• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Could OpenAI and Anthropic keep their best models private? Posts question the economics

    One post argues that Anthropic’s enterprise application revenue is unlikely to replace API sales anytime soon—a central concern in the discussion about withholding its best models.

    X(
    RP
    AA
    10 Sources, ,

    TLDR

    Posts discussing whether OpenAI and Anthropic might keep their best models to themselves focus on the revenue trade-off. One questions how they would maintain model sales revenue, arguing that Anthropic’s enterprise applications are unlikely to replace API sales anytime soon. A related comment argues that owning enough businesses to replace model revenue while continuing to grow would be very difficult.

    Combined views

    106.9K

    10 Sources, first seen 23d ago

    Combined views

    106.9K

    10 Sources, first seen 23d ago

    506 likes
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    23d ago
    first seen 23d ago
    506 likes
    59 comments
    93 saves
    58 reposts
    59 comments
    93 saves
    58 reposts

    Sentiment

    Positive48.9%51.1%Negative

    Summary

    Sentiment

    Positive48.9%51.1%Negative

    Positive accounts defended AI labs' model sales revenues against profitability doubts, while negative replies countered that labs still burn cash without profits and questioned Zuckerberg's business ownership claims.

    Based on 48 sentiment-bearing replies from 45 accounts across 3 conversations.

    Summary

    Positive accounts defended AI labs' model sales revenues against profitability doubts, while negative replies countered that labs still burn cash without profits and questioned Zuckerberg's business ownership claims.

    Based on 48 sentiment-bearing replies from 45 accounts across 3 conversations.

    10 Sources

    @anton_d_leichtLabs would love to keep their best models to themselves: makes them a lot harder to catch and less politically exposed. But it’s unstable because anyone can defect and win the outside market with a frontier model. At least as long as USG doesn’t accidentally stop defectors! There's an obvious upside to restriction, and I suspect we're partly seeing its effects already. Labs don't like fast-followers that eat at their margins, and they don't want to cause huge scandals, so they're incentivised to limit access to their models. Restricting broad access to leading models hamstrings fast-followers, especially those that distil. It also reduces the most obvious misuse-related model risks, especially as we move toward even more capable cyber models and potentially dangerous bio models. It also helps avoid harsh policy intervention. For the time being, labs that just want to push the frontier unencumbered might think that internal-only deployments keep them largely out of policymakers' acute attention. Hugging Face has changed that a little bit, but I'm not sure it was as legible to DC as Mythos was. The revenue case, at least in the medium term, seems pretty stable to me: internal deployments are very useful for accelerating R&D, labs are quickly moving toward very sophisticated vertical integrations. And you can still roll out limited-access programs under a restricted framing: you probably get a large part of frontier revenue simply by offering BioMythos to a handful of pharma firms, no actual release necessary. As long as all frontier labs do this, they can all just drip-feed distilled versions of their internal frontier models to stay visibly ahead and deploy the rest internally. Drip-fed versions would still be ahead of open-source alternatives far enough to get a slice of outside revenue. But as soon as any one lab defects and releases a true-frontier model, it could plausibly monopolise a large share of the outside market, and enable faster outside catch-up at the same time. Defecting like that is very costly. It exposes the defector specifically to a lot of regulatory and political risk all at once: any misuse incident reflects poorly specifically on the one open frontier developer, and they'd also become the prime target both for distillation attempts and for the US government's displeasure with inviting them. It'll be interesting to see how that plays out in the specific context of the current frontier labs. It's easy to see how Anthropic would bias toward locking down access and vertically integrating - because of their belief in RSI, their vertical integration, and the fact that they're least likely to survive political risk. In that scenario, OpenAI has conflicting incentives: do they take a risk and undercut, or accept the two-player oligopoly? Meta and xAI seem much less likely to deploy only internally. It's an open question whether they'll fully catch up at any point, but they'll likely release something good whenever they have it. (I also have no clue what GDM is up to or what they'd do). In that case, the two frontier labs face a tricky choice between frontrunning their releases, or tolerating the impression that they're not as far ahead as they think. National security concerns and USG pressure could actually serve as a coordinating function, holding back defectors due to misuse concerns. That, again, is very hard to predict: you could imagine USG letting xAI or Meta, but not Anthropic, release a model of the same capability level. That political volatility makes me think most purely economically derived arguments on this are a bit overconfident. The 'solution', if there is one, is the same as always: the same industry framework for everyone, assured access frameworks for the rest of the world, much more policymaker awareness of the risks of internal deployments. I'm not sure we're getting that in time, so the market dynamics until then will be very interesting to watch.
    @krishnanrohitAnthropic makes a 100 billion from its model sales, more than their revenues of any pharma company. Salesforce makes half that. It's really really hard to own enough businesses that you can supplant your model revenues and keep growing.
    @xuanalogueRT @anton_d_leicht: Labs would love to keep their best models to themselves: makes them a lot harder to catch and less politically exposed.…
    @rohanpaul_aiOpen-weight models may undermine AI lab revenue while continuing to feed cloud demand. OpenRouter’s data suggests proprietary models’ share of routed queries fell from roughly 60% to 25% within months. Its routing system directs each request to the lowest-cost model expected to produce an adequate answer, letting open models gain usage without customers selecting them manually. e.g. AT&T is moving a rapidly growing share of its AI work from proprietary models to cheaper open models, and reports cost savings of up to 80%. If other large companies follow, OpenAI and Anthropic could lose much of the routine enterprise usage they currently charge for. FT published a piece
    @alexatallahThis is an insight from our State of AI report last year that remains true today: https://openrouter.ai/state-of-ai
    @thdxralso every time i talk about how it's hard to host models at certain prices people keep telling me how dumb i am and linking this provider came out today that provider was wrapping openrouter and routing requests to cheaper models
    @lessinZero-Day models to hold banks and countries for ransom is the only busness model that makes sense... everything else is just commodity https://x.com/i/article/2099619883998965760
    @svpinoRT @svpino: Open-weight models are right behind frontier models in capabilities. I think they have an advantage, though: The more Frontie…

