Businesses are using AI more than ever, but spending on AI isn’t increasing.
A new report from Ramp shows that AI spending leveled off towards the end of the summer despite increased token volumes, indicating that price competition and smart customers may be limiting OpenAI and Anthropic’s revenues. According to Ramp data, token spending increased by only 1.5% from August to September. By the way, the number increased by 42% in July compared to the previous month.
As Anthropic prepares for an IPO, data suggests investors may be wrong to assume the AI company will continue to grow at breakneck speed.
“Nearly all of the spending on AI is coming from venture-backed, high-growth technology companies,” Ara Kharazian, principal economist at Ramp, told TechCrunch. “As long as spending on AI continues to increase, that’s fine, but if something in the market or some other factor starts to slow down AI spending, everything starts to slow down at the same time, and that can cause other cascading effects.”
Harajian said he has already seen evidence that spending among the top 1% of companies (the most “AI-immersed”) stagnated slightly in August and September, leaving them just below the $10,000 spent per employee they reached during the summer peak. He said the group’s slowdown is “sufficient to completely slow down earnings, despite the fact that sales volumes are increasing.”
RAMP data shows an increase in spending for the median and top 10%, but not enough to make up the difference. The median company spends $12 to $15 per employee per month, and the top 10% spends $600 to $1,000.
“If you look at today’s use cases, it’s not necessarily clear that we’re going to reach the 1% level of $10,000 per employee per month,” Karazian said. While there’s a lot of demand for AI coding agents in industries like software and finance, “it’s not necessarily true that a doctor’s office or a car dealership needs to spend that much money on AI.”
“The era of high growth may be over, and we now know there may not be as much power to extract from the top 1%,” he continued.
While October data does show a re-acceleration, with token spending increasing 8.9% from September 23 to October 23, Karazian said headwinds to spending growth are strong.
Karazian, who publishes a series of blogs called “Cracks in AI Theory,” said the slowdown is not due to demand issues. Again, the volume of AI is increasing. Instead, he points to the competitive nature of the market.
“OpenAI and Anthropic’s price cuts just made it cheaper to use AI,” he said. In September, OpenAI cut the API price of GPT-6 Sol and GPT-6 Luna in half. To combat price competition, Anthropic has significantly reduced the prices of Fable 5.1 and Opus 5.5. The decision, which the company announced in a recent filing, reduced revenue growth by about $600 million. Both companies also offer corporate discounts meant to lock in spending, but actually meet revenue targets.
Karazian said this dynamic means it truly is a buyer’s market.
Subsidizing the use of expensive technology to lock in customers has always been part of Big Tech’s strategy. Remember how cheap it used to be to take an Uber or Lyft? But now that there are few other ride-hailing companies, the cost of getting from one place to another has increased significantly. Harajian says these dynamics don’t work in the AI world because there’s too much competition.
“Meta and Nvidia are also introducing enterprise products with fairly inexpensive models and open source models,” Kharazian said. He also pointed to various open model companies in China and the US, as well as a new company called Jev that is becoming a top performer with Ramp, a lightweight tool for choosing between a number of predetermined outputs.
“Many AI use cases require a simple yes/no,” Kharazian says. “So when you move an already cheap model to Jev, it makes an already cheap model even cheaper.”
The ability to complete workloads at a much lower cost is another factor driving the slowdown in spending. This is partly due to discounts offered by AI companies, but also because companies are becoming more frugal.
“As we get very sophisticated in using AI, we also get better at understanding how to control it,” Karazian says. “You’re more likely to use defaults that prevent you from using very expensive models for long-running tasks. You’re more likely to impose some budget on your employees. You’re more likely to offload some of your spending from standard and lightweight models to even cheaper lightweight models through your router.”
Notion co-founder Akshay Kothari recently posted internal data that aligns with Ramp’s insights. He said simply: “The tokens are decreasing rapidly. The cost is so high.”
“A lot of the assumptions we had about the trajectory of AI spending are starting to break down,” Karazian said. “You can use more AI, but the cost of using it is lower. That’s great for customers, but not necessarily great for model companies. It also raises new questions about revenue trajectory.”
Bloomberg reported Friday that Anthropic’s potential valuation has fallen to just $1.2-1.3 trillion from the $2 trillion estimated in early October. The company is also reportedly expected to receive a high “junk” rating from credit reporting agencies. OpenAI also lowered its proposed valuation from $1.4 trillion to $1.25 trillion, cutting its latest funding goal by about $10 billion.
Karazian says Lamp’s data set is likely to convey many investors’ concerns. But there is one potential bright side to the spending trajectory, albeit a speculative one.
Much of the theory around hyper-growth spending is that AI will undoubtedly introduce new use cases that will accelerate things. AI coding emerged just a year and a half ago as the most commercially viable product that led Anthropic and OpenAI to multi-trillion dollar companies. The hope among investors is a new revolution that will encourage more spending.
Some believe that personal agents will become the next coding agents. OpenAI, SpaceXAI, and Meta have all released their own personal agents, along with startups like Instinct. But Harajian is skeptical that personal agents, which are “not as economically productive as coding,” will provide the boost the industry needs for continued growth.
“We anticipate the introduction of new products that will drive further growth,” Karazian said. “But based on what we can currently observe, what products are available and what is commercially advanced, spending on AI is starting to slow down.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
