How Much OpenAI Spends on Inference and Its Revenue Share with Microsoft
wheresyoured.atThanks for sharing! It's pretty rare to see a balanced take from Ed Zitron.
> I also cannot reconcile these numbers with the reporting that OpenAI will have a cash burn of $9 billion in CY2025. On inference alone, OpenAI has already spent $8.67 billion through Q3 CY2025.
This is insane.
Explains why Sam was so panicked when asked about revenues recently.
> It's pretty rare to see a balanced take from Ed Zitron.
How are you understanding the word 'balanced'? Do you mean _tone_, or something? Like, this is unusually dry and non-grumpy for Ed, but it's if anything more critical than his normal output.
It's insane if the data is accurate. Only time will tell
You just have to believe. Believe in Sam Altman and everything he says and imagines. Believe in him when he says that he needs to spend 7% of world GDP on datacentres to pump out Sora slop and get his LLMs to type out the cure for cancer. Just believe. Believe harder.
So they are basically Enron 2.0 cooking the books.
- OpenAI claims to have spent $2.5 billion on inference in H1 of 2025, report claims it actually spent $5.02 billion
- OpenAI claims to have made $3.7 billion in revenues in 2024, report claims it actually made $2.469 billion
- OpenAI claims to have made $4.3 billion in revenues in H1 of 2025, report claims it actually made $2.273 billion
Some more insights in this FT piece:
How high are OpenAI's compute costs? Possibly a lot higher than we thought
https://www.ft.com/content/fce77ba4-6231-4920-9e99-693a6c38e...
Depends if these numbers are counting the free Azure credits that Microsoft invested in them.
There's a story here that's in some ways bigger than OpenAI or Anthropic's finances. Someone is leaking very sensitive and private financial information to Ed. They're clearly getting these numbers from somewhere and given the monthly breakdowns Ed posted previously for Anthropic, they are likely coming from a billing dashboard of some kind inside the big clouds. It's not very likely the leaks are coming from inside the AI labs themselves because of how cloud specific and incomplete they are.
For a big cloud to have a rogue insider like this is huge. It's really rare for big tech firms to leak private data and this report suggests MS or whoever has this problem hasn't been able to find the leaker, which is amazing. These numbers can't be that widely distributed surely? If companies like OpenAI aren't safe from leaks then nobody is.
No — the big story is definitely that the people essentially propping up the US investment market are bullshitting about the most consequential expenses they have for their long-term viability. The big story isn’t that someone was sneaky getting that information out to people.
Although not mentioned, I think the exponentially rising costs have something to do with Sora.
This along with the fact that it is easier than ever to switch models today. And the fierce competition from Google and Anthropic.
Cannot wait for the pump and dump that the OpenAI IPO is going to be.
How can have this had so little votes and comments?
That is pretty interesting. Even Ed's comment section is empty. I wonder if by toning the histrionics down he just fails to engage?
I’m just seeing it 5 days later, weird honestly.
NVIDIA earnings call is November 19th
brb buying puts
Nvidia is getting all the money though. It’s not their problem that people want to burn cash in their direction in order to build products that only lose money. At least it’s not their problem right now…
The problem is that all the AI companies have been getting incredibly intertwined, see [0] for example. Nvidia isn't just selling chips - it is also actively investing in its own customers in order to drive up demand. This is giving us neat headlines like "NVIDIA intends to invest up to $100 billion in OpenAI as the new NVIDIA systems are deployed". The Coreweave deal is even worse: Nvidia is investing in Coreweave, so that Coreweave can buy GPUs from Nvidia, so that Nvidia can rent compute from Coreweave.
Sure, Nvidia is making crazy money right now, but what's going to happen to all those deals when the market blinks and some of those lesser players start falling over?
[0]: https://nymag.com/intelligencer/article/ai-investment-is-sta...
To some degree this is normal - for example I worked at a med device startup where we got investment from one of our manufacturing partners. That helped with growth which in turn meant more manufacturing work for them, and presumably they'd be happy to invest in another round and keep the flywheel going as long as there is some real outside demand.
I honestly think it's a great model for incentive alignment and not that sketchy on the surface. For the manufacturer, it's guaranteed revenue with upside convexity. For the startup, it's better terms and priority from the manufacturers since they have a stake in your success.
the oldheads are snickering at dotcom 2.0 from behind their gray beards.
Bitcoin miners have also been sitting on the sidelines waiting for this too.
Compute is a hard resource, inextricably linked to money, time, and energy.
The math doesn't work in Sam's favor, no matter how much smoke he blows up your ass.
It's going to be interesting when all these GPUs are repurposed to mine Bitcoin, and people try to forget falling for the hysteria that somehow you can arrive at AGI from a glorified markov bot.
Is there any difference in a GPU that's good at bitcoin mining versus one that's good for AI work? Or to ask another way is all the compute being built-out now able to be repurposed for mining?
> Is there any difference in a GPU that's good at bitcoin mining versus one that's good for AI work?
No GPU is good for bitcoin mining; that's all been ASICs for a long time. Even before anyone got around to making ASICs for it, FPGA-based designs had displaced GPU mining. Bitcoin mining is very, very simple.
Some altcoins use GPUs.
And slowly building up their shorts and keeping cash on the sidelines to short harder when it finally all comes crashing down.
The problem with shorts is that bubbles can expand well beyond rationality. Cash or bonds is what I am doing.
That's not a problem at all, you just roll some of your bull market gains into close dated ATM puts periodically. If the market is sideways sell calls.
Genuinely curious, why do you need close dated ATM puts. If all you care about is protecting upside can't you just put in limit orders.