Tokenmaxxing: why Salesforce is spending $300M on Anthropic (and your budget is next)

Salesforce is on track to spend roughly $300 million on Anthropic tokens this year. Uber blew through its annual AI budget by April. A four-person Silicon Valley startup is paying Anthropic $125k per month. The token budget is the new data-centre budget — and almost everyone is underestimating it.

Salesforce: $300 million, and that's just the start

On the All-In podcast last week, Marc Benioff said Salesforce is on track to spend roughly $300 million on Anthropic tokens this year. For context: that's about 4.5% of the $6.7 billion Salesforce spent on cost of revenues last year — the pot every third-party tech supplier is paid from. One AI vendor is now taking a serious bite.

The vast majority of that flows to coding agents. "These coding agents are awesome," Benioff said. "Everything's going to be cheaper to make. It's more efficient. I can do things that I just could not do before." The ROI justifies the spend — for now. But in the same breath he hinted he wants to wean Salesforce off that exclusive Anthropic dependency:

"The vast majority of those tokens don't need to go to Anthropic. There needs to be some intermediary layer that's saying, oh, that one has to go to Anthropic, but these ones can be handled by smaller models."

Benioff speculated that "a hot new company" would soon emerge to build it. It already exists: OpenRouter recently raised $120 million from an Alphabet fund at a $1.3 billion valuation. We wrote earlier about Cortecs.ai as a European, GDPR-native alternative in that same router category.

Uber: annual budget gone by April

Salesforce isn't alone. Uber blew through its annual AI budget in April 2026. The internal post-mortem that leaked wasn't "we under-budgeted", it was "we didn't understand what an agent costs once it loops more than twice per ticket". A loop with three tool calls and a retry strategy can cost ten to fifteen times more in production than the same flow as a single-shot prompt.

Helloprint, Swan and the $125k startup

The numbers scale with the company:

• Helloprint does €80M in revenue and recently disclosed €25k in annual AI costs — modest, but coupled to a headcount reduction from ~300 to ~120 FTE. Here the tokens are literally a substitute for people.
• Swan AI, an AI fintech, reportedly spends $113k per month at Anthropic. That's $1.3M per year, with one vendor, in one company.
• A four-person Silicon Valley startup is burning $125k per month at Anthropic. Per person, that's $375k in tokens per year. More than their salary.
• Jensen Huang recently said matter-of-factly that his most productive engineers cost "$250k per year in AI tools". Solberg's CEO went further, saying his team is burning $4M per month in tokens.

Recent figures put Anthropic on a run rate approaching $30 billion. That money is coming from somewhere. It's coming from these companies.

What is actually happening here

Three things are running in parallel:

1. The budgeting model is broken. AI budgets are still planned like SaaS budgets — a fixed amount per seat per month. But agent tokens scale with workload, not seats. One engineer with a tight Claude Code loop can burn more tokens in a week than a whole team burns in a month.

2. ROI masks the inefficiency. Benioff is right: these agents are awesome and they make work cheaper. That's exactly why nobody is challenging the token spend hard. As long as every €1 of tokens saves €3 of human labour, it feels like a good deal. But that's not a reason to route everything to the most expensive model.

3. Routing is becoming its own discipline. Not every prompt belongs on Sonnet 4.5. A classification step is fine on Haiku or a local Qwen. A drafting agent for boring emails can run on a cheaper Gemini model. The "intermediary layer" Benioff described isn't a future vision — we build it for clients today in n8n or OpenClaw, with routing rules driven by task, sensitivity and cost ceiling.

The Ramp AI Index in context

The Ramp AI Index recently showed that 50.4% of U.S. companies now hold a paid AI subscription, with spend rising sharply every quarter. What the index doesn't show is how much of that is agent tokens versus chat seats. The token portion is almost certainly growing faster than any other slice — because agents, by definition, consume more tokens than a human who occasionally opens a chat window.

How we deal with this in practice

A few things we build in by default the moment an agent goes into production:

• Per-tool and per-run cost tracking in the orchestration layer (n8n or a Python pipeline). Not just $-totals at month end, but per tool call, per agent step. That's how you discover that one retry storm caused 80% of your weekly bill.
• Model selection as config, not as hardcode. A classification step, an extraction step and a copy-edit step are three different cost categories and should be able to call three different models. Switching between GPT-5-mini, Claude Haiku, and a local Qwen via Cortecs or OpenRouter should not require a refactor.
• A hard cost ceiling per agent per day. No circuit breaker = open tap. For long-running loops we routinely add a ceiling that stops the agent or downgrades it to a cheaper model once a threshold is hit.
• On-prem for whatever is repetitive and sensitive. For most clients 70-80% of the tokens go into dull, repetitive work that runs just fine on a local Llama or Qwen via OpenClaw. Not for the wow work — for the bulk underneath it.

Finally

The irony of Benioff's quote — "I think that's just the moment of time we're in right now" — is that he's right, but only once a routing layer sits in between. Tokens don't get cheaper because Anthropic makes them cheaper. They get cheaper because you learn to send the right prompt to the right model.

CFOs haven't put a red pen to the AI budget yet. That's a matter of time. Whoever sets up their architecture for model flexibility now — instead of blindly funnelling everything to one vendor — is the one who'll be able to take the corner without panicking.