AI | The Sovereign Stack

1. Specification

Foundation models are reached through overseas APIs, and national compute depends on overseas accelerators (key finding 6). The sovereign approach does not require building a frontier model. It requires that any model used in a public service can be inspected, pinned to a known version, evaluated against the public body’s own tests, and run on infrastructure the buyer controls when the data demands it.

2. Layer profile

Id	Criterion	Standard
AI-1	The model version in use is pinned and the buyer is told before it changes.	Immutable model identifiers; change notice (general criterion G5)
AI-2	Open-weight models can be substituted for hosted ones without rewriting the application.	A provider-neutral inference API; ONNX for portability of smaller models
AI-3	Prompts, outputs and fine-tuning data belong to the buyer and are not used to train the provider.	Contract terms and a technical control, not a policy statement
AI-4	The model can be evaluated with the buyer’s own test set before and after any change.	Open evaluation harnesses; published evaluation results
AI-5	Sensitive workloads can run on UK infrastructure under the buyer’s control.	Self-hostable serving; Infrastructure profile
AI-6	The provenance of training data and the licence of the weights are stated.	Model cards; open-weight licences read by a lawyer, not a press release

Already adopted in UK government

No Open Standards Board profile covers models, inference or evaluation (Open standards for government). Adoption is de facto and set by whichever hosted API a department first integrated. Every line in this profile is therefore new ground for a buyer.

3. Open-source solutions that meet the profile

Not yet assessed. Candidates: open-weight model families with permissive licences; vLLM and similar serving stacks; open evaluation harnesses.

4. Public services implementing the standard

Not yet recorded. Candidate: the national AI compute facilities, which the evidence finds UK-owned and operated.

5. Skills and horizon

Horizon 2: UK-hosted AI. Horizon 3: sovereign compute. See the roadmap.

What this section rests on 36 findings

Every published finding in the Body of Knowledge for the layers this section covers. The Handbook recommends; the Findings show where we are. Open any row in the Stack Playground to see its sources.

