChatGPT to be completely mediocre when it comes to technical stuff, I have poured hours into discussing the humanities with it with astounding results, at least for myself, and I would hate to see the knowledge it has of our conversations vanish into thin air (or rather become unusable for me and become only a commodity for whoever buys them out). It would be a lot of time lost on my side and it would suck if I have to start on square one

      • How much are you willing to pay for your AI girlfriend?

        • Re: (Score:2)

          by lsllll ( 830002 )

          Funny how I started referring to it as "he" and I'm completely heterosexual. But to answer your question, I'm almost to the point of jumping in at their lowest tier, which I think is $20/month. It's supposed to be my father's day present to myself.

      • Agreed. Another instance of finding myself shaking my head a bit at the prevailing Slashdot negativity on AI. I don't even need to try having it create/maintain large or complex codebases to know current LLMs are going to choke on such a task. As a programmer though I still find it astounding in it's capabilities, including taking rambling natural language prompts and generating impressively functional chunks of code or scripts. I guarantee 90% of Slashdotters know the size and intended usage of a technical

        • Re: (Score:2)

          by lsllll ( 830002 )

          Case in point, I was working on a project where I needed to do .csv file processing in a BASH script and I knew AWK was the perfect tool to do it, but I had never written an AWK script. So I went through a couple of iterations with ChatGPT, asking it to create me a script and then I modified much of it to work in the way I wanted it to work for my process. Without ChatGPT, it would definitely have taken more time to write something from scratch and I may not even have been inclined to use AWK had I not ha

      • It only keeps its "memory" as long as the session is open.
        As soon as you close it, everything is lost.

        And: it is only tokens, no knowledge in the LLM.

        • Re: (Score:2)

          by lsllll ( 830002 )

          That's if you don't use an account when you use it. I use an account associated with it. Recently, as an exercise to see what it had learned about me, I asked it to write an obituary for me. It was spot on and referred to conversations I had with it from last year.

          • I did the same with google Gemini a few days ago: it told me it can not store anything about the session.

            But I am in Europe.

        • Codex CLI stores full conversations locally. So does Claude code. It doesn't mean the LLM will remember - it has a context limit. But it is there for you to search. It is PITA to migrate all this to another system, though. There is no officially supported way. LLMs were able to create scripts to move my side projects conversations from a WSL Ubuntu VM to a more manageable lxc container on my proxmox server, though.

          • Yes, I guess you need a local client for that. I do not know if it would work via a web site.
            Point is, the LLM does not "learn", but the hoster might store your interaction and use it to train the next generation.

            • It very much does learn even within its local session. For example, I have been teaching it to ssh into my raspberry pi hooked up to the carrier hvac bus via rs485 for the last few months for reverse engineering purposes. I iterate every week or so with new tracesx making it updtae the code, and run the nee version. The additional prompts I type eerkpy are very short, as it has all the conversation history. Even pre migration from wsl to lxc. Sometimes it does forget certain detailsx like any LLM. I do not

              • That is not what "learning" means.

                You have a local session, and over time you have configured it to understand and do what you want to do in that context.

                Close the session and try again, and it is back where it was before.

                However you likely remember its misunderstandings and formulate your new requests different and avoid its pitfalls, and hence you have the impression it learned.

                What LLM are you using?

                • It is chatgpt codex with gpt 5.5 cli.
                  Its reasoning is pretty good. Makes far fewer mistakes than the corresponding AI chatbot, which causes me a lot of facepalms.

                  • Interesting, good to know.

                    • Re: (Score:2)

                      by madbrain ( 11432 )

                      Yes. These agents/LLMs really are optimized for coding, especially Python, which I have been using for my pet projects.

                      When it comes to chat bot and non-coding topic, it's a lot less accurate. If you are an expert on some topic and discuss it with the bot, you will quickly see that it falls short.

                    • If you are an expert on some topic and discuss it with the bot, you will quickly see that it falls short.
                      Yeah, I had a chat with Claude I think. About a toy language of mine, that incorporates pre and post conditions for method calls, like in Eiffel.
                      It was kind of confused weather you can harden the precondition or the postcondition.
                      After I pointed out he is wrong, he first quickly assured I am right, and surprisingly jumped back on trail and supported my correction, correctly.

                      Bottom line they are simply in

    • Re: (Score:2, Troll)

      by fatwilbur ( 1098563 )
      "Nobody is asking for" - care to qualify what you mean there? A quick search says ChatGPT has 50 million paying subscribers and a billion (!) monthly active users. I think there's interesting debate on the ultimate potential profitability of it given the infrastructure and power costs. However it is simply detached from reality to think a service that hundreds upon hundreds of millions of people worldwide use on a monthly basis does not constitute demand. And that's just a single provider.

      My guess here i

  • Good Luck With the IPO (Score:5, Funny)

    by 0xG ( 712423 ) on Wednesday June 17, 2026 @11:21AM (#66197026)

    Good luck with the IPO, this might put a damper on things.

