AI usage rises while spending stalls, Ramp data shows

TechCrunch reports on Ramp data showing a split in the AI market: usage keeps climbing, but spending has stalled. The same data points to price cuts and more disciplined cost controls among the heaviest spenders as drivers of that gap. The report does not specify which vendors, contract sizes or time periods underpin the figures, nor how the spending trend is measured.

The timing matters because the item frames this against Anthropic's approach to an IPO, which makes the spending signal a question about the economics of AI supply rather than a curiosity about buyer habits.

What a flat spend line actually signals

Usage and spend are two different metrics, and they can move in opposite directions for mundane reasons. If unit prices fall, a company can send more requests through a model and still write a smaller cheque. If buyers consolidate vendors, route simpler tasks to cheaper models, or add caching and batching, the volume curve keeps rising while the invoice flattens.

For anyone building a product on top of model APIs, that is the relevant reading. Growth in tokens processed does not automatically translate into growth in revenue per customer. It is also well-established context that inference prices across major providers have fallen repeatedly over the past few years, so a divergence between usage and spend is plausible without any drop in demand.

What the report leaves open is the split between price-driven and discipline-driven savings. That distinction changes the outlook considerably: falling prices are a market-wide repricing, while cost controls among the biggest spenders suggest buyers are now treating AI as a line item to be managed rather than an experiment to be funded.

What this means for traffic and monetisation teams

Most teams reading this are not selling foundation models, but they are buying them, and their unit economics sit inside the same shift. A few things are worth checking this week.

• Pull the last two quarters of AI API and tooling invoices and compare them with actual usage volume, so you know whether your cost per task is falling or flat.
• Identify the workflows where a smaller or cheaper model would pass your quality bar, and test one of them.
• Check whether per-seat AI add-ons in your stack are still used enough to justify renewal.

On the monetisation side, the lesson cuts the other way. If your product bundles AI features, buyers now have benchmarks for what inference should cost and will push back on pricing that assumes 2023-level margins. Features priced as a premium because they use a model are exposed; features priced on the outcome they produce are not.

On acquisition, cheaper inference widens the range of personalisation and creative variants you can afford to test, which is a quiet advantage for teams that instrument their funnels well.

The macro question the data raises is whether AI revenue growth can keep pace with usage if buyers keep squeezing unit costs. The report does not answer that, and neither can a single dataset. But it gives a reason to treat "more usage" and "more spend" as separate assumptions in any plan built on this market.