Business adoption of AI tools slowed in August, according to spending data from 70,000 companies collected by payments firm Lamp. According to the latest research, 56% of Ramp customers paid for an AI product in August, an increase of just 0.4% from the previous month.
This isn’t the first time Ramp’s metrics have shown a slowdown in adoption. Last year, the company’s AI index saw little to no growth in adoption from August to October, but growth accelerated again towards the end of the year.
Still, AI is being built at such a rapid pace that even the slightest slowdown could be a cause for concern. The huge investments made by Frontier Labs and hyperscalers in AI infrastructure are based on the expectation that there will be enough revenue to pay them back. So far, usage has grown exponentially, especially as software engineers have adopted agent coding tools. But if that adoption slows, revenue could slow as well.
Ramp’s numbers likely overstate adoption overall, thanks to the company’s high-tech customers. According to the U.S. Census Bureau’s Continuing Survey on AI Deployment, updated on August 23, only 22% of companies report using AI. Although Ramp’s research is not necessarily representative of the market, it is one of the few direct spending data sets available and can be a leading indicator.
Granted, this data is from August, when much of the industry is on vacation. That may explain the downturn. But Lamp economist Ara Karazian says there are other warning signs for companies that rely on token spending.
First, AI spending per employee among the top 1% of AI-using companies in his sample decreased significantly, by nearly 10% to $7,205. That may be a factor in vacation tokens, but it’s also related to lower token costs. As OpenAI and Anthropic lowered their prices, the average token cost has fallen to $0.68 per million tokens, compared to the 2026 peak of $1.15 per million tokens in March.
The data suggests that the institute has not yet made up for price cuts associated with increased volumes. And with similar incentives, many customers are choosing to use older, cheaper models like OpenAI’s ChatGPT 5.6-Terra or Anthropic’s Sonnet instead of more powerful Frontier releases. Frontier Labs employees said much of the training costs can be recouped in the first few weeks of a new model’s release, but delays in implementation could threaten that dynamic.
Yet, despite the buzz about open-weight models threatening Frontier Labs, only 6.4% of AI spending companies used a model serving or inference platform in August. This rate is steadily increasing, but not fast enough to drive broader business adoption dynamics.
“We are showing that the competition between OpenAI and Anthropic is making AI more accessible and lowering the price for businesses. Not only are prices coming down, but they are also reducing spending for the top 1% of businesses, where the market was traditionally expected to drive much of its future growth,” Karazian said.
This also helps explain AI Lab’s focus on acquiring non-technical users of AI collaboration tools.
Can we even call this data point a “blip”? — If you’re a model builder or a hyperscaler with hundreds of billions of chips on order, this could be a bad sign. But Karazian says, “It depends on who you are in the market for. If your company is using AI, that’s great.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
