At this year's Dreamforce conference, Salesforce shipped two products that point in opposite directions. One is Claudeforce, built with Anthropic, which lets customers operate Salesforce directly from inside Claude. The other is Koa, a reasoning model built with NVIDIA on open-weight foundations — and, by Salesforce's own framing, designed to replace closed models like Claude and GPT on a meaningful slice of enterprise tasks.
One hand embraces the frontier labs; the other hand works to route around them. Most coverage treated this as a product story. It isn't. The two launches appearing side by side is the signal.
Here is the structural shift: for the first time, a company's accumulated domain knowledge can be installed inside the model itself. That gives enterprise software vendors a credible path around the frontier labs — and it changes who captures the value of enterprise AI.
1. Rethinking What a Vertical Model Is
Two assumptions dominate the conversation. First, that vertical models are just small models doing easy work — a compromise for those who can't afford frontier compute. Second, that frontier models will eventually absorb every vertical use case, and "every step forward for large models is a step back for SaaS."
Koa dismantles both.
It is not a small model trained from scratch. Salesforce post-trained NVIDIA's open-weight Nemotron 3 Super using supervised fine-tuning (SFT) followed by reinforcement learning with Group Relative Policy Optimization (GRPO), on tooling that includes NeMo RL, NeMo Gym, and NeMo AutoModel. General reasoning comes from the open foundation; domain competence comes from post-training. This is a complete paradigm — open base plus domain post-training — not a downscaled imitation.
The results tell the real story. On Salesforce's CRM Benchmark, Koa scored a weighted average of 8.6 — just below Claude Opus 4.8 at 8.7 and GPT-5.5 at 9.0, but ahead of GPT-4.1 at 8.1. Salesforce's official claim: Koa matches or exceeds leading models on CRM actions with three times fewer errors. Internal testing also cites 11% better tool-calling precision, 2.1x better context recall reliability, and 15% better long-conversation memory.
Read that scorecard carefully. On general benchmarks, Koa still trails the frontier. But on bounded, well-defined workflows — updating opportunities, routing cases, scheduling follow-ups — it is good enough. The competition has moved from "who is smarter" to "who fits the task."
One detail deserves special attention: the training data. Salesforce used no customer data at all. The corpus is entirely synthetic — simulated scenarios spanning more than 14 industries, from irate customer-service callers to sales professionals closing deals. Salesforce AI EVP Jayesh Govindarajan told TechCrunch why this hadn't happened before: there was no pre-trained base model that was sovereign, state of the art, and had clear data provenance. Until Nemotron, that foundation didn't exist.
2. Why Now: Three Forces
First, the rising cost of trust. The same week, another model vendor was caught packaging users' workspace files and uploading them to the cloud, with an affected company publicly threatening legal action. When model capabilities converge, data-boundary questions dominate procurement. Koa's answer: the weights, the post-training, and the inference all stay inside Salesforce's own trust boundary.
Second, economics. Frontier AI is a high-margin business — Anthropic reportedly expects 50% gross margin in 2026 and 77% by 2028. Meanwhile Salesforce's Agentforce + Data 360 revenue is approaching $3.9 billion per quarter. Salesforce has both the motive and the scale to pull inference workloads back in-house. TechCrunch adds a telling detail: NVIDIA says Nemotron's inference architecture is token-efficient — the same work burns fewer tokens, which matters when AI spending is measured in millions.
Third — and deepest — the SaaS industry's existential question. As models improve, customers could bypass software vendors entirely and build their own agents. Every SaaS company must now answer: how do we become the piece customers cannot route around? Salesforce CEO Marc Benioff put it plainly: "The most valuable thing Salesforce has built isn't our platform — it's the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself."
3. A Framework: The Reasoning Stack
Abstract Koa into a reusable tool. Enterprise AI inference needs come in three layers:
Bottom layer — domain action reasoning. Bounded tasks with standard actions and closed context: updating CRM records, routing tickets, rule-based approvals. The optimal model here is a post-trained vertical model — cheaper, more controllable, data never leaves the domain.
Middle layer — workflow orchestration. Multi-step, cross-system tasks that chain tool calls. This belongs to the platform — Agentforce's AI gateway routes each request to the model best suited for it, and Koa slots in as one option.
Top layer — open-domain reasoning. Open-ended research, cross-domain creative work, genuinely novel problems. Frontier models remain irreplaceable here for now.
The framework's value: it replaces the dead-end question "will vertical models replace frontier models?" with an allocation question — "who should own each layer?" A mixed-model strategy is not a transitional compromise. It is the steady state.
4. Will the Paradigm Spread?
NVIDIA is the paradigm's chief evangelist. Jensen Huang appeared at Dreamforce to support Koa personally, citing a statistic: open-model usage has grown from about 30% at the start of last year to roughly 70% today. His long game is "sovereign AI" — helping every company become an AI company. This does not conflict with NVIDIA's lab business as much as it seems: Anthropic reportedly runs over one million AWS Trainium2 chips and has committed to spending over $100 billion on AWS over a decade. If the labs are de-NVIDIA-fying, NVIDIA must find its next market in the enterprise base — which is vastly larger than a handful of AI labs.
Two constraints temper the forecast. Talent: people who can run high-quality post-training are concentrated in frontier labs and big tech; most software companies lack them. Economics: Salesforce can amortize post-training costs because it has the world's largest CRM workload and roughly $3.9 billion in quarterly AI-attached revenue. Most companies have neither.
But history suggests both constraints erode. When the iPhone launched, Salesforce scrambled to build mobile products and struggled for lack of iOS developers; within a few years, iOS talent was abundant, then oversupplied. With open foundations maturing and post-training toolchains like NeMo becoming productized, the barrier is dropping from "lab-grade" to "engineering-team-grade."
Extrapolate across industries: a bank can post-train a model on its compliance workflows. A hospital can install care-coordination reasoning on-premises. Governments get there fastest of all — Koa's sibling technology is already powering Missionforce, Salesforce's offering for private-cloud and air-gapped deployments, backed by a reported $1.6 billion contract with the U.S. Department of Veterans Affairs. Every company that owns distinctive process data is a potential vertical-model company.
5. What To Do About It
- If you build SaaS products: Inventory your process data. Which tasks have a "correct action sequence" that exists only inside your system? That is the moat your future vertical model stands on. Start small-scale post-training experiments on open foundations now — don't wait for a general model to solve it "eventually."
- If you buy enterprise AI: Segment your reasoning needs by layer before negotiating with vendors. Prefer domain-bounded vertical solutions for the bottom layer; pay frontier prices only for open-domain reasoning. And put data-boundary commitments in the contract, not the marketing deck.
- If you're an engineer: Post-training skills — SFT/GRPO, NeMo-style toolchains, synthetic data construction — are among the fastest-growing demands of the next few years. More valuable than learning yet another framework.
- If you track the frontier labs: Watch the bottom of your market get eaten — not by smarter models, but by sufficient-and-controllable ones. The labs' moat is shifting from raw capability to the extreme open-domain edge.
Launching Claudeforce and Koa together is not corporate schizophrenia. It is a clear-eyed positioning across the reasoning stack: keep renting frontier intelligence at the top, start owning it at the bottom. When sufficient vertical intelligence can be post-trained by anyone with process data, the power structure of enterprise AI finally begins to redistribute.