Salesforce launches Koa, an open-weight reasoning model post-trained on synthetic data to replace frontier API calls

Salesforce and Nvidia announced Koa, a reasoning model built on Nvidia's open-weight Nemotron base and post-trained using synthetic data for sales and support tasks. The release provides enterprise customers with a sovereign, token-efficient alternative to proprietary frontier models like Claude or ChatGPT, aiming to reduce inference costs and eliminate data leakage risks associated with sending customer prompts to external labs. This move structurally shifts revenue from pure API consumption toward platform-embedded, lower-margin inference services. For financial institutions, this signals potential margin compression for upstream AI vendors as enterprise buyers migrate high-value reasoning workloads to internal or partner-hosted open-weight alternatives.