    10 Sources

    @anton_d_leichtLabs would love to keep their best models to themselves: makes them a lot harder to catch and less politically exposed. But it’s unstable because anyone can defect and win the outside market with a frontier model. At least as long as USG doesn’t accidentally stop defectors! There's an obvious upside to restriction, and I suspect we're partly seeing its effects already. Labs don't like fast-followers that eat at their margins, and they don't want to cause huge scandals, so they're incentivised to limit access to their models. Restricting broad access to leading models hamstrings fast-followers, especially those that distil. It also reduces the most obvious misuse-related model risks, especially as we move toward even more capable cyber models and potentially dangerous bio models. It also helps avoid harsh policy intervention. For the time being, labs that just want to push the frontier unencumbered might think that internal-only deployments keep them largely out of policymakers' acute attention. Hugging Face has changed that a little bit, but I'm not sure it was as legible to DC as Mythos was. The revenue case, at least in the medium term, seems pretty stable to me: internal deployments are very useful for accelerating R&D, labs are quickly moving toward very sophisticated vertical integrations. And you can still roll out limited-access programs under a restricted framing: you probably get a large part of frontier revenue simply by offering BioMythos to a handful of pharma firms, no actual release necessary. As long as all frontier labs do this, they can all just drip-feed distilled versions of their internal frontier models to stay visibly ahead and deploy the rest internally. Drip-fed versions would still be ahead of open-source alternatives far enough to get a slice of outside revenue. But as soon as any one lab defects and releases a true-frontier model, it could plausibly monopolise a large share of the outside market, and enable faster outside catch-up at the same time. Defecting like that is very costly. It exposes the defector specifically to a lot of regulatory and political risk all at once: any misuse incident reflects poorly specifically on the one open frontier developer, and they'd also become the prime target both for distillation attempts and for the US government's displeasure with inviting them. It'll be interesting to see how that plays out in the specific context of the current frontier labs. It's easy to see how Anthropic would bias toward locking down access and vertically integrating - because of their belief in RSI, their vertical integration, and the fact that they're least likely to survive political risk. In that scenario, OpenAI has conflicting incentives: do they take a risk and undercut, or accept the two-player oligopoly? Meta and xAI seem much less likely to deploy only internally. It's an open question whether they'll fully catch up at any point, but they'll likely release something good whenever they have it. (I also have no clue what GDM is up to or what they'd do). In that case, the two frontier labs face a tricky choice between frontrunning their releases, or tolerating the impression that they're not as far ahead as they think. National security concerns and USG pressure could actually serve as a coordinating function, holding back defectors due to misuse concerns. That, again, is very hard to predict: you could imagine USG letting xAI or Meta, but not Anthropic, release a model of the same capability level. That political volatility makes me think most purely economically derived arguments on this are a bit overconfident. The 'solution', if there is one, is the same as always: the same industry framework for everyone, assured access frameworks for the rest of the world, much more policymaker awareness of the risks of internal deployments. I'm not sure we're getting that in time, so the market dynamics until then will be very interesting to watch.
    @krishnanrohitAnthropic makes a 100 billion from its model sales, more than their revenues of any pharma company. Salesforce makes half that. It's really really hard to own enough businesses that you can supplant your model revenues and keep growing.
    @xuanalogueRT @anton_d_leicht: Labs would love to keep their best models to themselves: makes them a lot harder to catch and less politically exposed.…
    @rohanpaul_aiOpen-weight models may undermine AI lab revenue while continuing to feed cloud demand. OpenRouter’s data suggests proprietary models’ share of routed queries fell from roughly 60% to 25% within months. Its routing system directs each request to the lowest-cost model expected to produce an adequate answer, letting open models gain usage without customers selecting them manually. e.g. AT&T is moving a rapidly growing share of its AI work from proprietary models to cheaper open models, and reports cost savings of up to 80%. If other large companies follow, OpenAI and Anthropic could lose much of the routine enterprise usage they currently charge for. FT published a piece
    @alexatallahThis is an insight from our State of AI report last year that remains true today: https://openrouter.ai/state-of-ai
    @thdxralso every time i talk about how it's hard to host models at certain prices people keep telling me how dumb i am and linking this provider came out today that provider was wrapping openrouter and routing requests to cheaper models
    @lessinZero-Day models to hold banks and countries for ransom is the only busness model that makes sense... everything else is just commodity https://x.com/i/article/2099619883998965760
    @svpinoRT @svpino: Open-weight models are right behind frontier models in capabilities. I think they have an advantage, though: The more Frontie…