Is there a business model for being profitable off training frontier open weights models?
Other people can host, fine-tune, consult etc. as least as cheaply as you can. There are no ancilary product sales & it is fantastically expensive to make compared to most open source work
Users agree there is a strong business case for training frontier open weights models because token costs remain extremely high.
No Digg Deeper questions have been answered for this story yet.
I know I keep harping on this theme but open weights models are quite valuable to AI users, and people seem to assume that they will always keep up as training costs grow, but I don’t really understand the incentives as opposed to open source, which was much clearer
Is there a business model for being profitable off training frontier open weights models?
Other people can host, fine-tune, consult etc. as least as cheaply as you can. There are no ancilary product sales & it is fantastically expensive to make compared to most open source work
Kind of a big deal that no one has been able to answer this question definitively?
(The only complimentary asset owner open weights makes sense for is Nvidia)
I know I keep harping on this theme but open weights models are quite valuable to AI users, and people seem to assume that they will always keep up as training costs grow, but I don’t really understand the incentives as opposed to open source, which was much clearer
@brandonsax52 People pay a lot of money for private benchmarks. and you can sell subscriptions for advanced benchmarks. and consulting.
Maybe Nvidia and Mistral collaborating on nemotron is a model. Maybe you share the cost of training and staffing, you get to control the model and Nvidia sells shovels for the gold rush and Mistral has a monopoly on consulting to a certain level sensitive EU partners?! In this case the economics of the model making might be similar to Android, Google only open sources it because it can make money off of the ecosystem.
GLM-5.2 suggests one model. Despite being available from several providers on OpenRouter, Z_ai has by far the greatest token share, possibly due to its high cache rate (meaning lower costs). So, one business model could be model makers using open weights to gain mindshare (and reduce fear of losing access, like with Fable) while using their knowledge of their own models to run the best API endpoints.
I suspect harnesses are going to be where it's at... and the open weight labs may just be able to carve out a defensible position there.
Anyone can use the model, but you have to subscribe with zAI to use it with their custom tuned harness for X, Y, or Z... It's better than you using it with OpenCode or Libra or something...
@emollick Meta has indicated they will likely go back to open sourcing again at some point in the foreseeable future - just not their most capable models.
@emollick selling an offering that's optimized for 10x inference speed or 10x guardrails' reliability. being upfront about it at model release would dissuade everyone else from building that, even though it would be technically possible for anyone to do it.
@emollick And if the US government has drawn a line in the sand at Fable-level capability, how can the frontier labs continue to afford to develop frontier models that won't be publicly available?
@emollick The Chinese have a strong incentive to produce and offer open weights models in order to commoditize their complement. The incentive need not be immediate profit from the models themselves.
@emollick Isn't Hugging Face already proving the model? They don't train frontier weights yeah but built a profitable platform
@emollick It can make sense for Nvidia and other hardware companies to train open weights models that are optimized for their hardware as a way to increase demand.
@emollick Training frontier open weights models shifts value toward proprietary datasets and specialized inference workflows rather than just raw compute access.
@emollick In principle you can crowd fund them! Society has ways of creating public goods: https://en.wikipedia.org/wiki/Public_good
As more hardware makers get good at running local models, more will be incentivized to open source. This isn't too far off the frontier. I'm certain GLM and Kimi have increased sales for Apples Mac Studios, and for Google Gemma is a no brainer for its edge devices. Maybe framework even?! https://youtu.be/N5xhOqlvRh4?t=872&is=LaInUli_oUY7ImDx
@emollick my best guess (tinfoil hat) was China gov funding OS model training since strong open models weakens the ability for US labs to raise capital, thus giving you an edge on AI race until youre good enough to close source your efforts.
@emollick The current open weights releases might be a temporary strategic move.
@emollick I think the model is have a profitable other business and provide inference for your free model that chokes out the closed model company. Meta was in a perfect place to do it until they flubbed llama 4. Google could do it, but don't seem to wanna
@emollick There are only two OS efforts.
China -> It's in the CCP's best interests to not let US labs runaway with it
Nemotron -> NVDA owns the compute layer. Makes perfect sense for both parties to go medal to the metal on OS
@emollick Teniendo la mejor inferencia y una pila de productos basados en el modelo, es lo que han hecho Kimi y los creadores de GLM. Por cierto, Zhipu fue la primera empresa AI que salió a bolsa y les ha ido bien