The numbers don't lie, but executives sometimes do. In this case, Marc Benioff is being brutally honest about Salesforce's AI spending—and it's staggering.
The enterprise software giant will spend $300 million on Anthropic tokens this year, Benioff revealed on the All-In podcast, primarily for coding applications. That's not capital expenditure on servers that depreciate over years. That's pure operating expense—tokens consumed and gone, like jet fuel burned at altitude.
"These coding agents are awesome. Anthropic is awesome," Benioff said with characteristic enthusiasm. "I am going to probably use $300 million of Anthropic tokens this year at Salesforce. Coding."
This is a watershed moment for how AI costs hit corporate P&L statements. While competitors like Microsoft and Google build their own models and can bury costs in R&D, Salesforce is essentially renting AI by the drink. Every API call is a line item. Every coding task solved by Claude is metered and billed.
The business case, according to Benioff, is productivity gains north of 30% among Salesforce's approximately 15,000 engineers. The company froze software engineering hiring in January 2025 after seeing those results—a decision that speaks louder than any earnings call platitude about "AI transformation."
"I can do things that I just could not do before. I can go faster than ever before," Benioff explained, crediting Anthropic's focus on coding agents as the right strategic bet.
The financial payoff appears real. Agentforce, Salesforce's AI agent product line, has reached $800 million in annual recurring revenue with 29,000 closed deals. That's actual revenue, not pilot programs or proof-of-concepts that never convert.
But here's the question analysts should be asking: What happens when every enterprise software company follows this playbook? If Salesforce, Adobe, ServiceNow, and SAP all start burning hundreds of millions on AI tokens, that's a permanent structural cost increase in an industry that Wall Street values on margin expansion.
Benioff also revealed plans for coding tools within Slack, with announcements coming soon. And he showed some cost discipline, advocating for routing simpler tasks to cheaper models while reserving Claude for complex reasoning—the kind of optimization CFOs will demand as these bills come due.
The numbers are real. The productivity gains appear real. But $300 million a year in tokens? That's not a rounding error. That's a bet that AI-augmented engineers can generate enough incremental revenue to justify what amounts to a permanent tax on software development.
Cui bono? Anthropic, certainly. And if Benioff is right about those productivity gains, Salesforce shareholders too. But the broader question is whether AI becomes a competitive necessity that compresses margins across the industry—great for customers, challenging for investors who've gotten used to software economics.
The era of cheap code is over. The era of AI-powered code is expensive in a whole new way.
