Salesforce Trained Its Own Reasoning Model and Stopped Calling OpenAI
Salesforce and Nvidia shipped a reasoning model called Koa on September 15, built on Nvidia's open-weight Nemotron and aimed squarely at Agentforce workloads — sales, marketing, customer support. Coverage at https://techcrunch.com/2026/09/15/salesforce-and-nvidias-new-reasoning-model-is-everything-the-ai-labs-should-fear/. The quote that should make a lab executive put down their coffee comes from Jayesh Govindarajan, Salesforce's EVP of AI: reasoning has always been something we relied on the frontier model providers for, until now.
The strategic logic is not about beating GPT-6 on anything. It is that Salesforce does not need to. An enterprise agent closing a support ticket is doing a narrow, repetitive, well-specified job, and for that job a task-trained open-weight model is cheaper per token, faster to first token, and runs inside Salesforce's existing data governance without a new vendor contract. Salesforce also makes a point of saying Koa never ingested actual customer data, which removes the single objection that kills the most enterprise deals.
Token efficiency is the quiet kill shot. Frontier labs price per token and have every incentive to have your agent think longer. A customer whose vendor is also their model provider has no leverage on that. A customer running a model they control on a stack they control can just decide a support triage does not need 4,000 reasoning tokens. Multiply by the volume Salesforce processes and the savings are not a line item, they are a budget.
This is the second time this week the same pattern showed up from the opposite direction. [Jev's whole argument is that most production calls never needed a prose model](https://clauday.com/article/f5fb35c2-fd03-4b36-b7a4-1ded3631d078), and Koa's argument is that most enterprise reasoning never needed a frontier one. [Nemotron has been built for agents rather than chat since the 550B release](https://clauday.com/article/c666eeaf-4621-435c-9033-0ba3f7ca15ce), and Nvidia's Kari Ann Briski framing it as sovereign AI plus time-to-first-token plus efficient reasoning tells you who Nvidia thinks the buyer is. It is not the labs.
The thing to watch is whether Salesforce publishes numbers. Right now there is no benchmark and no parameter count in the announcement, which is a choice, and a convenient one. But the direction is unmistakable: the largest enterprise software vendor on earth just moved reasoning in-house on an open-weight base. Every other application company with a Nvidia relationship is reading that press release very carefully this morning.
← Back to all articles
The strategic logic is not about beating GPT-6 on anything. It is that Salesforce does not need to. An enterprise agent closing a support ticket is doing a narrow, repetitive, well-specified job, and for that job a task-trained open-weight model is cheaper per token, faster to first token, and runs inside Salesforce's existing data governance without a new vendor contract. Salesforce also makes a point of saying Koa never ingested actual customer data, which removes the single objection that kills the most enterprise deals.
Token efficiency is the quiet kill shot. Frontier labs price per token and have every incentive to have your agent think longer. A customer whose vendor is also their model provider has no leverage on that. A customer running a model they control on a stack they control can just decide a support triage does not need 4,000 reasoning tokens. Multiply by the volume Salesforce processes and the savings are not a line item, they are a budget.
This is the second time this week the same pattern showed up from the opposite direction. [Jev's whole argument is that most production calls never needed a prose model](https://clauday.com/article/f5fb35c2-fd03-4b36-b7a4-1ded3631d078), and Koa's argument is that most enterprise reasoning never needed a frontier one. [Nemotron has been built for agents rather than chat since the 550B release](https://clauday.com/article/c666eeaf-4621-435c-9033-0ba3f7ca15ce), and Nvidia's Kari Ann Briski framing it as sovereign AI plus time-to-first-token plus efficient reasoning tells you who Nvidia thinks the buyer is. It is not the labs.
The thing to watch is whether Salesforce publishes numbers. Right now there is no benchmark and no parameter count in the announcement, which is a choice, and a convenient one. But the direction is unmistakable: the largest enterprise software vendor on earth just moved reasoning in-house on an open-weight base. Every other application company with a Nvidia relationship is reading that press release very carefully this morning.
Comments