SemiAnalysis founder Dylan Patel says revenue-generating AI startups are shifting resources toward pre-training their own models

if you’re looking to do this we can help 🙏

@dylan522p Who's leading the rebound off the open — my rotation read → https://www.marketopenlist.com/t/market-v3-20260629-t27uu.html

i think we're maybe the only company in the world that credibly offers "pretraining as a service". we work with your team to scope compute + experiment ramp-up. we have a bunch of tooling for health checks, alerts, restarts, node swaps etc. we advise on arch + params if you want.

Well I’d like this to happen, but still very early at best.

@dylan522p People shouldn't worry about the US banning Chinese models. They should worry about the CCP banning them. If there is just one cyberattack against the CCP that was assisted by an open weight model, I'd bet all Chinese AI will go dark overnight.

Unless you extend to post-train/RL, only saw actual runs from a few industries (Finance, Cyber, Telco…)

@dylan522p If Ant doesn’t lower prices dramatically the enterprises will have an incentive to build their own models, too.

@Dorialexander re the first point this feels like a campaign. not sure if useful idiot or on purpose.

@dylan522p Any examples? I am only aware of Cognition and Cursor going this route really

@dylan522p Who's going to fund all these pre-trains?

@dylan522p Pretrainitus ends up with a Kimi Model in a Composer trenchcoat

@dylan522p Those of us concerned about the Chinese Open Source being banned by either the USA or China are not the same group of people building their own custom models for their companies.

@dylan522p Yes correct and at the end it is net positive for the AI infra companies!

@dylan522p Pretrainitus captures it. $10-13B per gigawatt to rent what the frontier labs already have this year, and now startups want in on that game with revenue nowhere near it. Ban anxiety is almost beside the point if your own model is already training.

@willccbb Can’t wait for the Vera Rubins to get in for this.

@dylan522p "Everyone" is doing a lot of heavy lifting here my desi brother. Not every revenue generating startup has the VC runway or capital to pretrain from scratch

@dylan522p I was not concerned about it; thx. for a new fear 😭

@dylan522p What does this mean?

@dylan522p applied AI co's are training their own models from scratch? dont you need like a couple 100m for that?

@dylan522p Elon was ahead of his time catching g that