Ramp's card and bill-pay data puts about 80% of OpenAI's and Anthropic's enterprise revenue with 1% of the businesses that spend on them — and Ara Kharazian says that 1% has started spending less.
Concentration on its own is normal for a new technology. What makes this different, on his reading, is that the customers are all the same kind of company.
"A bigger problem is that within that, those companies tend to be very correlated with each other. It's by and large high growth tech companies, other AI startups as well. Now that's fine as long as AI spend continues to increase."
Kharazian is lead economist at Ramp, whose AI Index tracks what American businesses actually pay the model companies, months before either lab has to file a public financial statement.
I listened to the full segment so you can skip it.
Here are the 6 numbers that matter.
👤 Guest: Ara Kharazian, Lead Economist at Ramp, who publishes the firm's AI Index on corporate AI spending
🎙️ Host: Ed Elson, who presents Prof G Markets for Prof G Media
📰 Published: 10 September 2026 on YouTube (Prof G Markets)
🔴 YouTube | ⏱️ 9 min
Key Takeaways
About 80% of the two labs' enterprise revenue comes from 1% of the businesses spending on AI, and Kharazian says the market is underpricing that
In software outside AI you have to reach the top 10% or 20% of customers before the concentration looks like this
The 1% are not a diversified base: they are mostly high-growth technology companies and other AI startups, so they turn at the same time
Spend per employee among the heaviest users fell about 10% in a month, from $8,000 to $7,200
Summer seasonality is part of it but not all of it, because token usage rose over the same period
The price war between the two labs is pushing business spend toward their cheaper tiers rather than their frontier models
He is neither a bull nor a bear, and he does not think Chinese or open-source models take the enterprise market
His bullish case is the same shift: lower-margin models at much higher volume
The concentration exists because AI works best on engineering work, and the labs have not yet shipped something the rest of the economy wants
1. 1% Drives 80% of Revenue
Elson opened on the two IPOs nobody can see inside: OpenAI and Anthropic have both filed confidentially with the SEC, and until the filings are public, third-party data is what there is. Ramp's latest AI Index, titled "cracks in the AI thesis," is where the 80% figure comes from.
Kharazian's first move was to warn against reading one metric. His caution was that with a data set this new you do not want to focus too heavily on a single number or over-highlight one trend, because noisy data invites exactly that
What changed his mind is that several metrics moved together. "What's what's different about this is that we are seeing a number of metrics now move in a negative direction for the AI model companies and we are seeing a great amount of concentration risk as far as their customer base. So if you are someone who is you know involved in the AI trade that might be something that you want to take a look at that I think the market is relatively underpricing."
The headline number is a customer count, not a revenue split. "So you're right to point out, you know, a line share of the AI company's enterprise revenues is coming from a specific segment of customers, 1% of customers driving 80% of that spend."
The correlation inside that 1% is what he calls the bigger problem. "A bigger problem is that within that, those companies tend to be very correlated with each other. It's by and large high growth tech companies, other AI startups as well. Now that's fine as long as AI spend continues to increase."
2. The Top Spenders Slowed
The new finding in that week's index is not the concentration, which Ramp has published before, but that the concentrated customers have started cutting.
The drop is a month-on-month figure on a per-head measure. "But then our latest finding today is that the top 1% of spenders in terms of per employee per month spend are also slowing down their spend. So came down about 10% month over month from 8K per employee per month down to 7.2K. Still a lot of spend."
He pre-empted the seasonal explanation and then part-rejected it. "You know people can see that metric and they think oh that's because of summer seasonality. And that is somewhat true, but it is not a full explanation."
The reason it is not the full explanation is that usage went the other way. "Token volume and usage of these models actually rose in the same period." Businesses ran the models more and paid less for doing it
3. The Price War Below
If usage rose while spending fell, something changed about what businesses are buying. Kharazian's answer is that both labs are competing on price at the top and shipping cheap models underneath, and the buyers are moving down.
Both are cutting at the frontier and building below it. "One is that OpenAI and Anthropic are in this price war. So both introducing frontier models and cutting prices for them, trying to get the best model at the best price. Simultaneously introducing highly performant, cheaper models at the standard and light tier."
The mix has shifted, and he named the models. "And then you see an increasing share of that volume of AI spend volume by businesses going toward the standard and light models as opposed to the sort of frontier models. So businesses deciding they get pretty good ROI and performance from Sonnet and Terra over Astra, Opus, Fable, and GPT 5.6 Soul."
His summary of the picture is that several readings point the same way. "And so you have a couple metrics sort of moving in a negative direction."
4. Worse Than Any Software
Elson quoted the report back at him, on a "concentration risk unseen in any other software category" Ramp tracks, and asked how it compares with what Ramp sees elsewhere.
