Jim Cramer told CNBC’s Squawk on the Street audience on July 9 that Anthropic has emerged as the enterprise AI profit winner, according to 247 Wall St’s coverage of the segment. His argument: as enterprises slash tech budgets, the money is flowing to the model layer, not the application layer that dominated enterprise software for decades.

Cramer’s Enterprise AI Map

Cramer laid out an enterprise AI value chain with four roles, as 247 Wall St reported: Anthropic captures workflow through recurring token revenue at the model layer. Micron supplies the memory hardware. CrowdStrike handles endpoint security. And Salesforce must grow its AI agent revenue faster than customers cut traditional CRM license spending.

“Anthropic is the one that’s actually making a lot of money,” Cramer said, per the report.

The core of his thesis: the model layer, not the application layer, is where enterprise AI profit concentrates. Token inference costs are recurring. Every Slack message routed through Claude, every agent task executed against Claude’s API, generates revenue for Anthropic. Application vendors collect the user relationship but pay the inference bill.

The Inversion

For three decades, the application layer captured the lion’s share of enterprise software value. Salesforce, SAP, Oracle, and ServiceNow sold seats, licenses, and platform access. The infrastructure underneath was a cost center. Databases, compute, and storage were commoditized inputs.

AI agents invert this. The model is the product. The application becomes a routing layer that sends prompts to a model provider and renders responses. The more an enterprise deploys agents, the more token volume flows to model labs. Application vendors get squeezed between customer demands for lower per-seat pricing and rising inference costs they cannot fully control.

Salesforce’s own behavior supports the observation. The company is actively promoting Anthropic’s Claude Tag agent inside Slack, its $27.7 billion acquisition. Employees have expressed confusion about why Salesforce is embedding a competitor’s AI agent alongside its own Agentforce and Slackbot products. The answer is practical: Claude’s model quality drives engagement that Salesforce’s own models cannot yet match. But every Claude Tag invocation is revenue for Anthropic, not Salesforce.

What the Thesis Gets Right

The capital flow pattern is real. Enterprise IT budgets are increasingly denominated in tokens, not seats. A company that previously paid Salesforce $150 per user per month for CRM access now pays that plus a variable inference cost that scales with agent activity. The inference cost flows to model providers.

Anthropic has positioned itself to capture this flow. Claude powers agents across Salesforce, Amazon (via Bedrock), and dozens of enterprise platforms that chose Claude as their primary model. Each integration creates a token revenue stream that grows with usage, not headcount.

For agent builders, this has a direct consequence: if you are building on top of Claude, your margins are structurally linked to Anthropic’s pricing power. Every Claude API price increase or rate limit change hits your unit economics directly.

What It Misses

Cramer’s framing simplifies a messier reality. Anthropic is not yet profitable. The company’s last reported revenue run rate was approximately $2 billion annually, but its compute costs are enormous. Model training runs cost hundreds of millions of dollars. Inference infrastructure requires continuous capital expenditure. Revenue is not profit.

The model layer is also not a monopoly. OpenAI’s GPT-5.6 Sol launched this week with aggressive enterprise pricing. Meta’s Muse Spark 1.1 just went to public API with token costs roughly 70% below Claude’s equivalent tiers. Chinese labs like DeepSeek and Zhipu AI are offering comparable quality at 60-90% discounts. If the model becomes commoditized, the profit concentration Cramer describes evaporates.

There is also the question of whether application vendors will accept permanent margin compression. Salesforce, Microsoft, and Google are all investing heavily in proprietary models. Microsoft has already begun replacing OpenAI and Anthropic models with internally built MAI models across Excel and Outlook. If platforms can train competitive models in-house, the token revenue redirects internally.

The Uncomfortable Position for Builders

If Cramer’s thesis holds, builders face a strategic question: should you build on top of a model provider that captures an increasing share of your value chain?

The alternative is emerging. Open-weight models from Meta, Alibaba, and Mistral allow self-hosted inference at fixed compute cost rather than variable token cost. The trade-off is quality and reliability, but the gap is closing. Agent orchestration frameworks like OpenClaw, LangGraph, and CrewAI are model-agnostic by design, letting operators swap providers without rewriting their agent logic.

Cramer may be right that Anthropic is winning today. The question for anyone building enterprise AI agents is whether that winning position is structural or temporary. Token pricing pressure from open-weight competitors suggests the model layer’s margin advantage has a shelf life. The application vendors are not going to accept being dumb pipes forever.

The model layer captured the margin. The question is how long it keeps it.