Salesforce Koa Stops Renting Reasoning

How Salesforce and Nvidia post-trained an enterprise reasoning model on open weights — and why every SaaS company sitting on a data moat is about to run the same play.

For three years the deal was simple: application vendors owned the workflow, frontier labs owned the thinking, and the labs got paid per token. Salesforce just stopped paying.

Until this week, when a Salesforce agent hit a problem it couldn't solve in one step, the request left the building. Agentforce's gateway would route the hard part — the multi-step reasoning, the long-running task — to Claude or ChatGPT, and Salesforce would pay for the privilege of having someone else think about its customer's support ticket.

At Dreamforce this week, that stopped being the only option. Salesforce announced Koa, its first reasoning model, built on Nvidia's open-weight Nemotron base and post-trained jointly by the two companies to handle sales, marketing and customer-support work. It slots into Agentforce alongside the frontier models rather than replacing them.

The modest framing undersells what happened. A $200 billion application company just demonstrated that the reasoning layer is contestable, and it did so using a base model it did not train, on an architecture it does not own, for a fraction of what building a frontier model costs.

What Koa actually is

Koa is a post-trained derivative of Nemotron, Nvidia's open-weight model family. Post-training is the step that converts a general-purpose system into a specialist: you take a model that knows a great deal about everything and teach it the vocabulary, patterns and failure modes of one job.

The interesting engineering choice is what Salesforce and Nvidia used as training data. Not customer data — none of it. Instead they generated synthetic material, simulating a customer-service environment with invented personas, including the irate callers that make support work hard, and sales professionals working a deal to close. Jayesh Govindarajan, Salesforce's EVP of AI, described the approach to TechCrunch, and the reason it matters is legal rather than technical. A model that has never ingested customer data cannot leak customer data. That removes the single most common objection in enterprise AI procurement.

Govindarajan was also direct about why this wasn't possible before. Salesforce had built plenty of small task-specific models, but reasoning always got outsourced, because there was no suitable base to start from. What changed is the availability of a state-of-the-art pre-trained model with clear data provenance — something he contrasted pointedly with Chinese open-weight alternatives, where the training corpus is unknown.

That is the whole unlock. Not a breakthrough in architecture. A base model with a clean chain of custody.

The five things Salesforce is actually selling

Strip away the conference framing and Koa is a bundle of five procurement answers, each aimed at an objection a CIO has already raised.

It is open-weight, so customers are not permanently leasing capability from a vendor whose pricing they don't control. It is task-trained rather than benchmark-trained, tuned for closing tickets rather than for the competitive-math problems the frontier labs optimize against. It has no customer data in it. It burns fewer tokens for the same work, which turns an operating expense into a smaller operating expense. And it runs inside Salesforce's existing security and data-residency perimeter, routed automatically through the Agentforce gateway depending on what the task needs.

Nvidia's Kari Ann Briski framed the efficiency angle as a combination of sovereign AI, time to first token, and reasoning that doesn't waste tokens. Sovereignty is doing real work in that sentence — "American pre-trained model with clear provenance" is a procurement category now, not a talking point.

Who this hurts, and who it doesn't

For OpenAI and Anthropic, the threat is narrow but structural. Nobody is switching away from frontier models for hard novel problems. What is leaving is the high-volume, low-variance middle — the routine multi-step reasoning that made up the boring bulk of enterprise token spend and, not coincidentally, the most predictable revenue. Losing the easy work while keeping the hard work is a margin problem disguised as a loyalty win.

It is also not a clean break. Salesforce simultaneously announced Claudeforce, letting companies use Claude as their interface while data stays in Salesforce's system of record. That is the actual posture: Salesforce keeps the data and the perimeter, and lets customers choose whose brain to rent for which task. The application layer becomes the router, and routers decide who gets paid.

For Nvidia, this is strategy compounding. Selling GPUs is the near-term business; making the open-weight base model the default starting point for every enterprise that wants its own model is the durable one. Every Koa-style derivative is trained and served on Nvidia silicon by default. Nemotron is not a side project — it is demand generation with a model card.

For every other SaaS vendor, Koa is a template with the steps written down. If you have a defensible workflow, proprietary understanding of a domain, and the ability to generate realistic synthetic training data for it, you can now own your reasoning layer. The barrier was never the compute. It was the absence of a clean base model. That barrier is gone.

For buyers, the honest caveat: no independent benchmarks exist yet. Salesforce says Koa is better at its customers' tasks and cheaper in tokens than routing to Claude or ChatGPT. Both claims are plausible and neither is verified. Token efficiency in particular is measurable, and should be measured on your own workloads before it appears in a business case. We have written before about how AI's productivity gains keep failing to reach the P&L — vendor efficiency claims are exactly where that gap opens.

Why this was predictable

The open-weight ecosystem has had a monetization problem for two years: enormous adoption, very little revenue capture. We called it Open Weights Won the Tokens, Not the Money. Koa is what resolution looks like. The money doesn't accrue to whoever released the weights. It accrues to whoever owns the workflow the weights get pointed at.

That reframes where value sits in the stack. The scarce asset is not the model. It is the combination of a proprietary workflow, the data exhaust it produces, and the customer relationship that makes deploying against it possible. Everything else is increasingly a commodity input — which is also why the expensive part of enterprise AI keeps turning out to be the installation rather than the model.

There is a second-order effect worth watching. If a dozen large application vendors each post-train their own domain model, enterprise AI fragments into vertical specialists rather than consolidating around three general-purpose systems. That is a better outcome for buyers on price and on lock-in. It is a considerably worse one for anyone who raised capital on the assumption that intelligence would be centralized and metered.

The company that owns the workflow eventually owns the model that runs it. Everything above the workflow is rented.

Salesforce did not build a better model than OpenAI. It built a cheaper model that is better at one specific job, on top of someone else's research, without touching a byte of customer data. That is not a frontier achievement. It is a supply-chain one — and supply-chain moves are the kind that reprice an industry quietly, one procurement cycle at a time.

Frequently Asked Questions

What is Salesforce Koa?

Koa is Salesforce's first reasoning model, announced at Dreamforce in September 2026. It is built on Nvidia's open-weight Nemotron base model and post-trained by both companies for sales, marketing and customer-support tasks. It runs inside Agentforce as an alternative to routing requests to external frontier models.

Was Koa trained on Salesforce customer data?

No. Salesforce and Nvidia post-trained Koa using synthetic data that simulated customer-service and sales environments, including invented personas such as irate callers and sales staff working a deal. Because no real customer data was ingested, the model cannot leak customer data to other users.

Does Koa replace Claude or ChatGPT in Salesforce?

No. Koa is offered alongside existing frontier models in Agentforce, and the platform's AI gateway routes each request to whichever model suits the task. Salesforce separately announced Claudeforce, which lets companies use Anthropic's Claude as an interface while their data remains in Salesforce's systems.

Why does an open-weight base model matter for enterprises?

Open weights let a company post-train and run its own specialist model rather than permanently renting capability at a vendor's price. Salesforce cited a second reason: clear data provenance. Knowing what the base model was trained on is a procurement requirement that many open-weight alternatives cannot satisfy.

Editor's note — sources: TechCrunch (September 15, 2026), Salesforce's Koa and Claudeforce product pages, and Nvidia's Nemotron documentation. Executive quotes are as reported by TechCrunch. Benchmark and cost claims are Salesforce's and Nvidia's own and have not been independently verified. Analysis is Edgewisely's.