Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
Sep 15, 2026, 5:00 AM · TechCrunch
At Dreamforce, Salesforce unveiled Koa — a Nemotron-based reasoning model post-trained for sales and support without touching customer data — and aimed it at Claude and ChatGPT token bills.
Why it matters
Salesforce’s first reasoning model, Koa, is built on Nvidia’s open-weight Nemotron and post-trained with Nvidia for sales, marketing, and customer-support work. It ships inside Agentforce as an alternative to routing long multi-step tasks to frontier APIs like Claude or ChatGPT.
EVP Jayesh Govindarajan said Salesforce wanted a sovereign American pre-trained base with clear data provenance — something he said they lacked until Nemotron, contrasting with uncertainty about what Qwen trains on. Training used synthetic customer-service and sales personas, not real customer data.
We’re reading Koa as enterprise AI peeling away from “upload everything to the frontier lab” as the default.
From the desk
We’re watching the buyer rebellion get productized. Frontier labs want enterprises to pour files, prompts, and feedback into closed agents at millions a year. Salesforce is answering with open weights, task-specific reasoning, lower token burn, gateway routing, and security that stays inside Salesforce’s system of record.
Nvidia’s Kari Ann Briski frames the pitch as sovereign AI, fast time-to-first-token, and efficient reasoning — the tokenomics trifecta. That is a direct attack on the margin structure of closed reasoning APIs for rote CRM work.
Salesforce is not dumping Anthropic. It also announced Claudeforce so companies can use Claude as the interface while data remains in Salesforce. Useful specialized models that keep customer data local deserve adoption for the workloads they own. The downside if this scales is a fragmented agent stack where every SaaS vendor trains a pet reasoning model — great for switching costs, messy for governance. I’m watching whether Koa actually displaces frontier calls on multi-step Agentforce jobs in production spend reports.
Context
Before Koa, Agentforce relied on frontier providers through an AI gateway when agents needed long-running reasoning. Salesforce already ships many small task-specific models; reasoning was the missing piece.
Who feels it
- Enterprise AI buyers
- A credible path to cut frontier token spend on sales and support agents without abandoning closed models entirely.
- Frontier labs
- Open-weight sovereign bases plus vertical post-training threaten high-margin reasoning traffic.
- Security and compliance teams
- Synthetic-data post-training and in-Salesforce residency are the selling points to pressure-test.
What to watch
- Production share of Agentforce calls routed to Koa versus Claude/ChatGPT.
- Whether other SaaS platforms announce Nemotron-style vertical reasoning models.
- Customer reports on token cost and quality versus frontier baselines.
Companies: NVIDIA