Palantir and Nvidia formalise the sovereign AI operating system category

On 12 March 2026 Palantir and Nvidia published a Sovereign AI Operating System reference architecture, then on 29 June 2026 added an engine for running Nvidia Nemotron open models on the same stack. The category now has a name. UK regulated buyers now choose between a US controlled reference stack and a UK built, model agnostic alternative.

How the market named the category

The Palantir and Nvidia announcements did something more consequential than any single product launch. They gave the category a name. From that point, procurement teams in banks, hospitals, ministries and utilities could go to a vendor board and say the words sovereign AI operating system and expect the room to understand what was being described. Palantir's reference architecture pairs AIP, Foundry, Apollo, Rubix and AIP Hub with Nvidia Blackwell Ultra systems and Spectrum-X Ethernet networking, tested and qualified as an Nvidia Enterprise Reference Architecture. The June follow on lets customers run Nvidia Nemotron open models on that stack inside air gapped, sovereign environments, on the customer's own metal, with the model weights and the orchestration software both living inside the customer's boundary.

What Palantir and Nvidia actually shipped

The reference architecture is a tested, qualified pattern rather than a product. Nvidia supplies the hardware topology and the systems software. Palantir supplies the software substrate, spanning AIP, Ontology, Foundry, Apollo, Rubix and AIP Hub. The combination targets customers with existing GPU infrastructure, latency sensitive workflows and data sovereignty requirements. The 29 June 2026 engine takes the further step of hosting Nemotron open models in the same environment, closing the loop from silicon to application, and letting agencies train on their own data and retain full ownership of the resulting models. It is a well engineered answer to a real problem. It is also, in every layer, a stack designed, published and governed by two US listed companies.

Why a category name changes a boardroom conversation

Six months ago a CIO explaining on premise AI to a board committee had to spend twenty minutes defining the shape of the thing. Today the CIO says sovereign AI operating system and the finance director types four words into a search engine and lands on the Palantir reference. That reference will now anchor every RFP in this space through 2027. The category has three tests. Does the software run entirely inside the customer's boundary. Does model inference happen on hardware the customer owns. Can the customer verify what the system did without trusting the vendor. Any product that answers yes to all three is in the category. Any product that answers no to one of them is something else.

The procurement fork for UK regulated buyers

For UK regulated buyers the fork is now legible. A Palantir and Nvidia deployment is a US controlled stack running on US designed silicon, orchestrated by US software, in a category defined by two US listed companies. That is a legitimate choice for some buyers and a difficult one for others. A council, an NHS trust or a mid market defence supplier operating under the National Security and Investment Act 2021 will read the same catalogue of provisions and reach different conclusions from a large national programme with US alliance obligations.

Where a UK built alternative fits

MICKAI is a Sovereign Intelligence Operating System built in the United Kingdom on hardware the customer owns. It runs air gapped and on premise with no data egress. Every consequential action is signed into the Open Audit Record, a post quantum tamper evident ledger that any auditor can verify offline in a browser with no network and no trust in the vendor. Sixty three studios sit on one operating system today. Ten are production ready at launch and 53 are in development. The design goal is model agnosticism. The customer chooses the weights, whether from Mistral, IBM Granite, Cohere on premise, an open Llama derivative, an in house model or a future European sovereign model. The operating system provides the runtime, the audit trail and the studio applications.

Why patents matter to a regulated buyer

A regulated buyer signing a five year contract cares about vendor risk. Patents are one of the crude but effective ways to price that risk. Mickai LTD has 104 UK patent applications filed with 2,340 claims across 13 families. The families cover the architectural pieces a buyer would otherwise have to license from the incumbents: audit record cryptography, agent orchestration, air gapped model swapping, sovereign inference telemetry and the runtime primitives that make an air gapped studio surface behave like a network application. For a buyer the value is defensive. If the incumbent stack changes its licensing terms in 2028, an alternative that owns its own IP surface is easier to keep on the roadmap.

What every CIO should ask this quarter

• Does our vendor's software run entirely inside our boundary, or is there any control plane call home
• Can we verify what the system did without asking the vendor for logs
• If we swap the underlying model in twelve months, does the audit record and the studio surface stay the same
• Where is the software written and what jurisdiction governs the source code
• Which patent estate protects the architecture we are betting on for the next five years

When did Palantir and Nvidia announce their Sovereign AI Operating System reference architecture?

Palantir and Nvidia published the reference architecture on 12 March 2026. On 29 June 2026 Palantir added an engine for deploying Nvidia Nemotron open models inside sovereign, air gapped environments on the same qualified stack, closing the loop from silicon to application.

Is the Palantir and Nvidia reference architecture only for US federal customers?

No. The architecture is a general on premise pattern for customers with GPU infrastructure and data sovereignty requirements. However the code, the corporate parent, the silicon vendor and the model provider are all US entities, which is a material fact for buyers under UK, EU or Commonwealth sovereignty rules.

What are the procurement differences between a US-controlled and a UK-built sovereign AI operating system?

A US controlled sovereign AI operating system runs the customer's data on the customer's hardware, but the software, silicon and model provenance are governed by US law and US listed companies subject to the CLOUD Act 2018 and Section 702 of the Foreign Intelligence Surveillance Act. A UK built alternative such as MICKAI runs on the customer's hardware with British owned software and a British owned patent estate, and can host any compliant open weight model the customer chooses. For a UK council, NHS trust, defence supplier or bank operating under FCA rules and the National Security and Investment Act 2021, the second option removes a layer of extraterritorial reach. Neither is inherently safer. The right answer depends on which jurisdiction the buyer's regulator answers to and which extraterritorial exposures the buyer is willing to carry.

What is MICKAI?

MICKAI is a Sovereign Intelligence Operating System built in the United Kingdom, running on hardware the customer owns, on premise and air gapped, with no data egress. Every consequential action is signed into the Open Audit Record, a post quantum tamper evident ledger any outside party can verify offline in a browser. Sixty three studios sit on one operating system, ten production ready at launch and 53 in development, protected by 104 filed UK patent applications across 2,340 claims.