Salesforce just admitted Claude token spend is eating its margin. $300M/year on Anthropic and they're pulling back.

Salesforce didn't raise margin guidance this year.

Not because revenue fell. Revenue hit $11.3B last quarter, up 11%.

The reason was Claude.

Deputy CFO Mike Spencer told the Deutsche Bank Tech Conference that Salesforce "unleashed Claude in R&D" six months ago and the token spend was so large it prevented them from raising GAAP margin guidance above 20.1%.

Marc Benioff previously said they planned to spend $300 million with Anthropic in 2026.

Now they're entering what Spencer called "refinement mode."

Translation: they're picking cheaper models for cheaper tasks. OpenAI for some jobs. Cursor for others. Claude for the heavy lifting. Grok for experiments.

"You don't need to use the latest and greatest model for every single task," Spencer said.

This is the moment every enterprise CFO has been dreading. Your AI vendor is telling you that throwing frontier models at every problem is a money pit.

Salesforce's AI ARR is approaching $4B. Agentforce alone hit $1.5B in ARR, up 240% year-over-year. The revenue is real.

But the cost of generating it is compressing margins in a way that GAAP accounting can no longer hide.

Audit your own token spend. Map which models your teams are using for which tasks. The companies that optimize model selection by workload will save 30-60% on inference costs. The ones that don't will watch their margins evaporate like Salesforce almost did.