The capability spread between closed source and open source AI is only part of the story; the monetizable spread matters as much, and no one is paying attention to it
Excellent analysis; I like how these articles focus on topics and questions in the AI space that rarely arise in mainstream discussions, and with a level of intellectual thoroughness that it is almost never seen.
There are substantial economies of scale in serving ai though, open source or closed. We should expect the largest ai companies to also have the best margins on open source tokens. So regardless, the big gpu owners win, which at this point includes anthropic and OpenAI. But more importantly this analysis does not give enough credit to the data flywheel (Google is “just a text box” as well after all) and the vast economic space that lies above current model capabilities still — all of knowledge work, as a start (40% of gdp). Can open source keep up as the training data shifts from public internet tokens to complex proprietary knowledge workflows?
I agree that the data flywheel deserves more attention, and I'll probably write about it separately. But as far as the Google comparison goes: its moat was its search index, not the text box. The frontier labs need to figure out what their equivalently irreplicable asset is. Proprietary workflows is a future moat that hasn't formed yet, and the valuation question I discuss in this post is about the next 2-3 years, not the next 10.
Frontier tasks (agentic reasoning and complex workflow) is not significant in total pie today but will become the major growing tasks in the near future. The real competition is there.
Great analysis and framing, very interesting to read.
For the model capability levels that have been commoditized, what do you think about the opportunity to generate ad revenue? I think you would find my “Why AI Chatbots Can Be the Next Great Ad Platform" interesting, and I’d love to hear your pushback on the points I’m making in it. https://csuiteofone.substack.com/p/why-ai-chatbots-can-be-the-next-giant
Excellent analysis; I like how these articles focus on topics and questions in the AI space that rarely arise in mainstream discussions, and with a level of intellectual thoroughness that it is almost never seen.
There are substantial economies of scale in serving ai though, open source or closed. We should expect the largest ai companies to also have the best margins on open source tokens. So regardless, the big gpu owners win, which at this point includes anthropic and OpenAI. But more importantly this analysis does not give enough credit to the data flywheel (Google is “just a text box” as well after all) and the vast economic space that lies above current model capabilities still — all of knowledge work, as a start (40% of gdp). Can open source keep up as the training data shifts from public internet tokens to complex proprietary knowledge workflows?
I agree that the data flywheel deserves more attention, and I'll probably write about it separately. But as far as the Google comparison goes: its moat was its search index, not the text box. The frontier labs need to figure out what their equivalently irreplicable asset is. Proprietary workflows is a future moat that hasn't formed yet, and the valuation question I discuss in this post is about the next 2-3 years, not the next 10.
This shows why capability isn’t the same as paying power in AI
Frontier tasks (agentic reasoning and complex workflow) is not significant in total pie today but will become the major growing tasks in the near future. The real competition is there.
Great analysis and framing, very interesting to read.
For the model capability levels that have been commoditized, what do you think about the opportunity to generate ad revenue? I think you would find my “Why AI Chatbots Can Be the Next Great Ad Platform" interesting, and I’d love to hear your pushback on the points I’m making in it. https://csuiteofone.substack.com/p/why-ai-chatbots-can-be-the-next-giant