WEBVTT - Exclusive: Here's How Much OpenAI Spends On Inference And Its MSFT Revenue Share

0:00:03.160 --> 0:00:06.760
<v Speaker 1>Media. Hello, and welcome to a very special episode of

0:00:06.800 --> 0:00:09.440
<v Speaker 1>Better Offline. I'm, of course, your host ed ze trunk.

0:00:17.880 --> 0:00:20.400
<v Speaker 1>For years, I've been hunting down the core details behind

0:00:20.440 --> 0:00:23.159
<v Speaker 1>open aiyes costs and revenues, and today I'm going to

0:00:23.160 --> 0:00:25.159
<v Speaker 1>bring you some of them. A lot of what I

0:00:25.160 --> 0:00:27.120
<v Speaker 1>say today is going to be reflected in my newsletter,

0:00:27.120 --> 0:00:29.840
<v Speaker 1>which I'll link to in the notes. Based on documents

0:00:29.920 --> 0:00:32.839
<v Speaker 1>viewed by my newsletter, I'm able to report open aiy's

0:00:32.840 --> 0:00:35.760
<v Speaker 1>inference spend on Microsoft asure in addition to its payments

0:00:35.800 --> 0:00:38.640
<v Speaker 1>to Microsoft as part of its twenty percent revenue share agreement,

0:00:38.800 --> 0:00:41.480
<v Speaker 1>which was reported in October twenty twenty four by The Information.

0:00:42.159 --> 0:00:44.920
<v Speaker 1>In simple terms, that last bit means that Microsoft receives

0:00:44.960 --> 0:00:48.120
<v Speaker 1>twenty percent of open eyes revenue in addition to whatever

0:00:48.200 --> 0:00:52.159
<v Speaker 1>it spends on GPUs and servers. As a reminder, influence

0:00:52.240 --> 0:00:54.240
<v Speaker 1>is the process through which a model creates an output,

0:00:54.240 --> 0:00:56.480
<v Speaker 1>which I'll be reminding you of a few times because

0:00:56.560 --> 0:01:00.120
<v Speaker 1>it's actually kind of important. Now a few notes. I

0:01:00.160 --> 0:01:02.520
<v Speaker 1>don't have open AI's training spend, nor do I have

0:01:02.600 --> 0:01:05.920
<v Speaker 1>information on the entire extent of open AI's revenues, as

0:01:05.959 --> 0:01:08.760
<v Speaker 1>it appears that Microsoft shares some percentage of its revenue

0:01:08.760 --> 0:01:10.840
<v Speaker 1>from being as well as twenty percent of the revenue

0:01:10.880 --> 0:01:13.600
<v Speaker 1>Microsoft receives from selling open AI's models on as you.

0:01:14.480 --> 0:01:17.720
<v Speaker 1>What I do have, as I've mentioned, is its inference spend.

0:01:18.240 --> 0:01:20.280
<v Speaker 1>And if you're new to this, like I said, this

0:01:20.400 --> 0:01:23.280
<v Speaker 1>means all the computation's open ai does when processing requests

0:01:23.319 --> 0:01:26.640
<v Speaker 1>sent to its services like chat, GPT, and Sora. Now,

0:01:26.640 --> 0:01:29.640
<v Speaker 1>before publishing, I asked a financial time supporter to help

0:01:29.680 --> 0:01:32.080
<v Speaker 1>corroborate some of the data in the documents. They reached

0:01:32.080 --> 0:01:36.160
<v Speaker 1>out to Microsoft and open Ai, who both declined to comment. Now,

0:01:36.200 --> 0:01:37.960
<v Speaker 1>the following will be a lot of numbers, and it

0:01:38.040 --> 0:01:40.520
<v Speaker 1>might be easier for you to read them. However, I'm

0:01:40.520 --> 0:01:42.160
<v Speaker 1>going to try and make things as easy and clear

0:01:42.160 --> 0:01:44.720
<v Speaker 1>as possible, because the documents I've seen call into question

0:01:44.800 --> 0:01:47.240
<v Speaker 1>what we actually knew about open AI's business and the

0:01:47.280 --> 0:01:51.760
<v Speaker 1>sustainability of said business. To keep things simple, all the

0:01:51.800 --> 0:01:55.000
<v Speaker 1>years in this piece are calendar years. Microsoft has fiscal years.

