Jensen Huang unveils Cua: Salesforce's first reasoning model, built on Nemotron 3 Super

NVIDIA founder and chief executive Jensen Huang took the Dreamforce stage alongside Marc Benioff to introduce Cua, Salesforce's first reasoning model for CRM. The model was created by post-training Nemotron 3 Super on synthetic data drawn from nearly three decades of enterprise CRM deployments. The joint appearance went beyond a handshake: Huang walked into the Moscone Center crowd, microphone in hand, while Benioff followed with "we're all following you, brother."

From electricity to the internet, and now planetary AI infrastructure

Huang framed the moment as a direct continuation of industrial revolutions. Electricity made it possible to power everything; the internet made it possible to find everything; artificial intelligence as a global infrastructure layer now makes it possible to "know everything and do everything." That line led straight to the announcement: Cua is built on an open NVIDIA model, which let Salesforce fine-tune it inside its own infrastructure, control the weights, and guarantee that no byte of customer data was used in training or inference — a point Salesforce platform and engineering president Rohan Kumar stressed explicitly before Huang appeared.

Safety as an engineering problem, not a barrier

When Benioff asked about AI safety, Huang was direct: "Safety is above all, it's mission number one, but it's an engineering problem." In his view, innovation and safety are not a false choice; you can move fast and stop only when the product feels out of control or unsafe. He dismissed the narrative that AI will kill the software industry or threaten jobs, arguing that productivity gains open an unlimited ceiling: "Work with AI, don't get left behind."

Trained on synthetic scenarios across 14 industries

Cua was developed through supervised fine-tuning and reinforcement learning using NVIDIA's NeMo RL, NeMo Gym, and Neo AutoModel tools. The training corpus covers synthetic enterprise scenarios across 14-plus industries — manufacturing, financial services, healthcare, travel, and others. On Salesforce's internal CRM Bench, which includes real-world tasks such as updating an opportunity, routing a case, or scheduling a follow-up, Cua already matches or outperforms leading models.

The trend: from closed models to customized open ones

Huang noted that open-model adoption has jumped from roughly 30% early last year to about 70% today. "Every software company becomes an AI company, and every organization will become an agentic organization," he said. NVIDIA itself runs on Salesforce for sales, service, marketing, and operations, and is already piloting Agentforce for customer-support workflows. Benioff closed: "The sky's the limit, especially for our trailblazers — the technology can collaborate with them in entirely new ways."