Nvidia Says SAP, Salesforce Build Agents on Its Open Models

Nvidia said on 21 October 2026, during GTC Berlin, that SAP, Salesforce, Shopify, Synopsys, WPP and Dassault Systèmes are among the companies working with its open Nemotron models and datasets “to create custom models and build domain-specific agents.” In the same blog post it introduced Kumo, open models for structured business data, and named Coinbase, Databricks, DoorDash and Revolut among the first to work with them.

In plain terms, an open model is one whose weights a company can download, adapt and run itself, instead of renting it through another firm’s service. Nvidia’s pitch for Europe is the “sovereign” version of that idea: the open models, its blog says, give organizations “a foundation they can customize with local data and deploy on infrastructure they control.”

Business Pill · OPEN WEIGHTS

A Business Engineer explainer on open weights: renting a model or running it yourself. It covers the concept in general and says nothing about Nvidia.

Watch · The Keynote

NVIDIA GTC Berlin 2026 keynote with CEO Jensen Huang, 21 October, on NVIDIA’s YouTube channel. Quotations in this piece come from Nvidia’s blog posts and the slides shown, not from the spoken keynote.

What Nvidia Announced

• Nemotron 3.5. New Nemotron 3.5 Super VL and Nemotron 3.5 Bolt language models, which Nvidia says support text and images for agents. Super VL is available on Hugging Face, OpenRouter and Nvidia’s build.nvidia.com; Nemotron 3.5 Bolt E4B 4GB “is expected to be available on Friday, Oct. 23,” according to the blog.
• Nemotron Retriever. New open models for embedding, reranking and document parsing; in plain terms, the steps that help an agent find the right passage in a company’s documents.
• Kumo. Open foundation models, libraries and tools that, Nvidia says, help agents “forecast outcomes, score fraud risk and recommend actions from structured business data.”
• Cosmos and Alpamayo. Cosmos 3.5 world models for robots and vehicles, a new Cosmos Data Factory for curating and generating training data, and Alpamayo 2 Edge, a compact driving model that the blog says will be available “later this month.”
• Ising. A family of open models for quantum computing, from hardware calibration to error correction.

Who Nvidia Says Is Using Them

For Nemotron, the blog names Dassault Systèmes, Domyn, Dream, Palantir’s customers, Salesforce, SAP, Shopify, Synopsys and WPP, plus the professional services firms Accenture and EY. It also lists 22 post-training partners that help enterprises customize Nemotron, including CoreWeave, Fireworks, Google Cloud, Nebius and Scale AI.

For Kumo, Nvidia names Cognite, Coinbase, Databricks, DoorDash, Revolut and Tubi “among the first to work with Kumo models,” and says Snowflake customers are already using them in advertising, energy, financial services and retail.

For Cosmos Data Factory, the blog names Porsche with Voxel51, ANYbotics with Encord, Zalando and CEVA Logistics with Sereact, and Siemens. For Ising, it names IQM Quantum Computers, SEEQC and AWS, which it says is integrating the models into AWS Parallel Computing Service.

What the Keynote Slides Showed

On stage, a keynote slide headed “Turn Proprietary Data Into Sovereign Intelligence” described Nemotron as “Incredibly Customizable.” Under the Harvey and Trajectory logos, it set an accuracy bar for GPT 5.5 beside one for Nemotron 3 Ultra marked “+30.” A second slide, under the Alice and AI21 logos, showed Nemotron 3.5 Super VL marked “+41” beside GPT 6.1 Sol.

The slides, as shown on Nvidia’s keynote stream, do not name the benchmark or task behind the accuracy scale, so the figures cannot be checked against published results. AI21, Alice and Trajectory are also on the blog’s list of post-training partners.

Why the Kumo Piece Matters for Business

Kumo is aimed at structured business data, such as tables of transactions or orders. Nvidia says Kumo “shortens time to prediction and reduces the cost and complexity of building a separate machine learning pipeline for every task.”

The blog describes the intended workflow: developers start with pretrained Kumo models, “then fine-tune them using proprietary data when higher performance is critical.” The blog lists Kumo Structured, for relational and tabular prediction, forecasting and anomaly detection, and Kumo Engine, an open library stack.

Where Europe Comes In

The blog’s title frames the models as advancing Europe’s “sovereign intelligence.” It says Nemotron is built with data, research and expertise from the Nemotron Coalition, “including European partners Black Forest Labs, H-Company and Mistral.”

On stage in Berlin, Huang said: “To be part of this AI economy, you don’t want to just be consuming the AI; you want to be producing the AI,” according to Nvidia’s live blog of the keynote. “We could do all of that here in Europe, and in fact, we’ve already started.”