    • Signed integer (Score:5, Funny)

      by goombah99 ( 560566 ) on Wednesday June 17, 2026 @11:47AM (#66197080)

      well once they crossed 2 billion the signed integer representation made losses into profits

      • Re: (Score:2)

        by bn-7bc ( 909819 )
        only if you use 32 bit integers, come one it's 2026 don't be stingy with the bits, allso I would trust absolutely no financial sw the operated with integers (unless ofc they scaled every value by 100 to account for cents) or you'd run into a lot of rounding errors. and no I'm not advocating for floats either they have their own issues, which is why we have specific types of fixed point decimals (that use wya more than 32 bits in these systems)... Oh you where making a joke sorry for taking you seriously

        • Re: (Score:2)

          by BKX ( 5066 )

          Fixed-point decimals and scaled integers are fundamentally the same thing. The only real difference is that decimal data types use base-10 at the representation level (aka binary-coded decimal, or even straight ASCII numerals. You can do math directly on both using techniques that are practically identical, save a few differing constants.). Of course, I doubt you meant actual decimal data types. Fixed-point binary and scaled integers are exactly the same thing.

          Historically, money was computed using only BCD

  • Oh, they're loosing *so* much less money this year than last... I find it mind-boggling that anyone in a back office hasn't attacked the CEO who wants to spend money on this, or had them committed.

    • Where the LLMs shine is they all know that "loosing" is actually "losing." Not only that, but they all know that the 3rd person possessive is "its" and not "it's." They are good for certain things.

  • Normally, I'm not one to root for a company to go under, but I'll make an exception for this one. I hope their losses keep going up exponentially!
  • So, will Sam make it to IPO and cash out time before it all goes titsup ?

    Let us all, on behalf of Open Source software and Non Profit companies everywhere, hope not.

  • Spending 3X your revenue (Score:5, Informative)

    by OrangeTide ( 124937 ) on Wednesday June 17, 2026 @11:39AM (#66197070) Homepage Journal

    Spending more than your revenue with no answer to when it turns around is a recipe for a bubble.
    Yet OpenAI is not the worst offender by a long shot.
    ref: Is AI Profitable Yet? [isaiprofitable.com]

    • The question in addition to "Are they profitable yet?" is "Do they have a path to profitability?" Amazon was unprofitable for the nine years of its existence, but it built the technical infrastructure and used ruthless marketing to emerge into profitability. That's questionable for any of the hyperscalers, the new AI model companies, and the data center operators. I'd like that "Are they profitable yet?" site to address the "Do they have a path to profitability?" issue, based on an financial analysis.

      • Re: (Score:2)

        by allo ( 1728082 )

        People were laughing when someone told them Facebook could ever make money.

        The companies are planning for more than two years and the investors know that. And currently, the investors seem to believe that the companies will make it, as they unfortunately didn't read the analyses in the Slashdot comments.

        • Re: (Score:2)

          by dfghjk ( 711126 )

          "People were laughing when someone told them Facebook could ever make money."
          Citation please. Facebook was a copycat, and was of course stolen as well. Also, "make money" is probably not a good way to put it.

          "The companies are planning for more than two years and the investors know that."
          Thieves plan too, doesn't mean they are respectable.

    • Re: (Score:2)

      by bn-7bc ( 909819 )
      Oh AI iksvery profitable, but not for the people you think, Just ask manufacturers of storrage an ram + amd, Intel and Nvidia. Oh I nearly foget CISCO and the contractors building the dcs and the power companies selling GWHs worth of power to the madness they all love their AI related profitts.
  • Sam Altman used to run Y Combinator, that makes him supreme king of the Startup Bros, he knows what he's doing!

  • Zitron misses the bigger picture (Score:5, Funny)

    by zmollusc ( 763634 ) on Wednesday June 17, 2026 @12:04PM (#66197116)

    Ed Zitron completely fails to take into account the four factors which will inevitably lead to massive profitability for the AI firms. These are:
    1 Stuff
    2 Things
    3 Misc
    4 Other

    This is not just my view, it is the informed opinion of experts in the AI field.

  • OpenAI spent more than $10 billion on training its AI models? That's more than the annual tuition for over 380,000 college students (I have not estimated how many Olympic swimming pools full of cash this is, but I'll leave that to the reader).

    Is there some expectation that this training will be mostly *done* at some point, or at least significantly done so they can reduce spending on model training and recoup this investment through future subscriptions? Or is this $10 billion + a likely annual cost of ke
    • LLMs need a constant flow of new training data; otherwise, the models suffer from well known problems: - Knowledge staleness - Data distribution shift sources: - https://apxml.com/courses/how-... [apxml.com] - https://arxiv.org/pdf/2307.090... [arxiv.org]

      • And it's going to get worse. Much of what they've trained their models on is user created information from a lot of sites that are now seeing sharp dropoffs in traffic because people aren't going there for info/answers any more, so there's less incentive to post new information and many of those sites will just go completely stale.

        A perfect example is projectors. 5 years ago if you wanted a projector for a price range, you'd go to projectorcentral.com and a few other sites for detailed information. Now m

        • That is not how it works.
          LLMs are not trained on thousands of projector models, how do you come to that idea?

