Jensen Huang Says AI Will Let Us “Know Everything and Do Anything” as NVIDIA and Salesforce Unveil Koa
NVIDIA CEO Jensen Huang used the Salesforce Dreamforce stage Tuesday to paint an ambitious picture of the AI-powered enterprise, saying artificial intelligence could eventually allow people and organizations to “know everything and do anything.”
Huang appeared alongside Salesforce CEO Marc Benioff at San Francisco’s Moscone Center as the companies announced Koa, Salesforce’s first CRM reasoning model for Agentforce. Koa was built by post-training NVIDIA’s Nemotron 3 Super model using synthetic enterprise scenarios based on nearly three decades of Salesforce CRM knowledge.
The announcement puts specialized AI reasoning at the center of Salesforce’s push toward what it calls the “Agentic Enterprise,” where AI agents can reason through multistep business processes and take actions rather than simply generate text.
Jensen Huang’s vision for the AI era
Huang framed the development as the latest step in a much longer technological transformation.
“With electricity, we could power everything,” Huang said, describing earlier industrial revolutions. “With the internet, you can find anything. And now, with artificial intelligence as an infrastructure layer across the planet, we can know everything and do anything.”
He also encouraged companies to embrace AI rather than wait on the sidelines.
Huang argued that every software company and enterprise will increasingly become an AI company, with AI agents taking on more complex tasks across organizations.
That vision is reflected in Salesforce’s Koa announcement, which focuses specifically on AI that can reason through enterprise workflows and use the appropriate tools to complete them.
Salesforce introduces Koa CRM reasoning model
Koa is designed specifically for Salesforce’s Agentforce platform.
Rather than training it as a general-purpose chatbot, Salesforce post-trained NVIDIA Nemotron 3 Super on a proprietary synthetic dataset representing enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel.
The training process used supervised fine-tuning and reinforcement learning through NVIDIA’s NeMo RL, NeMo Gym and NeMo AutoModel technologies.
The goal is for Koa to do more than produce an answer. Salesforce says the model is designed to reason through a task step by step and select the appropriate actions and tools.
In Salesforce’s CRM benchmark, which includes tasks such as updating opportunities, routing customer-service cases and scheduling follow-ups, the company says Koa matches or exceeds leading model performance on CRM actions while producing three times fewer errors.
No customer data used to train Koa
One of the major selling points of Koa is control over data and the model itself.
Salesforce says no customer data was used to train the reasoning model. Instead, the training corpus consists of synthetic scenarios designed to reproduce the types of reasoning, tool use and decision-making required by CRM agents.
Salesforce also controls Koa’s model weights and performs post-training and inference within its own infrastructure, keeping the model inside what the company describes as its trust boundary.
That approach could be particularly important for businesses handling sensitive customer, financial, healthcare or operational information.
Salesforce and NVIDIA are also extending the same model-control approach to Missionforce, where NVIDIA models and accelerated computing will support specialized AI deployments for government and regulated organizations, including private clouds and air-gapped environments.
Jensen Huang says AI safety is an engineering problem
AI safety was another major topic during Huang’s conversation with Benioff.
Huang described safety as a top priority but argued that companies do not necessarily have to choose between rapid innovation and safer products.
“Safety is paramount in a lot of ways. It’s job one,” Huang said. He added that if a company is not confident in a product’s functionality, capability or safety, it should not release it.
His broader message was that companies should move quickly while being prepared to pause when a system is not under adequate control.
Huang also pushed back on the idea that AI will simply eliminate software development or broadly make human work irrelevant. Instead, he emphasized the potential productivity gains from combining people with increasingly capable AI systems.
Koa is already being tested
Koa is not just a future concept.
Salesforce says it is already being used internally, including through an employee agent in Slack that can help workers find information and complete everyday tasks. The model is now moving toward customer pilots.
The pilot group includes Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero.
Salesforce expects Koa to become generally available in U.S. regions in winter 2026.
For businesses, the significance is less about another chatbot and more about whether specialized reasoning models can reliably execute real operational work.
NVIDIA is betting on specialized AI
The Salesforce partnership also highlights NVIDIA’s broader strategy around open models.
Nemotron 3 Super is part of NVIDIA’s Nemotron family of open models designed for agentic reasoning and multi-agent applications. NVIDIA says Nemotron 3 Super uses a hybrid mixture-of-experts architecture and is designed to support complex reasoning while allowing organizations to customize and deploy the model on their own infrastructure.
For Salesforce, that openness provides the foundation to adapt a model to its own enterprise expertise instead of relying exclusively on a general-purpose AI model.
For NVIDIA, the partnership demonstrates another way its AI technology can become part of enterprise software infrastructure beyond the traditional GPU market.
Huang’s Dreamforce appearance therefore offered a glimpse of the direction both companies are pursuing: AI systems that are increasingly specialized, capable of reasoning through business processes and able to act within controlled enterprise environments.
The immediate test will be whether Koa can translate its benchmark results into reliable performance across real customer workflows as Salesforce expands its pilots and moves toward broader availability.