0:01:55.040 --> 0:01:57.000
<v Speaker 1>I'm not going to play that game. It's impossible to

0:01:57.000 --> 0:02:01.200
<v Speaker 1>follow along with nobody thinks this way. Now we've done that,

0:02:01.400 --> 0:02:04.720
<v Speaker 1>let's get to him. According to the document's view by

0:02:04.760 --> 0:02:07.840
<v Speaker 1>my newsletter, open Ai spent five point oh two billion

0:02:07.880 --> 0:02:10.880
<v Speaker 1>dollars on inference alone with Microsoft Azure in the first

0:02:10.880 --> 0:02:14.680
<v Speaker 1>half of calendar year twenty twenty five. This is a

0:02:14.720 --> 0:02:17.040
<v Speaker 1>pattern that has continued through the end of September twenty

0:02:17.080 --> 0:02:19.440
<v Speaker 1>twenty five, by which point open ai had spent eight

0:02:19.480 --> 0:02:23.920
<v Speaker 1>point six seven billion dollars just on inference. Open AI's

0:02:23.919 --> 0:02:27.120
<v Speaker 1>inference costs have risen consistently over the past eighteen months too.

0:02:27.560 --> 0:02:30.560
<v Speaker 1>For example, open ai spent three point seven six billion

0:02:30.600 --> 0:02:33.600
<v Speaker 1>dollars on inference in twenty twenty four, meaning that open

0:02:33.639 --> 0:02:36.240
<v Speaker 1>ai has already more than doubled its inference costs in

0:02:36.360 --> 0:02:39.919
<v Speaker 1>just the first nine months of twenty twenty five. These

0:02:39.960 --> 0:02:43.520
<v Speaker 1>costs are dramatic and significantly higher than has been previously reported.

0:02:44.320 --> 0:02:47.639
<v Speaker 1>According to the Information, open AI's computer run models, which

0:02:47.680 --> 0:02:50.320
<v Speaker 1>I understand to mean inference, was two billion dollars in

0:02:50.320 --> 0:02:54.040
<v Speaker 1>twenty twenty four. Additionally, an April twenty twenty five piece

0:02:54.080 --> 0:02:56.920
<v Speaker 1>from the Information stated that open AI's inference costs for

0:02:56.960 --> 0:03:00.480
<v Speaker 1>twenty twenty five would be around six billion dollars, or

0:03:00.560 --> 0:03:03.200
<v Speaker 1>roughly two billion dollars less than open ai appears to

0:03:03.200 --> 0:03:05.600
<v Speaker 1>have spent through the end of September. I want to

0:03:05.600 --> 0:03:08.000
<v Speaker 1>be clear as well, I'm just reporting what these documents

0:03:08.040 --> 0:03:10.880
<v Speaker 1>have said. This is not a statement about the information.

0:03:10.960 --> 0:03:14.800
<v Speaker 1>They do great reporting. But then there's the issue of

0:03:14.840 --> 0:03:18.440
<v Speaker 1>the revenue share. As I've previously stated, the following numbers

0:03:18.440 --> 0:03:20.640
<v Speaker 1>are based on the revenue share paid to Microsoft as

0:03:20.720 --> 0:03:23.600
<v Speaker 1>part of its deal with open Ai, where it gives

0:03:23.600 --> 0:03:27.040
<v Speaker 1>Microsoft twenty percent of its revenues. According to the documents,

0:03:27.080 --> 0:03:30.200
<v Speaker 1>Microsoft received four hundred and ninety three point eight million

