/Tech2h ago

Developer teortaxesTex argues high proprietary LLM prices and low throughput suggest providers run massive batch sizes to maximize profit

Creator swyx warned the chart's underlying data remains unverified

4862106.1K

#225

Original post

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex#501inTech

> but this would mean the margin on these models is CRAZY compared to oss models

…yes, it is Furthermore, given that they have vastly better hardware than the Chinese, but WE don't get 300 t/s on GPT 5.5, they must serve this crap with insane batch sizes it's a money machine

elie@eliebakouch

i have a hard time believing this. at the same time it would make sense that the CEO of cursor know the size of closed models, but this would mean the margin on these models is CRAZY compared to oss models.

this plot doesn't even account for the fact that serving at scale reduces cost (much higher batch, better hardware) and that they have many more engineers working on optimizing the inference stack ect.. (this is a log-log plot btw)

11:15 AM · Jun 17, 2026 · 3.5K Views

Sentiment

Negative users dismiss claims that closed AI models deliver massive margins over open-source alternatives, arguing the gains are merely incremental from extra training on scaled models.

Pos

0.0%

Neg

100.0%

1 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS1.7K

swyx@swyx

@eliebakouch yeah also doubt. people just be vibecoding charts all over because who's really checking

swyx@swyx

uhhh

did Mustafa just leak the Mythos FLOP count??

was this public knowledge before, even if its an estimate i dont get what you gain out of this

2h1.7K101

BOOKMARKS3LIKES16

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

absolutely disgusting tbh, their performance is roughly what you'd expect from a V4-scale model that got another couple epochs and more RL. They're really not in a rush huh "GPT-4 will be 100T params" my ass This is also how Westoids treated rare earths btw Just give China GPUs

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> but this would mean the margin on these models is CRAZY compared to oss models

…yes, it is Furthermore, given that they have vastly better hardware than the Chinese, but WE don't get 300 t/s on GPT 5.5, they must serve this crap with insane batch sizes it's a money machine

2h1.1K163

levzzz@levzzz5154

@teortaxesTex ye u can estimate real costs by just looking at the sub quotas, i think they serve them at break even in the worst case scenario https://she-llac.com/claude-limits

2h271