At its GTC Berlin event on October 21, Nvidia expanded two parts of its enterprise AI stack: the open-source Agent Toolkit, which now includes a layer meant to ground agents in company data, and NVIDIA Confidential Computing, which gains a deployment option called Confidential Bare Metal. Both announcements are about the same worry from different angles: agents are only useful if they know a company’s data, and only trustworthy if they run inside boundaries the company controls. Everything below comes from Nvidia’s own posts, so performance and adoption claims are vendor-reported.
The Agent Toolkit gets a knowledge layer
Nvidia describes the Agent Toolkit as an open-source stack covering the models, harness, runtime, data and tools that builders need. The new piece is the NVIDIA Agent Knowledge Foundation, which Nvidia says helps enterprises turn their data into context that agents can use to find information, understand what it means for the business and make predictions. According to the company, it can be used without moving data, rebuilding workflows or changing governance controls.
The foundation draws on Nvidia’s CUDA-X libraries, including cuDF, cuGraph, cuML, cuOpt and cuVS, plus a new Auto-Ontology library. Nvidia says Auto-Ontology translates raw company data into familiar business concepts such as customers, products and orders and maps how they relate. It also bundles open models: Nemotron Retriever and a new Kumo family covering tabular data, relational data, time series and forecasting.
The rest of the toolkit is a set of components Nvidia lists: the OpenShell secure agent runtime, NemoClaw blueprints for secure, specialized agents (the newest is a blueprint for agent optimization), and tools such as NeMo Relay and NeMo Compass.
Who is using it, according to Nvidia
Nvidia names several early users. Its own supply chain organization is deploying agents through Palantir Foundry, Ontology and AIP. Databricks pairs Kumo Anomaly with its Genie Ontology, and Snowflake is letting customers run the Kumo Engine and Kumo Relational models in its notebooks. For the retrieval models, Nvidia reports that Amazon OpenSearch Service improved relevance with fewer search steps, SAP’s fine-tuned reranker ran nearly three times faster, and ServiceNow improved reranking quality while cutting latency by nearly a quarter. Zoom is using the models for document understanding and search. We could not check these figures independently.
Confidential Bare Metal for model builders
The second announcement targets companies that build frontier models but rent the hardware. Nvidia says Confidential Bare Metal lets them run sensitive workloads on dedicated third-party infrastructure while keeping control of their models and data. New platform security controls protect data in transit, measure platform integrity and enforce hardware-level access restrictions, and let builders operate their own host and platform software on third-party cloud, neocloud and enterprise infrastructure.
Nvidia’s Confidential Computing now offers three deployment options: Confidential Bare Metal, Confidential Containers for Kubernetes, and Confidential Virtual Machines. Support extends to Grace Blackwell NVL72 rack-scale systems and to x86-based HGX and DGX Rubin NVL8 systems, and will integrate with the NVIDIA DSX OS reference architecture for cloud providers building AI factories. The company says the technology relies on hardware-rooted isolation and cryptographic attestation.
Providers and partners
Cohere plans to integrate Confidential Bare Metal into Cohere Model Vault. Google Cloud is extending support to A4X Max instances powered by GB300 NVL72 systems. Nvidia lists CoreWeave, Firmus, GMI Cloud, Google Cloud, IREN, Lambda, Microsoft Azure Infrastructure, Nebius, Nscale and Vultr as supporting Confidential Computing deployments, with Cisco, Dell, HPE, Lenovo, Supermicro and Bull supporting it through certified systems.
Why it matters
Frontier labs have recently drawn attention to how their training environments are secured; see our report on OpenAI moving frontier training offline after sandbox escapes. Nvidia’s announcement is a hardware-and-runtime answer to a related question: how do you let valuable models and sensitive company data run on infrastructure you do not own? Our earlier coverage of the same event covers the physical-AI side, in Cosmos Data Factory. For background on the term, see AI agent.
What to watch next
- Independent testing of the retrieval and forecasting claims from Nvidia’s partners.
- Which providers ship Confidential Bare Metal and when it reaches general availability.
- Whether the OpenShell runtime and NemoClaw blueprints are adopted outside Nvidia’s partner circle.
- How cloud providers expose attestation to customers who want to verify the claims.
Key takeaways
- Nvidia added an Agent Knowledge Foundation to its open-source Agent Toolkit to ground agents in company data.
- The Kumo model family covers tabular, relational, time-series and forecasting tasks.
- Confidential Bare Metal joins Confidential Containers and Confidential Virtual Machines as a third deployment option.
- Cohere and Google Cloud are among the first named adopters; performance figures are Nvidia’s own.
- Both announcements frame agent safety as a mix of data grounding and hardware-enforced boundaries.
Frequently asked questions
What is the Agent Knowledge Foundation?
Nvidia describes it as a set of CUDA-X libraries and open models that let agents find and interpret company data without that data being moved or governance being changed.
What does Confidential Bare Metal protect?
According to Nvidia, it protects models and data while they run on dedicated third-party hardware, using hardware-rooted isolation and cryptographic attestation.
Are these products available now?
Nvidia’s posts describe the toolkit as in use at named enterprises and Confidential Computing as supported by listed providers. Cohere’s integration is described as planned, so availability varies by provider.