0:03:30.240 --> 0:03:32.720
<v Speaker 1>dollars in revenue share payments in twenty twenty four from

0:03:32.760 --> 0:03:36.320
<v Speaker 1>open Ai, implying revenues for twenty twenty four for open

0:03:36.360 --> 0:03:38.680
<v Speaker 1>Ai of at least two point four to six nine

0:03:38.720 --> 0:03:41.400
<v Speaker 1>billion dollars, or around one point twenty three billion dollars

0:03:41.520 --> 0:03:44.400
<v Speaker 1>less than the three point seven billion dollar number that's

0:03:44.440 --> 0:03:48.160
<v Speaker 1>been previously reported in multiple outlets. Similarly, for the first

0:03:48.200 --> 0:03:51.040
<v Speaker 1>half of twenty twenty five, Microsoft received four hundred and

0:03:51.080 --> 0:03:54.200
<v Speaker 1>fifty four point seven million dollars as part of its

0:03:54.200 --> 0:03:57.240
<v Speaker 1>revenue share agreement, implying open AI's revenues for that six

0:03:57.280 --> 0:03:59.600
<v Speaker 1>month period or at least two point twenty seven three

0:03:59.640 --> 0:04:02.640
<v Speaker 1>billion dollars or around two billion dollars less than the

0:04:02.680 --> 0:04:05.480
<v Speaker 1>four point three billion dollars previously reported for that period.

0:04:06.640 --> 0:04:10.080
<v Speaker 1>Through September, Microsoft's revenue share payments total eight hundred and

0:04:10.120 --> 0:04:13.240
<v Speaker 1>sixty five point eight million dollars, implying open AI's revenues

0:04:13.280 --> 0:04:15.800
<v Speaker 1>are at least four point three to two nine billion

0:04:15.840 --> 0:04:18.120
<v Speaker 1>dollars through the end of Q three twenty twenty five.

0:04:19.080 --> 0:04:22.040
<v Speaker 1>To be clear, and I'm going to say this, Microsoft

0:04:22.120 --> 0:04:24.440
<v Speaker 1>also pays open Ai a cut of bing's revenues under

0:04:24.440 --> 0:04:27.479
<v Speaker 1>certain circumstances I could not confirm, as well as a

0:04:27.520 --> 0:04:30.800
<v Speaker 1>cut of about twenty percent of all open ai models

0:04:30.800 --> 0:04:33.120
<v Speaker 1>sold through a zore. Just to be clear, Microsoft is

0:04:33.160 --> 0:04:35.640
<v Speaker 1>the only party that can sell open AI's models other

0:04:35.680 --> 0:04:38.560
<v Speaker 1>than open Ai. I don't have the details on those payments,

0:04:38.640 --> 0:04:40.480
<v Speaker 1>like I said, but I am skeptical that they can

0:04:40.480 --> 0:04:42.720
<v Speaker 1>account for the massive difference between those numbers that have

0:04:42.760 --> 0:04:45.560
<v Speaker 1>been leaked and the ones in the documents in question.

0:04:46.320 --> 0:04:48.320
<v Speaker 1>I do not know, nor will I speculate on why

0:04:48.360 --> 0:04:51.480
<v Speaker 1>these differences are so distinct. What was important about today

0:04:51.520 --> 0:04:53.240
<v Speaker 1>was getting you these numbers and shedding light on the

0:04:53.279 --> 0:04:56.039
<v Speaker 1>differences I see between the story told about open ai

0:04:56.240 --> 0:04:59.720
<v Speaker 1>and the reality of its spend and potential revenues. You've

0:04:59.720 --> 0:05:01.800
<v Speaker 1>also I probably noticed that this podcast is a bit

0:05:01.839 --> 0:05:05.000
<v Speaker 1>of a different tone to the usual no insults, no jokes,

