One percent of companies pay most of the AI bill. Cheaper tokens aren't changing that.
24 Oct 2026 · News
Everyone is arguing about whether AI is a bubble. The sharper question is how many customers are actually paying for it. Company card data says: not many.
The AI bubble fight usually circles chips, power and debt. Spending data from Ramp points somewhere simpler: about 1% of businesses account for roughly 80% of what companies pay OpenAI and Anthropic. Tokens just got a quarter cheaper in a month, and the same short list is soaking up the extra. Our call: that list is still this short in March.
The number
Ramp, a corporate card company, publishes a monthly AI Index built from transactions at more than 85,000 businesses. In its October letter, lead economist Ara Kharazian repeated a figure the firm had shared before: "80% of spend on OpenAI and Anthropic is driven by 1% of firms."
The median business in Ramp's data spends $15.70 per employee per month on AI. The median firm in the top 1% spends $9,567.53, per the live index. That's not a bigger customer. It's a different animal.
One caveat Ramp states itself: it only sees what companies pay for. Free tools and personal accounts don't show up.
Cheaper, not wider
The newer data is about price. Among businesses using Ramp's token tracking product, the blended price of OpenAI and Anthropic tokens fell 25.4% in four weeks, from $0.86 to $0.64 per million tokens (week ending 11 Oct versus week ending 13 Sep). Volume rose 52.4%. Spending rose only 23.0%. Ramp says that slice isn't fully representative, though the direction matches.
So companies are buying far more for a bit more money. Kharazian's read is that business AI is near "the top of the S-curve, where the growth in the market will have to come from more intense adoption rather than more customers." Paid adoption sits at 55.9% of businesses. He also writes that on contracts, "the customer is in the favorable negotiating position," with labs handing out discounts.
Odd setting for a gold rush: falling prices, a flattening customer count, buyers with leverage.
Who's on the short list
Apollo's chief economist Torsten Slok added a twist on 22 Oct: "80% of the top 1% of AI spenders are VC-backed, compared with only 13% of the median spender." His worry is that "a pullback in venture funding could in turn weaken demand for compute." Translation: much of the bill is startups spending investor money.
Anthropic's own paperwork says the same thing from the other side. Reuters, which saw its draft prospectus, reported in September that nearly a quarter of its 2025 revenue came from two customers, and that many of its largest clients weren't locked into long-term contracts. The Wall Street Journal's look at a later draft this week led with "a narrow customer base."
Plug versus payer
On Thursday we wrote that AI building now runs backwards: "buy the brains first, then wait for the plug." That still holds. Business Insider reported Amazon parking billions in Nvidia racks in warehouses because some sites lack power to run them.
Yet Gartner's chief forecaster, John Lovelock, told CNBC on Monday: "At the moment, we have more compute capacity than we have demand, and we have a lot of the tokens and processing power being given away in freemium solutions." Gartner's 14 Oct forecast puts more than $3.5 trillion of AI infrastructure spending at risk of being written down by the end of 2030, and says over 80% of neoclouds, the newer GPU rental firms, will "cease independent operations" by then.
Both are true, at different layers. Power is short at the building. Paying customers are short across the market. Earlier this month we looked at consumer AI, where the top 1% of payers were 19.5% of spend. On the business side, the top 1% carries about four times that share.
Why it matters
When the price falls and the customer list doesn't grow, each big account matters more every month. If one of them switches labs, builds its own model or loses its funding, the hole is large and can open fast, since many aren't locked into long contracts. Cheaper tokens don't fix that. They mostly let the same few buy more.
The bubble question is less about how many chips get plugged in than how many names are on the invoice.
Our call
By Ramp's March 2027 AI Index, the top 1% of businesses will still account for at least three quarters of company spending on OpenAI and Anthropic. And when Anthropic's prospectus goes public, its two largest customers will still add up to at least a sixth of revenue. What would prove us wrong: Ramp putting that share below 75%, or the public filing showing no pair of customers that big.