Salesforce and Nvidia jointly unveiled Koa at Dreamforce on 15 September — a reasoning model post-trained on Nvidia's Nemotron open-weight foundation and purpose-built for enterprise sales, marketing, and customer support workflows. Koa never ingested real customer data during training, uses fewer tokens per task than frontier models, and runs inside Salesforce's own infrastructure rather than routing requests to external API endpoints. Salesforce's announcement framing was pointed: "Reasoning has always been something we've relied on frontier model providers for. Until now."

The competitive implication is explicit. Salesforce processes customer interactions, sales records, and proprietary CRM data for hundreds of thousands of enterprises. The prior architecture — sending that data to OpenAI or Anthropic APIs for reasoning tasks — required trust in a third party's data handling practices and accumulated per-token costs at scale. Koa internalises that reasoning capability. At Salesforce's user base, a meaningful shift from external frontier APIs to in-house reasoning models represents a structural challenge to the enterprise API revenue thesis that underpins current frontier lab valuations — and a template that other large enterprises with proprietary data environments and sufficient engineering capacity will evaluate. The "Claudeforce" multi-model architecture Salesforce runs in parallel, routing different task types to different models, extends the pattern: enterprise AI at scale is increasingly a portfolio of models, not a single API endpoint.