Great post! I like to think there is a Hofstadter's law equivalent of TAM expansion in AI services: "it'll be a larger TAM than you expect, even if you've taken into account this law" :)
I would like to think that AI services will accelerate the transition from on-prem to cloud, especially if we get AI agents to a point where languages like COBOL can easily be translated to more modern languages
Great article, Kevin. While overall thesis makes sense, 1) how about the open-source AI ecosystem disrupting the market for commercial inference services? 2) What if some inference providers develop end-user applications or industry-specific solutions? - this would change their value prop and potential TAM.
Kevin, great article! I am wondering if we can translate your blog into Chinese and post it on AI community. We will keep the original link and state where it is translated from. Thank you!
Coming to this late but it's aged remarkably well. The convergence thesis on price and performance turned out to be spot on.
What I didn't expect to see was flat subscription proxies entering the picture. I've been routing my coding agents through one for a few weeks now and it completely changes the maths. $30/month flat, entire model library, no per-token charges. I put the rate limit comparison together here https://reading.sh/how-to-get-3x-claude-rate-limits-for-30-a-month-1d3fdb8658df if you're curious how the numbers stack up against Claude Pro and Max.
Do you reckon flat-rate models are another vector of commoditisation, or more of a niche that gets squeezed once the bigger inference players start bundling?
What other offerings do you think inference providers can bundle in for revenue expansion? I get the full-stack play but finetuning, monitoring and data management seem like obvious next steps.
Great article! Just help me understand where companies like Groq and Crebras fit in your framework. Is it "Compute Substrate' layer, where they fit in?
Great post! I like to think there is a Hofstadter's law equivalent of TAM expansion in AI services: "it'll be a larger TAM than you expect, even if you've taken into account this law" :)
c.f. we still haven't scratched the surface of moving on-prem services to the cloud: https://x.com/umang/status/1811270475365994520
I would like to think that AI services will accelerate the transition from on-prem to cloud, especially if we get AI agents to a point where languages like COBOL can easily be translated to more modern languages
100%
Great article, Kevin. While overall thesis makes sense, 1) how about the open-source AI ecosystem disrupting the market for commercial inference services? 2) What if some inference providers develop end-user applications or industry-specific solutions? - this would change their value prop and potential TAM.
Nice blog Kevin!
Great analysis.
Kevin, great article! I am wondering if we can translate your blog into Chinese and post it on AI community. We will keep the original link and state where it is translated from. Thank you!
Just sent you a message!
Coming to this late but it's aged remarkably well. The convergence thesis on price and performance turned out to be spot on.
What I didn't expect to see was flat subscription proxies entering the picture. I've been routing my coding agents through one for a few weeks now and it completely changes the maths. $30/month flat, entire model library, no per-token charges. I put the rate limit comparison together here https://reading.sh/how-to-get-3x-claude-rate-limits-for-30-a-month-1d3fdb8658df if you're curious how the numbers stack up against Claude Pro and Max.
Do you reckon flat-rate models are another vector of commoditisation, or more of a niche that gets squeezed once the bigger inference players start bundling?
Loved this read
What other offerings do you think inference providers can bundle in for revenue expansion? I get the full-stack play but finetuning, monitoring and data management seem like obvious next steps.
What do you think?
Great article! Just help me understand where companies like Groq and Crebras fit in your framework. Is it "Compute Substrate' layer, where they fit in?