Other software has a power law; it does not have this one. "Well, we looked at some of the other large spend categories for businesses. So software excluding AI, you still see some power law, you know, but you don't see it to the extent of 1% driving 80% of revenues, you know, you have to get to sort of the top 10, top 20% before you start to see something of that magnitude."
Advertising is the flatter comparison. Digital advertising, he said, tends to be "more broad and spread out"
He repeated that concentration by itself is not the risk. "So I think it's normal to have this in a new technology to be clear and it's not necessary problem unless if those companies in that customer base are highly correlated and simultaneously end up drawing down their spend" — which is the combination he says the data now shows
Elson's framing was harder than Kharazian's. The host put it that the economics of the labs only work with a subsidized handful of customers providing the growth, and said the same pattern shows up on the supplier side, with Anthropic and OpenAI accounting for 70 to 80% of the AI revenue at Microsoft, Amazon and Google. That is the host's own claim, not Ramp's data
5. Not a Bear Case
Asked to draw the conclusion for the AI trade, Kharazian declined to take a side and then laid out the bullish reading of his own numbers.
He put himself outside the argument. "Well, I'm I'm neither on the side necessarily of the AI bears or the AI bulls, right?"
The bull case is growth at a slower rate and a bigger share of the wallet. "You can still see a bit of a bullish thesis come through here, which is that OpenAI and Anthropic will continue to draw more and more spend, maybe at a slower rate than what they've seen so far. And most notably that they will continue to draw a greater share of spent toward their models. Maybe it's going to be their standard and lighter models which are lower margin but they will get higher volumes there."
He does not think the open-source threat is the real one. "And that's an important part of the thesis right because lately a lot of the concern has been about the threat of Chinese model companies or open- source models eating at the market share of OpenAI and Anthropic and I don't think that's necessarily going to happen. I do think OpenAI and Anthropic will be in a very good place to command enterprise market share for AI at least in the United States for the foreseeable future."
What changes is the mix, not the winner. "It's just likely going to be shifting over towards standard and lighter models rather than the frontier models that have generated so much of the attention so far."
6. AI Hasn't Left Engineering
Elson's last question was how seriously the concentration risk should be taken against any other business risk. Kharazian's answer explained why the customer base looks the way it does — and what would have to change for it to broaden.
The 1% are technology companies because that is where the product works. "Part of the reason why other AI companies and tech companies are the biggest spenders on AI models is because AI models are most commercially advanced and most productivity enhancing today for technical professions, for software engineers. It's coding agents, it's engineering tasks, it's it's particularly AI sort of works especially well and is diffused especially effectively in these sort of highly technical industries."
That is not where the labs say they are going. "You know, if you want to follow the goals of the AI model companies as far as their public statements, it's that they want to see proliferation across the US economy into non-technical work into broader areas of white collar work eventually into areas of manufacturing automation."
He gives two reasons it has not happened, and only one is about time. "You know, the reason why we haven't seen that uptake in part it's because tech is pretty laggy tech. It takes time for technology to proliferate through a society. But it's also because the model companies themselves have not yet introduced a sufficiently enticing commercially advanced version of AI that works for the general population."
He leaves it open. "And so I think they're working on that, but you know, time will tell if they're going to be successful."
It is a risk the labs themselves want to fix, in his account, separately from any worry about a market turning against AI
Bonus Insights
Elson's closing note on the data was that he cannot wait to see the S1 filings, and that until then third-party providers like Ramp are how anyone learns what is happening inside the two companies
The same episode carried a market read in which Brent crude topped $101 a barrel for the first time since May and the 10-year Treasury yield hit its highest level since 2023
This is one of three interviews in the same Prof G Markets episode, which also carried Alex Kantrowitz on the Anthropic extinction warning and Patrick McGee on Apple's foldable iPhone, each written up on its own
Kharazian's bottom line is that the AI labs' enterprise business is real, growing and dangerously narrow at the same time: it rests on 1% of customers who are mostly each other, those customers have just cut per-head spending by about a tenth while using the models more, and the broadening the labs are counting on waits on a product the rest of the economy actually wants.
Products, Companies & Tools Mentioned
Ramp (Kharazian's employer; its corporate card and bill-pay data is where every figure in this segment comes from)
OpenAI and Anthropic (Both have filed confidentially with the SEC; between them they are the enterprise AI market Ramp is measuring, and they are in a price war at the frontier)
Microsoft, Amazon and Google (Elson's point, not Ramp's data: he says the two labs account for 70 to 80% of the AI revenue at the big data center owners)
Books & Resources Mentioned
Ramp AI Index (The live tracker behind the 1%-of-customers and per-employee spending figures)
Ramp AI Index, September 2026 (The specific report Elson calls "cracks in the AI thesis," published the day before the episode)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: