Almost everything in the AI economy is a bet about the future.
When Nvidia reports its quarterly earnings, its backlog — orders planned but not yet filled — matters nearly as much as revenue. Anthropic and OpenAI’s IPO chatter and “valuations” are bets on what they’ll earn as much as decades from now. Companies like Coreweave finance data centers before they have tenants. Everyone is pushing capital through the same bottleneck, roughly 2% of GDP a year, on a simple premise: demand for AI compute is close to infinite, so either businesses will pay more for smarter models, or they’ll use so much of them that it won’t even matter.
Jensen Huang, CEO of Nvidia, calls that the “two exponentials” driving the price of AI compute; models are growing more complex, and more people and agents use them. Either way, the idea is that the labs capture that surplus and send it back through the ecosystem to cover their debts. There’s just one problem: as impressive as the new model releases are, they don’t seem to be causing sustained spikes in the price of AI compute—in fact, the AI token is getting cheaper, fast.
That’s according to new data from Ramp, the corporate spending platform, published Wednesday, showing the effective price that American businesses pay per a million tokens has fallen about 41% from its peak in March, from $1.15 to 68 cents. The share of usage going to frontier models is dropping, too; about 53% in early August to 45% by September. And the top 1% of spenders, the cohort that drives about 80% of OpenAI and Anthropic’s enterprise revenue, cut per-employee spend by nearly 10% in August.
It’s not a disaster or the bubble bursting but it is a “crack in the AI thesis,” Ara Khazarian, the Ramp chief economist who runs the Index, told Fortune. Rather than unleashing a gush of demand for the best models, tokens are starting to be priced more like a commodity– as interchangeable as salt or wheat. And commodity owners aren’t valued at $2 trillion. Morgan Stanley has flagged vulnerability for up to $300 billion in bonds financing neocloud buildouts—CoreWeave-style companies that borrowed to build data centers before signing tenants—if token prices don’t keep up.
Ramp isn’t the only one flagging the trend. Citadel Securities noted in June that a separate measure, Silicon Data’s LLM Expenditure Index, started to fall because of a “bifurcation” between frontier AI, concentrated among the few tech-heavy firms that can afford it, and the “everyday” AI the rest of the economy runs on.
“You have multiple metrics now starting to move in a negative direction,” Kharazian said.
He said that the price decline reflects a mix of labs being forced to cut prices—OpenAI slashed the cost of its GPT-5.6 Luna model by 80%, and Anthropic announced its own cuts last month—and customers trading down to cheaper and simpler models. Which makes it threatening, he added, to anyone “who’s expecting a full dream scenario where the AI companies grow with nothing curbing their enthusiasm.”
That was the mood in the Spring, as “tokenmaxxing” entered the tech lexicon, the media told stories of token-usage dashboards and Nvidia’s Huang insisted a $500,000 engineer should burn $250,000 a year in tokens. But by the summer, cost discipline set in; Amazon and Meta killed its own leaderboards in May, while Microsoft cancelled Claude Code subscriptions. Khazarian said he’s now hearing the opposite of tokenmaxxing from businesses: companies are imposing defaults that steer employees away from frontier models entirely.
“Companies are increasingly starting to use Terra and Sonnet,” he said; the mid-tier models that are “highly performant and also cheaper.” Ramp’s top 1% of firms, the most AI-intensive in the country, now spend about $7,200 per employee per month on AI—roughly a third of Huang’s target, and tapering off.
Some analysts blame the rise of open-source models, a hot-topic issue a few weeks ago for AI companies that has now faded to the background among the high tenor of AI discourse drama. Only 3.6% of businesses on Ramp’s platform use open source or Chinese models. And even if they did, there’s good evidence that Deepseek, Tencent, Alibaba, and other major Chinese companies are fighting a brutal pricing war that’s depressing token prices too.
So it’s an international phenomenon. Back home, OpenAI and Anthropic are in a lopsided competition, Kharazian said. Since Aug. 1, OpenAI’s effective price has fallen 38%, to 48 cents; Anthropic’s has fallen 22%, to 90 cents. Anthropic has charged nearly double of what OpenAI has all year and has held down a floor near 90 cents since June, but OpenAI’s price has kept falling. That suggests Anthropic “probably has some pricing power,” Kharazian said, but “that edge is wearing down” as OpenAI takes a growing share of tokens on price.
The token pricing patterns tell a similar story; Anthropic’s price spiked when Fable 5 launched in March, and OpenAI spiked with Sol in July. But they weren’t consistent, and prices eventually faded back down. “It’s not that businesses aren’t willing to pay high prices,” Kharazian said. “Prices are relative to the other products available on the market,” and the mid-tier products “are also really good.”
OpenAI’s finance chief described that dynamic on Tuesday as she spoke at a Goldman Sachs conference. CFO Sarah Friar said the company had cut the price of its GPT-5.6 Luna model by 80% since its launch, cheaper than some Chinese open-source weight models. She added that she’d like to get away from token counting altogether, and move enterprise customers to paying only for completed work. They claim that OpenAI’s enterprise revenue grew 32% from June to July, she said, but that growth seems to come from increased share.
“I would love,” Friar said, “to get us away from token-counting.”