          An LLM that is used as a search engine will use Retrieval Augmented Generation: RAG.

          That means it gathers the relevant information at the moment you start your query. /FACEPALM

          If it hits projectorcentral.com or other web sites it considers more relevant, I can not tell.

          • Re: (Score:2)

            by dfghjk ( 711126 )

            A more appropriate answer to the original question is that training hasn't even started yet on tomorrow's relevant AI networks. Today's LLMs are psychopaths that need to be taken out back and shot. Training of today's models are of no use when those models are superseded.

            • Well,
              people train new models with their own private data.
              They want to use them tomorrow. Not in 5 years when we have new kinds of models.

              Training of today's models are of no use when those models are superseded.
              Let's say they are to costly, no use is exaggerated. They still cand do the same things they can do right now.

          • "That is not how it works.
            LLMs are not trained on thousands of projector models, how do you come to that idea?"

            Where in the hell did I say it was 'trained on thousands of projector models'? I implied it was trained on the ARTICLES at Projector Central. You know, the ones have relevant review data in them that the search engines also scrape that your LLM also relies on? The ones that are no longer being written at the same rate because their views and ad revenue are down, the ones that will eventually sto

            • Yes, you said that.
              And that makes no sense at all.

              The LLM is not trained on that data.

              What would be the purpose of that?

              You never could it ask a relevant question ... this is not training this farking DATA.

              The LLM is trained on what Projectors are - in a sense.

              And when you ask it: it searches the web for stuff that is interesting based on your question. It is trained to answer your question, not to memorize 1000ds of projector details.

              For that you would not need an LLM, but a database would be enough.

              Seriou

              • Re: (Score:2)

                by unrtst ( 777550 )

                The LLM is not trained on that data.

                You have a valid point about RAG usage, but it almost certainly was trained on that data.

                FYI, I tried the question GP posted on Google with AI results on: "I have $1000 to spend on a projector, what should I get"
                The answer included reference to one URL and 2 products:
                https://www.rtings.com/project... [rtings.com]
                A Google product link for Epson Lifestudio Flex Plus 4K PRO-UHD Projector
                A Google product link for ViewSonic PX701-4K Projector

                If/when those review sites stop posting such lists, where will the RAG get its resul

                • I guess I misunderstood the original point.

                  As in what a user would ask and what the answer is.

                  You are right, for answers like that the LLM would be trained on the review data.

                  A RAG will get its result from google searches and analyzing the pages the result point to ... actually there are special search APIs "to search the internet".

                  RAG basically only means: "find me the cheapest flight from LA to BKK", and figure how to get the data for the cheapest flight for the situation right now.

                  How to find a cheap fli

      • Unless LLM companies become strong enough to prevent new data from entering the public domain. When all the *.ai noise began I saw this constraint as a business requirement. Can the big.data companies force it ... as some sort of "national security issue" ? I believe once the limitations/expense  of *.ai become clear those companies will play the "security" card:  as brazen as ' think of the children' type arguments against on line privacy.
  • Since nobody wants your garbage products, take all that RAM you just bought and sell the modules back to DRAM makers. You'd make a killing!
  • They are marketing themselves as AI companies and all the have to sell are Massive Automation Machines sold as AGI.
    I am in SpaceX they have a future(I think). But I am staying clear of all the AI companies when they go IPO.
    The AI companies(marketing automation) I see so far are loaded with old and new debt which I don't see an ability to pay back at this point.
    And even if they can cover their debt, I wonder if sucking that much money out of their user base will be bad for their customers and the economy ove

    • Re: (Score:2)

      by allo ( 1728082 )

      Does any of them claim they would have AGI yet? Many talk a lot about achieving AGI (they probably won't), but I don' think they claim their AIs were more than language models.

      • Yea, I should have typed AI not AGI! me bad.
        But I do think a majority of people on the street believe they have AGI already due to their sales and marketing.
        But yes, they are careful to stay on the barely legal side of things with their official comments.

        • Re: (Score:2)

          by allo ( 1728082 )

          Let's say it the other way round: They have a pretty useful product, but they market it as something different what is only hype. People rather believe that the hype would be useful than the working tool. A LLM sounds boring, AGI/ASI is exciting. But in the end I'd rather have a good tool than a bot to debate with if it deserves labor rights like other sentient beings. :D

    • I'd be cautious on spacex. They were touting their AI in the IPO. spacex is severely over valued at this moment.

    • I think the sooner people realize/admit LLMs are not AI and have little to do with AI, the better we'll all be.

      LLMs are stochastic token generators. While machine learning might be used to train them, the LLMs do not actually perform anything related to AI. Claiming LLMs are AI powered is like claiming my car is AI powered because machine learning was used when designing for air flow.

  • They will make up the losses per transaction in volume.

Slashdot Top Deals

The party adjourned to a hot tub, yes. Fully clothed, I might add. -- IBM employee, testifying in California State Supreme Court

Close