Functionally, the entire AI industry is dependent on two companies that don't have any money — OpenAI and Anthropic — that don't have natural cash flow. They have signed $1.2 trillion worth of compute commitments based on the idea that everyone will pay for it. 1% of customers hold up everything else. These aren't profitable companies either. Even if they were profitable, this would be horrible. So, yeah, this is a fake industry. Everyone is just running some kind of scam almost where they're like, 'Yeah, it's the biggest thing ever.' Other than the fact we have to do black magic to keep it alive.
Hi, I'm Isaac, and on The Tech Report with me today is the writer of Where's Your Ed At and the host of the Better Offline podcast, Ed Zitron. Thanks for being here.
So according to some new Ramp data, 80% of Anthropic and OpenAI's enterprise revenue comes from just 1% of their customers. And I feel like that says it all. So let's just get straight into the questions. What does that statistic mean? And kind of why is it important?
So people will say, 'Oh, Ramp doesn't have everything. Ramp doesn't have...' So Ramp very likely doesn't have like Microsoft and Meta spend. They don't have like the hyperscaler spend. What they do have is they have both regular business users, like a ton of enterprise customers, but also AI companies, so AI startups that use AI tokens to power their services, usually subsidized, running their spend through Ramp. So it's very indicative of the fact that for the most part, most people are not spending much money on AI.
That there is a very small contingent of maybe a few hundred companies. I would say high hundreds. I'm just guessing, but this is what I'm intuitively getting from this, and it means that there is a small group of customers who are spending all the money. Now, people also argue, well, that happens with all sorts of business, happens with Salesforce, their big customers make up most of their revenue. Yeah, but it's not quite the same thing because this isn't like a software subscription. This isn't them buying extra seats every month. It isn't them paying more for services. It is a consumption-based model. It's how much they are spending on AI models.
This is very, very bad because it means that, first of all, just to keep things going at a normal pace, everyone needs to spend the same amount or more every single month, at a time when Meta is putting out competitive models that are much, much cheaper just to mess with people, where every week someone discounts something. You've got open-source models. And remember, I've been kind of cool on the open-source movement. Like, not to say it's a bad thing, I think it's a great thing. How much is actually taken out of people's pocketbooks? Seeing this, I'm a little bit more... I think it'd actually be a real challenge, because think of it like this. Let's say it's 900 customers and they're each spending millions, tens of millions of dollars on AI credits from OpenAI and Anthropic. Let's just say that these are easily the most perfect customers to move to open source. Why wouldn't they? If you were spending tens of millions of dollars on something that is increasingly commoditized, increasingly hard to tell the difference between an... not an Astra, sorry, Astra hasn't come out, uh, Sora or whatever Muse, Spark Meta has or Fable, what have you. Like, yes, you might be able to tell a few of the differences, but for the majority of tasks, you probably can't.
So, you've now got... this is not diverse revenue. It means most of the revenue that's being made on the enterprise is being made by a small group of customers who can and will go with whatever model developer does the task better. You've got Thomson Reuters that's doing it. You've got AT&T who's doing it. I don't know if they run through Ramp. It's also not getting better. The customers are not getting more diverse. It's not... you're still at the 80-plus range, which means that... I know I'm trying to find a really simple way of saying it. Um, it means these companies are reliant on a... well, I mean, all of this is... it's a few hundred customers that two AI labs are dependent on, and then you've got five more compute providers who are dependent on them for spend just by buying capex and spending money from two companies, Nvidia and Broadcom.
I mean, yeah, to kind of get you to go into a bit deeper than that, how does this sort of concentration stack with the fact that 70% of AI compute is purchased by Anthropic and OpenAI?
I mean, it's just more... it's revenue concentration risk.
It's because in a healthy industry, yes, you do... you get whales, you get your big power players, you get the big spenders. That always happens. Nothing wrong with that on its face. The problem is that... I mean, there's a really badly written Claude slop piece from a company called Groundbreaker about this that makes this point, but functionally the entire AI industry is dependent on two companies that don't have any money, OpenAI and Anthropic, that don't have natural cash flow. So you have this concentration risk now where two companies, OpenAI and Anthropic, are dependent on maybe 900 companies to keep paying them and to keep paying them more, by the way, because it's not expanding over time. It's not like the small... it's not like there are more of these big customers who are popping up. It's very clear that there's a roughly the same size group every single month.
So they are dependent on those companies continuing to spend more and more and more and more and more, and they need to spend more. Just where they are right now is not enough. They need to spend more because they're not expanding. They need these specific customers to spend more. This is terrifying because in software sales, upselling is how you grow revenue. Always has been. And you do that through either price increases, which nobody likes, new features, which people will like. But in this case, they're not really selling to these companies. It's not like Anthropic turns around and says, 'Oh, we have a new add-on that costs 10 bucks a month per seat, and this is what you pay for it.' They say, 'Oh, we have a new model. It's more expensive.' Except Fable has seen terrible pickup. And Fable 5.1 has not seen great pickup either. So, that is the more expensive model that would theoretically increase the costs on the customer base.
On the other... the other problem is, I don't know what percentage it would be, but some contingent of this 1% are actually AI startups. So I don't... I'm just using Perplexity as an example. I don't know if Perplexity is actually one of them, but you're an AI coding startup and you let your customers pay 200 bucks a month or 20 bucks a month. With that, they're likely burning hundreds of dollars of tokens a month. So some of this revenue is coming from AI startups that subsidize their token spend for their customers.
This again is unstable revenue because all of that money is dependent on venture capital funding. So every link of this chain at some point comes down to: will venture capital keep funding this forever? And so it's a very unstable industry, like deeply unstable. And so you've got the revenue at the top, the customers paying for AI tokens. You've got some instability there. You've got AI startups that are subsidizing their token spend. So that revenue goes away when venture capital goes away. Then you've got, okay, the model developers, AI labs, the Anthropics, the OpenAIs. Well, they can only keep running their services as long as venture capital and hyperscalers keep paying them. Then you've got hyperscalers that can only keep buying GPUs as long as they're able to raise debt. And then you've got Broadcom and Nvidia who can only hit their revenue targets if everything above them goes well. And so it's, um, very unstable. That's the way I'd put it.
For 80% of the demand to come from 1% of the customers, that must mean that in that 1%, the amount being spent on AI per employee must be astronomically high.
What's your response to those who would look at that figure and say the 1% are just the early adopters, the 99% will catch up, and the AI bubble will be saved?
I mean, my answer would be go and look at the chart. It isn't getting better. And when I say it isn't getting better, I mean they're not expanding that base. That 1% is still most of the revenue. 80% of it, in fact. It got slightly better. You'll notice on the chart in April when they moved everyone to token-based billing. So what they did there was they were like, 'Okay, now you're going to pay more.' And it's very clear that that spend went away.
