Ramp launches token spend management as AI costs hit 10% of payroll, growing 21x in a year

Ramp launches token spend management

Ramp's AI token spend hit nearly 10% of payroll in May 2026. That number, which Eric Glyman describes as both a product signal and a warning, is what pushed the company to build and ship tokenspend.fm, a token spend management tool that lets any company link up API keys and start monitoring, analyzing, and cutting AI spending within minutes.

The growth on the customer side is sharper: token spend across Ramp's customer base grew 21x over the past year. Glyman is clear that the goal isn't to suppress spending — "people aren't saying turn it off," he says — but to make sure every dollar spent on tokens is doing real work.

“Over the last year, the last full months alone, ramp customer spend on tokens has grown by 21 times... A few years ago, AI spend was a routing error to I think in May, it hit almost 10% of our payroll spend, the equivalent was on tokens on a payroll.”

What the product does

Companies plug in their API keys across providers — OpenAI, Anthropic, Gemini, Cursor, open-source models — and get a real-time consolidated view of what's being spent, by whom, and on what. The core value in the first version is visibility: knowing today, not at month-end billing, that a single employee has burned through $800 in an hour rather than their normal $1,000 a month. The product surfaces unusual spikes, flags inefficient configurations like leaving fast mode enabled (which can be several times more expensive), and highlights whether caching is being used. Glyman frames this as a direct analogy to how Ramp originally approached corporate cards — alerting first, then controlling.

The frontier vs. open source question

Glyman pushes back on the assumption that cheaper-per-call models always reduce total spend. In Ramp's own benchmarks, using an older, cheaper Anthropic model required more agent calls and longer reasoning chains, making the total cost higher than using a frontier model that solved the task faster. His framing: a smarter model's higher per-token cost can be offset by fewer tokens needed. The more durable optimization strategy, he argues, is benchmarking your own workload and routing tasks by complexity — using a small fine-tuned model for repetitive, well-defined tasks like customer service or accounting, and reserving frontier models for genuinely complex reasoning.

Visibility drives behavior

Even before any controls are applied, exposure to spend data changes behavior. Glyman draws on Ramp's card business to make the point: when employees are told once that a purchase was out of policy, out-of-policy spend drops sharply. The same dynamic is showing up in token spend — showing individuals their own AI consumption is already reducing waste. Someone checking the weather with an AI model at meaningful cost isn't doing it maliciously; they just didn't know.

The Ramp Econ Lab angle

Ramp tracks roughly 1% of all U.S. corporate spend, and Glyman argues that real-time visibility into that data lets companies adjust strategy faster than conventional economic indicators, which typically lag by a quarter. He cites AI spend as a case in point: it has moved from a rounding error to what he estimates could be close to 1% of U.S. GDP within the next twelve months. The research output from Ramp's economics lab has placed the company in the Financial Times and Wall Street Journal, which Glyman says matters competitively — when prospects already trust Ramp's analysis before a sales conversation begins, the deal moves faster.