The quarterly report Ramp publishes to show which software tools are winning corporate budgets carried a first on Monday. DeepSeek, the Chinese AI lab, topped the “breakthrough growth” ranking, the first time a company outside the usual American and European names has led that chart, according to Ramp.
Ramp defines “breakthrough growth” as paid usage that accelerated fastest over the quarter, and its lists have tracked the rise of everything from design software to internal chat tools in past editions. Finance chiefs read the report as a proxy for where budgets will move next, which is what makes the DeepSeek entry notable: it is a signal from the CFO’s office, not from a developer forum.
Ramp’s chief economist said the shift is no longer about experimentation. American companies that once downloaded DeepSeek’s open-source models and ran them on their own servers are now paying the company directly, with data moving across servers in China. He called it the most visible sign yet that U.S. businesses are hunting for cheaper alternatives to OpenAI and Anthropic.
The numbers behind that conclusion are stark. A Bain survey cited in the report covered 951 companies with annual revenue above $100 million and found cumulative corporate AI spending has passed $1 trillion. The cost savings those companies expected, Bain found, have come in well below plan. The gap between spending and payoff is what is pushing procurement teams to price-shop, the survey’s authors said.
The Bain finding deserves attention, analysts said. Companies that rushed to deploy AI assistants found the models cheaper to buy than the engineering needed to make them useful; integration, data plumbing and retraining absorbed the savings. With the easy wins already booked, unit price has become the main lever procurement has left to pull.
Some usage figures in the Ramp report are blunt. Uber burned through its full token budget for the year in the first four months, according to the report’s account of corporate usage patterns. Salesforce is on track to pay Anthropic about $300 million this year for model access. At that scale of consumption, a tenfold difference in unit price stops being an abstract question.
DeepSeek’s pricing makes the arithmetic obvious. V4-Pro, its flagship model, costs roughly one-tenth of what GPT-5.5 charges for a comparable call, and a price cut on May 22 pushed the ratio lower still. On OpenRouter, the marketplace where developers compare models, DeepSeek’s V4-Flash handled 3.43 trillion tokens in a single week, the most of any model in the world. Analysts who track the marketplace said the volume is a leading indicator of enterprise spending, because developers test on OpenRouter before finance teams sign contracts.
Security teams have not gone quiet. Compliance officers still raise questions about where data travels when a U.S. company pays DeepSeek directly, and the data does travel: the models run on Chinese infrastructure, and the company has been explicit about that. Procurement decisions, though, are being made by finance as much as by security, and the finance math is moving in one direction.
What is happening, analysts said, is a reordering of the AI market’s economics. OpenAI and Anthropic built their businesses selling frontier models at premium prices, and both have raised enormous sums on the assumption that corporate customers will keep paying for the best available capability. DeepSeek’s rise tests that assumption from below, the way discount carriers test network airlines: same destination, different fare, and a growing share of price-sensitive passengers.
There are limits to the analogy. Enterprises that run regulated workloads, such as banks, hospitals and government contractors, often cannot route data through Chinese servers regardless of price, and several large U.S. companies have told investors they will not use DeepSeek for that reason. The Bain survey suggests those restrictions apply to a minority of the companies it studied, and the majority is behaving like the Ramp chart suggests: trying the cheaper option first.
DeepSeek’s open-weight strategy helped create the trust path. Because the models could be downloaded and tested internally, companies got comfortable with the capability before any money changed hands. Paying for the hosted API was a small step after that, and it converted DeepSeek from a curiosity in developer forums into a line item in software budgets.
The pattern shows up in how fast the numbers moved. DeepSeek’s weekly token volume on OpenRouter passed its rivals months after its first models appeared, and the climb to the top of a software-trends chart moves the signal out of the developer community and into the finance office. Ramp’s report is read by finance teams precisely because it tracks what companies pay for, not what they announce.
For OpenAI and Anthropic, the response has been price cuts and volume discounts, and both have signaled they will keep investing in models that justify their rates. Whether that holds depends on the question the Bain data raises: if a trillion dollars of spending produced less savings than planned, will boards approve the next trillion at the same prices?
The report’s authors put it in terms finance understands. The security objections are real, the report concluded, but they are being outweighed, one invoice at a time, by the arithmetic of what frontier AI now costs. Companies that pay for AI with their own money are voting with the bill.
