Today we’re announcing a collaboration with Palantir to deliver Palantir’s Ontology for Cybersecurity via Cisco’s Secure AI Factory with NVIDIA as a preferred full-stack foundation for Palantir’s Sovereign AI OS. This will make it much easier for customers, including government organizations and large enterprises in regulated industries, to have operational control of their proprietary data and critical infrastructure in AI deployments, and to deliver the foundational infrastructure and IT Ontology for leading technical organizations. It’s something we hear often from our customers as they begin to scale their AI use cases, and we think this is a meaningful step forward for the industry.
I also want to share some perspective on what’s happening in AI that’s making this all possible.
Intelligence, Cost, and Control
When it comes to AI strategy, I think the relevant vectors are intelligence, cost, and control.
First, let’s look at intelligence. Models are getting incredibly powerful, even tackling some of the hardest math problems in history. But raw intelligence is not always the correct answer – often, post-training an open model like NVIDIA Nemotron with business data can deliver better accuracy at lower costs. Raw benchmark scores make for great slides, but the real test is applied intelligence built on specific enterprise context. In any AI deployment, we have to ask ourselves, how well does a model perform on the specific tasks that matter to my business in this use case? That requires custom evals, fine-tuning models with proprietary data, and deeply integrating with your real workflows.
Second, is cost. Price per token is important, and organizations need to be able to measure and track it with rigor and accuracy. But useful intelligence per dollar, per watt, per unit of infrastructure is the metric that will ultimately show up in the health of a business long term.
And finally, control means exactly what it sounds like. It’s the ability to customize a model and deploy it where you choose, in the cloud, at the edge, or on-prem, with control over your data, your security posture, and your operating environment. That’s a meaningful design requirement for protecting your proprietary data, intellectual property, and competitive advantage, while meeting compliance requirements.
So, when you put everything in this context, it’s clear that the best model for you is rarely the smartest model in the abstract. It’s the model that sits at the right point on the Pareto frontier of intelligence, cost, and control for your specific workload. To make it real, we know that a fraud-detection pipeline calls for a different model than a customer service agent, and a classified workload that can never leave a specific building calls for a different model and architecture altogether.
This only matters right now because we’ve hit a tipping point in open- source and open- weights AI models. Post-trained open-weight models are outperforming frontier models on real-world applications today.
That’s genuinely exciting, but more importantly, it empowers customers to diversify their AI strategy across their operations and be more strategic and flexible. They can use highly capable frontier models in the cloud for some more general use cases, while investing in custom training open-source models on their proprietary data and running on-prem AI factories for areas where they value a high-degree of control, and many flavors in between.
Defining a New Architecture for AI with Cisco, Palantir and NVIDIA
Cisco, Palantir and NVIDIA are teaming up to extend our secure custom AI platform to companies and countries. With a focus on cybersecurity, Cisco’s Secure AI Factory with NVIDIA will serve as a preferred full-stack foundation for Palantir’s Sovereign AI OS with NVIDIA Nemotron open models, powering and delivering Palantir’s Ontology for Cybersecurity and for the modern IT organization. It will offer customers a repeatable, governed path to production AI from day one, built for organizations that need a high degree of control over their own data and infrastructure.
Cisco will provide the secure, observable infrastructure with unified management in Cisco Cloud Control. Nemotron provides a platform for enterprises to post-train with their proprietary data, boosting accuracy and saving costs while protecting IP. Palantir will provide the operational AI layer with Palantir Foundry and Artificial Intelligence Platform (AIP), grounded in the Palantir Ontology, to create supply chain visibility, identify constraints, continuously codify operational expertise and guide decisions at machine speed that turns an organization’s data into decisions, workflows, and agentic action, backed by the Ontology and with a focus on security and IT organizations.
It will come together in a reference architecture that combines compute, networking, storage, security, observability, and AI workflows, all validated as one system. The focus is on defining an architecture that’s faster to deploy, simpler and cheaper to run, and more secure than what customers could assemble on their own. Forward-deployed engineers will help companies and countries build custom AI and engage the AI economy.
Every organization will need to deeply understand the levers of intelligence, cost, and control. That will never mean a one-size-fits all approach to tackling every use case, but we do believe that the infrastructure and platforms underlying AI deployments can get so much faster and easier to deploy.
We already know this architecture is producing results with NVIDIA and Palantir using it as the blueprint for one of the most complex supply chains in the world. The next step is taking it to more industries around the world.
