Their revenue in the entire first half per the article was $4.3B, even if we assume it was backloaded in Q2 that’s a -500% profit margin. I’m very excited to review their S-1.
Like, for every dollar you give them they spend $5 to deliver the product. They’re selling $10 bills for $2. They seem to have taken the joke about making up losses with scale pretty seriously.
What Theo claims on YouTube is that each model they make is profitable over its lifetime and the weird numbers for the company are because they keep training more models, having to pay up front to buy a bunch of future profit each time basically
The problem is that OpenAI will have to keep doing this basically indefinitely, as otherwise open-source commodity models will catch up with them and offer an equivalent product at a much lower cost. If the company is reliant on having to keep paying more than it's making on its current model to train new models, I don't see how it can ever become a sustainable business.
Not even just open source models - other commercial LLM models as well. All of the big LLM companies (Google, Anthropic, OpenAI) are basically ocked in a cold war of having to continuously outspend the other providers or risk becoming irrelevant.
The big question for me is if people will ever be happy with a model that is "good enough", and can thus be optimized and run profitably over time without faling behind. Time will tell!
This claim that if they stopped training new models wouldn’t the old ones become stale as things are updated? Not sure how quickly that would occur but it does seem likely as the world moves fast
this is bad reporting. microsoft's share of openai's cumulative loss over the lifetime of their investment is 11.5b. the microsoft share of openai's loss in the quarter was 4.1b, implying an 8b loss for openai itself assuming microsoft recognized 49% of openai losses per their ownership stake exiting the quarter. there may be some one time items that skew this, if not this is ain't good
Contrast this with Google who made $100Bn profit last quarter and on the side has Gemini with 650MAU and does 7B tokens/minute. It makes it clear that neither MS nor OAI know what they are doing.
Also most of it is from ads, not from their AI or cloud products. I wonder when OpenAI will start to serve ads as much as Google does, they certainly have people's attention a lot of times throughout the day looking at how people around me use their services.
Their revenue in the entire first half per the article was $4.3B, even if we assume it was backloaded in Q2 that’s a -500% profit margin. I’m very excited to review their S-1.
Like, for every dollar you give them they spend $5 to deliver the product. They’re selling $10 bills for $2. They seem to have taken the joke about making up losses with scale pretty seriously.
> Like, for every dollar you give them they spend $5 to deliver the product.
Well, it would be useful before concluding this to see a breakdown between OpEx, CapEx, and R&D.
What Theo claims on YouTube is that each model they make is profitable over its lifetime and the weird numbers for the company are because they keep training more models, having to pay up front to buy a bunch of future profit each time basically
The problem is that OpenAI will have to keep doing this basically indefinitely, as otherwise open-source commodity models will catch up with them and offer an equivalent product at a much lower cost. If the company is reliant on having to keep paying more than it's making on its current model to train new models, I don't see how it can ever become a sustainable business.
Not even just open source models - other commercial LLM models as well. All of the big LLM companies (Google, Anthropic, OpenAI) are basically ocked in a cold war of having to continuously outspend the other providers or risk becoming irrelevant.
The big question for me is if people will ever be happy with a model that is "good enough", and can thus be optimized and run profitably over time without faling behind. Time will tell!
Future profit? Isn't that what every company in the red claims?
Ofcourse in the 21st century none of this actually matters people just want to buy low sell high.
This claim that if they stopped training new models wouldn’t the old ones become stale as things are updated? Not sure how quickly that would occur but it does seem likely as the world moves fast
I would think that logic only works if opensource models like llama don't catch up? In other words the models don't become a commodity.
this is bad reporting. microsoft's share of openai's cumulative loss over the lifetime of their investment is 11.5b. the microsoft share of openai's loss in the quarter was 4.1b, implying an 8b loss for openai itself assuming microsoft recognized 49% of openai losses per their ownership stake exiting the quarter. there may be some one time items that skew this, if not this is ain't good
How long can this go on for?
Forever, tech has become a never ending sequence of bubbles and hype cycles.
The duration of this hype-cycle/bubble has exceeded my expectations.
As long as investors are willing to pay for it.
I can’t help but feel their push to go public is driven by their having milked the private markets dry.
[dead]
Contrast this with Google who made $100Bn profit last quarter and on the side has Gemini with 650MAU and does 7B tokens/minute. It makes it clear that neither MS nor OAI know what they are doing.
Alphabet (Google) made $102.35 billion in revenue last quarter, not $100bn in profit.
Also most of it is from ads, not from their AI or cloud products. I wonder when OpenAI will start to serve ads as much as Google does, they certainly have people's attention a lot of times throughout the day looking at how people around me use their services.
Not a single AI company is cash flow positive, not even Google's AI part of the business.
It's unclear if they even can be, since commodity models you can use for free catch up to them pretty close to yearly.