Salesforce and Nvidia's Koa Reasoner Shows Enterprises Can Sidestep the AI Labs

Salesforce's first reasoning model — post-trained on Nvidia's open-weight Nemotron — is a direct answer to the proprietary frontier labs and their data demands.

Published: 2026-09-16 Category: Quick Take Sources: TechCrunch — Salesforce and Nvidia's new reasoning model is everything the AI labs should fear

What happened

A new AI model called Koa is one of the biggest announcements from Salesforce at this week's Dreamforce conference. Koa is the company's first reasoning model, built on Nvidia's open-weight Nemotron model and post-trained by the two companies to excel at sales, marketing, and customer-support tasks. Salesforce will offer Koa as an alternative to the other models in its Agentforce platform, where customers build agents to handle rote work like answering service questions and scheduling appointments.

Until now, Salesforce had relied on frontier models for reasoning. "Reasoning has always been something that we've relied on the frontier model providers for. Until now," Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch. Previously, if an agent needed to reason through a long-running or multi-step task, those prompts would be routed through Agentforce's AI gateway to a frontier model like Claude or ChatGPT. The constraint, Govindarajan said, was the lack of a suitable pre-trained base model: "Until Nemotron came along, there was no sovereign American pre-trained model that was available… that was state of the art… that had clear data provenance. We have no idea what Qwen trains on," he said, referring to Alibaba's popular open-weight model. Post-training such a model means steering it from a general base toward specific enterprise competencies.

The divergence between enterprise and frontier

Koa is the clearest sign yet that enterprise AI needs have diverged from what the frontier labs are selling. Proprietary labs want enterprises to upload files, code, prompts, and feedback directly into their models and agents — and to spend millions doing so. That model turns the enterprise into both a customer and a data source, feeding the labs' models with real-world usage. Salesforce's answer is the opposite: build the reasoning capability the enterprise needs, on an open-weight base with clear provenance, and keep the data and the reasoning in-house.

The economics explain why this matters. Reasoning is the highest-margin, most strategically valuable layer of an AI stack, and the frontier labs have used it to keep enterprises locked into their platforms. A company like Salesforce post-training its own reasoner removes a dependency on Claude or ChatGPT for exactly the workloads that would otherwise route the most data — and the most revenue — to a lab. It's a de-commoditization of the reasoning layer, and it hits the labs in their most important position: enterprise lock-in.

Why this matters

Koa is a proof point that open-weight models have matured enough for a major enterprise to build on them. Six months ago, the conventional wisdom was that frontier labs had an insurmountable lead on open models, especially for reasoning. Salesforce betting its flagship agent platform on Nvidia's open-weight Nemotron — over Alibaba's Qwen, which it explicitly distrusts on provenance grounds — signals that the open-weights lane is now viable for serious enterprise work, provided the provenance story is clean.

That "sovereign American pre-trained model" framing is doing a lot of work. It's partly a real technical and compliance argument — enterprises genuinely want to know what their models were trained on — and partly a geopolitical one, a thumb on the scale against Chinese open models. Either way, it points to a future where enterprises don't have to rent their reasoning from a handful of labs. That's the "everything the AI labs should fear" part: not a single model, but the demonstration that the frontier's most valuable capability is now reproducible in-house. The labs' moat isn't only intelligence — it's been the enterprise's unwillingness to build it themselves. Koa says that willingness has arrived.

Reporting via TechCrunch's "Salesforce and Nvidia's new reasoning model is everything the AI labs should fear" (September 15, 2026).