Update, 21 October: Nvidia Wraps the Models in an Agent Toolkit

In a second GTC Berlin blog post on 21 October, Nvidia said its open source Agent Toolkit is “now expanded with new tools”. The post lists its parts: the new NVIDIA Agent Knowledge Foundation with CUDA-X libraries, open models including Nemotron Retriever and Kumo, the OpenShell secure agent runtime, NemoClaw secure agent blueprints, and tools such as NeMo Relay and NeMo Compass.

Nvidia says the Agent Knowledge Foundation helps enterprises “turn data into company intelligence agents can use”, and that companies can use it “without moving data, rebuilding existing workflows or changing governance controls.” A new Auto-Ontology library, the post says, “translates raw company data into familiar business concepts — like customers, products and orders — and maps how they relate.” In plain terms, grounding an agent in company data means having it look up the company’s own records before it answers or acts, rather than relying only on what the model learned in training.

For the retrieval models, Nvidia cites early results: using the embedding model, Amazon OpenSearch Service “improved relevance with fewer search steps and virtually no added latency”, SAP’s fine-tuned reranker “ran nearly 3x faster”, and ServiceNow “improved reranking quality while cutting latency by nearly a quarter.” The post says Zoom is using Nemotron Retriever models for Zoom AI Services, and that Nvidia’s own supply chain organization uses the foundation through Palantir Foundry, Ontology and AIP. It names AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data among those building on it. These are Nvidia’s figures; the post does not give the test conditions behind them.

Connecting the Coverage

Our Germany pact piece reported on 19 October that Germany’s economics and research ministries signed a letter of intent with Nvidia at the Berlin investment summit, which, according to AFP’s account of the government’s statement, covers the development of specialised, open models for industrial AI. The same piece reported that Siemens named Nvidia Cosmos world models among the technology behind new AI applications at its Erlangen factory; Siemens is on today’s Cosmos Data Factory list.

Our midtraining piece covered an Nvidia research paper of 15 October that, its authors say, lifted Nemotron 3.5 Lightning’s SWE-bench Verified score from 60.7 to 64.5. Our Santander piece reported the bank’s statement that, in selected experiments, smaller specialised models served on Nvidia infrastructure cut serving costs from around €5 to about €0.03 per interaction.

Our Pichai piece reported his statement that Google’s Gemini 4 Argon reached almost 25 trillion tokens in its first 72 hours. Today’s other GTC Berlin pieces, our DSX water piece and our DSX POD piece, covered the infrastructure half of Nvidia’s pitch. Our OpenShell piece covered the same day’s release in which Nvidia says enterprises can run agents from eight named model builders on its OpenShell reference architecture. What is new here is the model half: named companies building agents on Nvidia’s open models.

The Structural Read

As we read it, the blog’s sovereignty pitch is a control pitch: Nvidia says the open models are a foundation companies “can customize with local data and deploy on infrastructure they control,” which is the opposite of renting a model through another firm’s service.

As we read it, the longest list in the post is not users but helpers: 22 post-training partners, from CoreWeave and Nebius to Scale AI, that Nvidia says help enterprises customize Nemotron for domain-specific workloads.

As we read it, Kumo extends the same play to the tables companies already keep: Nvidia says it reduces the cost and complexity of building a separate machine learning pipeline for every task.

Three Implications

FOR ENTERPRISE SOFTWARE BUYERS Nvidia names SAP, Salesforce, Shopify, Synopsys and Dassault Systèmes among the companies working with Nemotron models and datasets to create custom models and build domain-specific agents.

FOR BANKS AND PAYMENTS FIRMS Nvidia says Kumo helps agents score fraud risk from structured business data, and names Coinbase and Revolut among the first to work with the models.

FOR AI CLOUD OPERATORS CoreWeave, Nebius, Google Cloud and Fireworks are on the blog’s list of 22 post-training partners helping enterprises customize Nemotron.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework states the pattern it sees in platform shifts: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”

As we read it, Nvidia’s GTC Berlin announcements fit that split: the models in this post are released openly, while the same day’s DSX announcements cover the AI factories they run in. The framework was published in March 2026 and does not discuss these models.

What Is Not Established

The blog, as we read it on 21 October, gives no download counts, prices or contract terms, and does not say which of the named companies run the models in production rather than testing them. For Kumo it uses the words “among the first to work with.” The counts of named companies in our poster are our arithmetic; groups such as “Palantir’s customers” and “Snowflake customers” are not counted.

The Bottom Line

Nvidia used GTC Berlin to name SAP, Salesforce and other large companies as builders on its open Nemotron models, and to launch Kumo for structured business data, with Coinbase and Revolut among the first to work with it. As we read it, Nvidia is offering European companies models they can adapt and run themselves, alongside the AI factories it announced the same day to run them on.