0:05:05.200 --> 0:05:07.680
<v Speaker 1>haven't called anyone clammy, haven't even said a swear word

0:05:07.720 --> 0:05:10.400
<v Speaker 1>for the first time in maybe one hundred episodes. The

0:05:10.440 --> 0:05:14.400
<v Speaker 1>reason simple. These numbers are serious and seriously different to

0:05:14.440 --> 0:05:18.600
<v Speaker 1>those reported. Open AI's costs are dramatically higher than previously

0:05:18.640 --> 0:05:22.000
<v Speaker 1>reported in thought, and based on the extrapolations from Microsoft's

0:05:22.000 --> 0:05:26.119
<v Speaker 1>revenue share, its implied revenues are also seemingly dramatically lower

0:05:26.160 --> 0:05:29.480
<v Speaker 1>than we knew. The ramifications of these numbers are severe.

0:05:30.120 --> 0:05:33.480
<v Speaker 1>Open AI's influence costs are incredibly high, absorbing any and

0:05:33.520 --> 0:05:36.320
<v Speaker 1>all revenues and seemingly scaling with every increase in chat

0:05:36.400 --> 0:05:40.200
<v Speaker 1>GPT's user numbers. As revenue goes up, so does their

0:05:40.240 --> 0:05:44.240
<v Speaker 1>inference costs. Conversely, if these implied revenues are indicative of

0:05:44.279 --> 0:05:47.880
<v Speaker 1>the larger financial picture, open ai is not as successful

0:05:47.880 --> 0:05:51.479
<v Speaker 1>as company as we previously believed. In any case, the

0:05:51.520 --> 0:05:54.760
<v Speaker 1>reality of the AI bubble is becoming clearer. Inference, the

0:05:54.800 --> 0:05:57.240
<v Speaker 1>process of creating outputs for a model appears to be

0:05:57.279 --> 0:06:00.159
<v Speaker 1>an incredibly burdensome cost. And if these implied revenue is

0:06:00.160 --> 0:06:04.400
<v Speaker 1>there any indicator the actual business of selling generative AI

0:06:04.560 --> 0:06:07.960
<v Speaker 1>services and models doesn't really seem to be as good

0:06:07.960 --> 0:06:11.760
<v Speaker 1>a business as we thought either. It's all looking a

0:06:11.760 --> 0:06:15.720
<v Speaker 1>little bleak out there. I don't want to editorialize too

0:06:15.800 --> 0:06:17.960
<v Speaker 1>much because I want this information to sit on its

0:06:18.000 --> 0:06:22.320
<v Speaker 1>own own self. But it's it's strange being here. It's

0:06:22.360 --> 0:06:25.320
<v Speaker 1>strange getting these numbers and seeing them myself, and I

0:06:25.360 --> 0:06:27.919
<v Speaker 1>have to wonder how things work out from here. I

0:06:27.960 --> 0:06:30.440
<v Speaker 1>truthfully have no idea, but I do know I'll be

0:06:30.440 --> 0:06:32.920
<v Speaker 1>happy to do this every week, and I will tell

0:06:32.960 --> 0:06:36.600
<v Speaker 1>you what happens now. These numbers allow us to kind

0:06:36.600 --> 0:06:39.840
<v Speaker 1>of see the real picture of the AI bubble, and

0:06:39.880 --> 0:06:42.320
<v Speaker 1>I have to wonder what other companies look like now

0:06:42.320 --> 0:06:46.720
<v Speaker 1>that I've seen these numbers. Email me contact me Esitron

0:06:46.800 --> 0:06:49.080
<v Speaker 1>dots seventy six on signal. If you ever want to

0:06:49.080 --> 0:06:51.680
<v Speaker 1>tell me anything, do you ever want to show me anything.

0:06:52.440 --> 0:06:55.159
<v Speaker 1>I'm always interested to hear, and I'm honored to do

0:06:55.240 --> 0:06:55.360
<v Speaker 1>this