Nvidia CEO Jensen Huang joined Salesforce CEO Marc Benioff onstage at Dreamforce on Tuesday to unveil Koa, Salesforce's first CRM reasoning model, post-trained from Nvidia's Nemotron 3 Super on 27 years of Salesforce CRM data. In Salesforce's internal CRM Bench evaluation — a suite covering tasks like updating an opportunity, routing a case, and scheduling a follow-up — Koa already matches or exceeds leading models with 3x fewer errors. The model runs entirely inside Salesforce's own infrastructure, with Salesforce controlling the weights.
Huang used the appearance at Moscone Center to frame AI as the next horizontal layer beneath the economy, alongside electricity and the internet. He walked into the crowd with a handheld mic, Benioff alongside, in a staging that made the partnership look less like a vendor deal and more like a joint pitch to Salesforce's enterprise base.
The choice of Nemotron matters because it is open. Salesforce fine-tuned and now runs the model on its own systems, and Nvidia's president of platform and engineering Rohan Kumar told the Dreamforce audience that no customer data touched training or inference. That structure is the crux of Salesforce's pitch to regulated buyers who have resisted sending CRM records to third-party model APIs.
Key facts
- 01Salesforce and Nvidia launched Koa, a CRM reasoning model post-trained from Nemotron 3 Super on 27 years of Salesforce CRM intelligence.
- 02Koa matches or exceeds leading models on Salesforce's CRM Bench with 3x fewer errors on real-world CRM actions.
- 03Open-model adoption climbed from 30% at the start of last year to roughly 70% today, per Huang.
- 04Customer pilots begin in October with Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero; GA is expected winter 2026 in U.S. regions.
- 05The training corpus spans synthetic enterprise scenarios across 14+ industries, built with NeMo RL, NeMo Gym and NeMo AutoModel.
Huang cited a broader shift to open weights: adoption of open models has moved from roughly 30% at the start of last year to about 70% now. He argued that closed-model consumption is still growing exponentially, but that enterprises are increasingly building custom models on top of open bases — a trend that directly benefits Nvidia's Nemotron family and its NeMo tooling stack.
Koa was built with supervised fine-tuning and reinforcement learning using NeMo RL, NeMo Gym, and NeMo AutoModel. Its training corpus draws from synthetic enterprise scenarios spanning more than 14 industries, including manufacturing, financial services, healthcare, and travel — sectors where Salesforce already has entrenched deployments and where hallucinations in customer-facing agents carry real cost.
“Not a single byte of customer data was used.”— Rohan Kumar, Salesforce president of platform and engineering
The model is already live inside Salesforce, powering an employee agent in Slack. Customer pilots begin in October as a selectable model inside Agentforce, with Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine, and Xero as the initial cohort. General availability in U.S. regions is expected in winter 2026.
Benioff steered the conversation to safety, and Huang answered squarely. He called safety "job one" but reframed it as engineering rather than governance, arguing that vendors who lack confidence in a product's functionality or safety simply should not ship it. He rejected the framing that speed and safety are in tension.
Huang also pushed back on the argument that AI threatens software jobs, a line he has repeated across recent appearances, including his speakerphone cameo with President Trump at the All-In Summit that we covered earlier this month. His Dreamforce version was blunter: engage AI or get left behind. "Every company, every enterprise, every country would become an AI company," he said. "It's going to be an agentic enterprise."
The Koa launch slots into a run of Nemotron-based enterprise wins Nvidia has been stacking. It follows Nvidia's DSX announcement of a 24% token-per-watt gain in AI factory throughput, and reinforces a pattern where Nvidia supplies the base model and the tooling while a partner owns the vertical data and the customer relationship. For Salesforce, it is a way to answer OpenAI and Anthropic without sending customer records outside its trust boundary.
The open question is whether Koa's 3x error reduction on CRM Bench — a benchmark Salesforce itself built — will hold up against the same tasks measured by third parties or against buyers running head-to-head pilots with Claude, GPT, or Gemini-powered agents. Vendor-authored benchmarks routinely flatter the sponsor, and Agentforce's early traction has been mixed on real-world deployment metrics. The October pilots with named customers will be the first honest read.
The strategic read for the AI market is that Nvidia is quietly becoming a model company as well as a chip company, and its distribution channel is every enterprise software vendor that wants to run AI on its own metal. Selling Nemotron to Salesforce is worth more than a one-off licensing fee — it locks Salesforce's agent roadmap onto Nvidia's stack and gives Nvidia a reference customer with 27 years of data and hundreds of thousands of enterprise seats. That is a moat competitors building closed frontier models cannot easily cross.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
