Nvidia Open Model-Based ‘Core’ Unveiled
CRM Specialized in Business to optimize cost and performance
Jensen Huang “From 30% to 70% of open models”
Amodei “Using only 5% to 10% of AI potential value”
Amid fierce competition to create the ‘smartest artificial intelligence (AI), a different trend is emerging in the corporate AI market. Instead of entrusting all tasks to one of the largest and most powerful AI models, there is a growing movement to bring relatively small and open AI models and specialize in their work. This is as companies that apply AI to actual tasks begin to consider not only performance but also cost, security, and suitability for work.
Salesforce unveiled its own customer relationship management (CRM) inference model “Koa” at its annual event “Dreamforce 2026” in San Francisco on the 15th (local time). CRM is a system in which companies manage customer information and sales, marketing, and service processes.
Core was developed by Salesforce to suit CRM work based on Nvidia’s open AI model ‘Nemotron’. It analyzes complex sales opportunities to determine what to do next, or to find the next step to solve problems in the customer service process. The focus was on performing certain tasks of the company rather than AI, which answers all general-purpose questions.
“Customers want to choose the best AI model in terms of price and performance based on the results they want,” said Rohan Kumar, President of Salesforce Platform and Engineering. Salesforce said it trained Core with synthetic data, not actual data from customers. This is to prevent companies’ sensitive customer information from being used to learn AI models.
This is an example of how “appropriate performance” is becoming more important than “best performance” in the corporate AI market. State-of-the-art AI models developed by OpenAI and Antropic can perform complex reasoning and various tasks, but require as much computing resources. On the other hand, if a company builds a model that fits a specific task directly, it can utilize AI by focusing on the necessary functions.
Nvidia CEO Jensen Huang stressed on the stage that these changes are already progressing rapidly. According to CEO Hwang, the size of tokens generated by AI has increased by about 25 times over the past year. In particular, he explained that the share of open models among all AI uses increased from about 30% to 70%.
CEO Hwang said, “It means that people are using closed models exponentially and creating their own customized AI at the same time,” adding, “All software companies will become AI companies, and all companies will become companies that use agents and AI.”
This means that companies’ interest in AI is shifting from “how smart the model is” to “how much value it creates in my work.” Dario Amodei, CEO of Antropics, also diagnosed that AI technology is not being fully utilized in the corporate field.
CEO Amodei said, “Even assuming that we freeze the technology as it is now, what we are using is only about 5-10% of the value that this technology can create.” Introducing the case where the sales manager connects Claude and Salesforce every day to analyze major contracts, possible contract failures, and sales strategies, he explained, “There are still fewer people who have experienced this being possible than those who can benefit from it.”
There is more unutilized data in industrial sites. “Currently, AI models have not even seen most of the industrial data yet,” Siemens CEO Roland Bush said on the same day. Siemens supplies industrial systems to manufacturing sites and energy, transportation, and medical care worldwide. CEO Bush predicted that combining AI, which deals with corporate customer and sales data, and industrial data such as factories and products will be a new opportunity.
Salesforce’s “AI Force,” which was released together on the same day, also shows the same direction. It is a platform that connects existing work data and software to AI, not just using the AI model itself. Salesforce called up the company’s sales data from Antropic’s “Claude,” analyzed sales targets and key contracts, and demonstrated the process of sending Slack messages to employees in charge.
Salesforce CEO Mark Benioff said, “You can’t run a company with models alone.” The AI model can produce answers, but it does not know the exact customer information and work rules, authority, and security system inside the company. Salesforce explained that it applied a “zero data retention” structure in which customer data is not used for learning external AI models while connecting corporate internal data to AI. [Silicon Valley correspondent Wonho-seop]
