If the leaked financial data is right, OpenAI is profitable on serving customers with 40%+ gross margins. But training remains incredibly expensive.
Automating AI research may also be a play to increase the efficiency of training: a superhuman researcher could do more with less.
Positive users view OpenAI's leaked 40%+ serving margins as validation of a viable business model, while negative users dismiss them as unsustainable or likely to result in higher rate limits.
No Digg Deeper questions have been answered for this story yet.
https://www.ft.com/content/e15b0d7e-ff6b-4f16-ba7a-4068feddb828?syn-25a6b1a6=1
If the leaked financial data is right, OpenAI is profitable on serving customers with 40%+ gross margins. But training remains incredibly expensive.
Automating AI research may also be a play to increase the efficiency of training: a superhuman researcher could do more with less.
@emollick Training costs per FLOP might drop significantly if automated researchers optimize architecture search and data curation pipelines more efficiently.
@emollick The real question I have here is when is it going to be considered "enough training" to where the companies are gonna say we no longer need "free users" for training data
@emollick Boom
@emollick how do they make a profit, by subscribing users or providing an API?
@emollick profitable serving customers, unprofitable creating the product
@emollick Those inference margins will go away once Cursor is in the picture and open models get good enough
@emollick What are the chances frontier LLM training experiences some serious diminishing returns in the not so distant future?
@emollick OpenAI's 40% margin is a snapshot of this generation.
Each new model resets the whole cost structure, training spend, inference price, everything. Serving margin today tells you almost nothing about what next year's model costs to run will be.
@emollick So the business sells shovels at 40% margin and uses the cash to build a robot that digs the next mine. Weirdly elegant. The pitch decks calling this "unprofitable" are reading the income statement like it's a lemonade stand.
@emollick The 40% margin is on inference — serving the model. The bleeding is all upstream in training the next one. So the real question isn't "is OpenAI profitable," it's "can each new model earn back a bigger training bill than the last."
@emollick 40% gross margin means the business works by automating research,
which means the moat compounds different problems, in the same direction
@emollick 40% gross margin on serving is a real business.
@emollick 40% margin sounds quite terrible TBH, expect more subscription rate limits
@emollick Who is paying you to post? Can you read? I know math is hard but this is propaganda. Are you trying to get people to invest in money losing businesses? Speak up
@emollick The real question is whether automating research compresses the cost curve fast enough to matter. Margins mean nothing if training costs scale faster than efficiency gains.
@emollick the superhuman researcher angle is cope though. they just want cheaper compute without paying the electric bill.
@emollick automating the research to cut training cost is interesting on the surface
but isnt that just letting the lab design its own bill