Published findings for the layers covered by AI
Finding	Organisation	Product	Vendor	Owner country	Index
BBC-054	BBC	ChatGPT and Microsoft Copilot (staff general-purpose GenAI)	OpenAI; Microsoft	United States	28 · Dependent
BBC-055	BBC	ElevenLabs (synthetic voice) and Runway (generative video)	ElevenLabs; Runway	United States	39 · Dependent
BBC-056	BBC	BBC-trained/fine-tuned LLM (Style Assist; R&D model on BBC archive)	BBC R&D on open-weight base model	United Kingdom	85 · Sovereign
DEF-058	Defence	OpenAI GPT model behind a DE&S cyber and security compliance chatbot hosted on MoD Cloud	OpenAI	United States	25 · Dependent
DEF-059	Defence	Microsoft 365 Copilot	Microsoft	United States	25 · Dependent
DEF-060	Defence	Chat-MoD secure generative-AI assistant (chatmod.desdigital.mod.uk)	MoD (DE&S and DAIC; underlying model not named)	United Kingdom	69 · Partly sovereign
DEF-061	Defence	Digital Decision Accelerators for Defence (DDAD) open framework – AI/ML decision-support software for ASGARD	26 suppliers incl. Anduril, Helsing, Faculty, Oracle, QinetiQ, Roke, Telicent	Mixed	45 · Dependent
DEF-062	Defence	AI technical and data science support to DAIC programmes (5 lots incl. Generative AI and AI Assurance)	Capgemini; Deloitte; Applied Data Science Partners; Rowden Technologies; Advai	Mixed	86 · Sovereign
DEF-063	Defence	Hadean AI-enabled spatial computing platform (digital twins, synthetic environments) – R&D enterprise agreement	Hadean	United Kingdom	71 · Partly sovereign
DEF-064	Defence	Dstl–Google Cloud AI collaboration (MoU): access to Google Cloud AI tools, Marketplace and training	Google Cloud	United States	31 · Dependent
DEV-007	Devolved Governments	Microsoft Copilot Chat - preferred corporate generative AI tool for all SG staff	Microsoft	United States	22 · Exposed
GOV-074	Central Government	Microsoft 365 Copilot (OpenAI GPT models via Microsoft)	Microsoft (OpenAI models)	United States	23 · Exposed
GOV-075	Central Government	Microsoft 365 Copilot rollout	Microsoft	United States	23 · Exposed
GOV-076	Central Government	Anthropic Claude for GOV.UK agentic assistant (DSIT)	Anthropic	United States	33 · Dependent
GOV-077	Central Government	OpenAI ChatGPT Enterprise (UK data residency option)	OpenAI	United States	35 · Dependent
GOV-078	Central Government	Google DeepMind / Gemini partnership	Google	United States	38 · Dependent
GOV-079	Central Government	i.AI 'Humphrey' tools (Consult, Minute, Redbox, Lex, Parlex)	Incubator for AI (DSIT, in-house)	United Kingdom	95 · Sovereign
GOV-081	Central Government	Sovereign UK AI compute (Stargate UK with Nscale/NVIDIA; Nscale-Microsoft supercomputer)	Nscale; OpenAI; Microsoft	United Kingdom	57 · Partly sovereign
HE-053	Higher Education	ChatGPT Edu (OpenAI GPT-5)	OpenAI	United States	26 · Dependent
HE-054	Higher Education	Microsoft 365 Copilot / Copilot Chat (OpenAI models via Microsoft)	Microsoft	United States	26 · Dependent
HE-055	Higher Education	Google Gemini Pro / NotebookLM free student offer	Google	United States	31 · Dependent
HE-056	Higher Education	AI Research Resource (Isambard-AI, Dawn) - UK-hosted national AI compute	DSIT/UKRI; Universities of Bristol and Cambridge	United Kingdom	64 · Partly sovereign
LG-061	Local Government	Large language model behind Magic Notes (provider not disclosed by councils)	Undisclosed LLM provider (via Beam)	Unknown	42 · Dependent
LG-062	Local Government	Microsoft 365 Copilot	Microsoft	United States	30 · Dependent
LG-063	Local Government	Microsoft 365 Copilot	Microsoft	United States	30 · Dependent
LG-064	Local Government	Magic Notes AI note-taking for social workers	Beam	United Kingdom	62 · Partly sovereign
LG-065	Local Government	Paid/procured AI tools across councils (LGA State of the Sector AI 2025)	Various	Mixed	31 · Dependent
NHS-053	NHS	OpenAI GPT models (via Microsoft 365 Copilot / Azure OpenAI)	OpenAI; Microsoft	United States	23 · Exposed
NHS-054	NHS	Microsoft 365 Copilot + Copilot Studio national deployment	Microsoft	United States	25 · Dependent
NHS-055	NHS	Ambient voice technology (AI scribes) – NHSE self-certified supplier registry	Multiple (e.g. Tortus, Heidi, Accurx, Microsoft Dragon Copilot/Nuance, others)	United Kingdom	52 · Partly sovereign
POL-066	Policing	Microsoft Copilot (OpenAI GPT models) for staff and investigations	Microsoft (OpenAI models)	United States	23 · Exposed
POL-067	Policing	Police.AI - National Centre for AI in Policing	To be determined	United Kingdom	67 · Partly sovereign
SCH-045	Schools	Google Gemini for Education / DeepMind curriculum-grounded teacher assistant	Google	United States	21 · Exposed
SCH-046	Schools	OpenAI GPT-4o (powering Oak's Aila lesson assistant) + Cohere Rerank	OpenAI; Cohere	United States	30 · Dependent
SCH-047	Schools	Teacher use of general GenAI (ChatGPT; Microsoft Copilot; Gemini)	OpenAI; Microsoft; Google	United States	33 · Dependent
SCH-048	Schools	Oak National Academy platform (Aila) - DfE ALB	Oak National Academy	United Kingdom	91 · Sovereign

Change this section

The recommendations in this section are made by The Sovereign Stack Consortium, now forming.

Sections are Markdown in a public repository. Propose a change as a pull request, or open a discussion first. Contributions are reviewed by the section's maintainers under the Handbook governance. Text is licensed CC BY 4.0.

AI