Sovereign AI research
The field guides, analysis, playbooks, tools and definitions MindMap Digital has published on sovereign AI, in one place. Each entry says what the piece answers, who wrote it and when, and comes with a stable link and a ready-made reference.
What is sovereign AI?
Sovereign AI is an architecture where (1) customer data never leaves the customer's network perimeter, (2) model weights run on hardware the customer controls, (3) inference logs stay in the customer's own SIEM, and (4) the entire stack can operate air-gapped with zero outbound calls. It is not a sovereign cloud region, not a Bring Your Own Key arrangement, and not a hyperscaler deployment in your country's borders.
- Customer data never leaves the customer's network perimeter
- Model weights run on hardware the customer controls
- Inference logs stay in the customer's own SIEM
- The entire stack can operate air-gapped, with zero outbound calls
Field guides and field notes
Longer pieces written from delivery work and public sources. They are practitioner's analysis, not studies.
- Lead field guide
- Saurabh Goenka
- Published July 2026
- Updated November 2026
- 10 min read
Sovereign AI Reference Architectures: A 2026 Field Guide
There is no single sovereign AI stack. The guide describes five reference architectures (lean single-rack, multi-tenant enterprise, air-gapped, GPU-on-edge, and hybrid sovereign cloud with on-premise) and how to choose the model family, vector database, inference server and integration pattern inside them. It covers what a cost per million tokens should include, what each architecture makes easier to evidence for a regulator, a default starting point and where the guide is likely to date. It is practitioner's analysis and reports no survey or deployment data.
Cite this
Goenka, S. (2026). Sovereign AI Reference Architectures: A 2026 Field Guide. MindMap Digital. https://www.mindmapdigital.ai/blog/sovereign-ai-architecture-report-2026
@misc{mindmap2026architecturereport2026, author = {Goenka, Saurabh}, title = {Sovereign AI Reference Architectures: A 2026 Field Guide}, howpublished = {MindMap Digital}, year = {2026}, month = {7}, url = {https://www.mindmapdigital.ai/blog/sovereign-ai-architecture-report-2026} } - Engineering guide
- MindMap Engineering
- Published July 2026
- Updated November 2026
- 8 min read
Agentic AI in Regulated Industries: Designing for Audit
An AI agent can be audited when its runtime is designed for audit from the start. The guide covers three patterns that can be made auditable (bounded ReAct, planner-executor, and multi-agent orchestration with hand-off contracts), three failure patterns to design against (silent loops, hidden tool calls, unbounded reasoning) and four engineering controls that separate a prototype from a production runtime. It is practitioner's analysis and reports no deployment statistics.
Cite this
MindMap Engineering. (2026). Agentic AI in Regulated Industries: Designing for Audit. MindMap Digital. https://www.mindmapdigital.ai/blog/state-of-agentic-ai-regulated-industries-2026
@misc{mindmap2026stateofagenticai2026, author = {{MindMap Engineering}}, title = {Agentic AI in Regulated Industries: Designing for Audit}, howpublished = {MindMap Digital}, year = {2026}, month = {7}, url = {https://www.mindmapdigital.ai/blog/state-of-agentic-ai-regulated-industries-2026} } - Field notes
- Saurabh Goenka
- Published July 2026
- 16 min read
What CRO Conversations on AI Look Like in 2026
Chief Risk Officers have moved on from asking whether to adopt AI. They now ask about vendor concentration risk, model lifecycle artefacts and the audit trail. A synthesis of the author's conversations with CROs across BFSI, healthcare, the public sector and pharma, organised as four shifts in how AI risk is framed.
Cite this
Goenka, S. (2026). What CRO Conversations on AI Look Like in 2026. MindMap Digital. https://www.mindmapdigital.ai/blog/cro-conversations-ai-2026
@misc{mindmap2026croconversations2026, author = {Goenka, Saurabh}, title = {What CRO Conversations on AI Look Like in 2026}, howpublished = {MindMap Digital}, year = {2026}, month = {7}, url = {https://www.mindmapdigital.ai/blog/cro-conversations-ai-2026} }
Analysis and engineering notes
Shorter pieces on one question each: architecture, cost, regulation, security.
- Analysis
- Saurabh Goenka
- Published 17 October 2026
- 7 min read
AI Has Access. Who Has Control? The Case for Sovereign AI
Sovereignty means being able to enforce a decision, not only choosing where a server sits. Written after CERT-In's October 2026 advisory on AI agents, the piece sets out three controls to review together (deployment control, operational control and evidence) and argues that the final permission check belongs outside the model.
Cite this
Goenka, S. (2026). AI Has Access. Who Has Control? The Case for Sovereign AI. MindMap Digital. https://www.mindmapdigital.ai/blog/sovereign-ai-agent-security
@misc{mindmap2026agentsecurity, author = {Goenka, Saurabh}, title = {AI Has Access. Who Has Control? The Case for Sovereign AI}, howpublished = {MindMap Digital}, year = {2026}, month = {10}, url = {https://www.mindmapdigital.ai/blog/sovereign-ai-agent-security} } - Analysis
- Saurabh Goenka
- Published May 2026
- 8 min read
The Sovereign AI Inflection Point: Why Regulated Enterprises Are Moving On-Prem
Regulated enterprises are moving AI onto their own infrastructure because of regulation, not fear of the cloud. The piece explains what sovereign AI means, why the cost argument has changed, and the reference architecture MindMap deploys.
Cite this
Goenka, S. (2026). The Sovereign AI Inflection Point: Why Regulated Enterprises Are Moving On-Prem. MindMap Digital. https://www.mindmapdigital.ai/blog/sovereign-ai-inflection
@misc{mindmap2026inflectionpoint, author = {Goenka, Saurabh}, title = {The Sovereign AI Inflection Point: Why Regulated Enterprises Are Moving On-Prem}, howpublished = {MindMap Digital}, year = {2026}, month = {5}, url = {https://www.mindmapdigital.ai/blog/sovereign-ai-inflection} } - Analysis
- Saurabh Goenka
- Published June 2026
- 8 min read
Sovereign AI vs Bring Your Own Key: Why BYOK Doesn't Close the Compliance Gap
Bring Your Own Key gives a customer cryptographic control over data at rest. It does not give control over the model, the inference or the audit trail. The piece covers the three questions BYOK leaves open, what regulators ask for, and when cloud with BYOK is still a reasonable choice.
Cite this
Goenka, S. (2026). Sovereign AI vs Bring Your Own Key: Why BYOK Doesn't Close the Compliance Gap. MindMap Digital. https://www.mindmapdigital.ai/blog/sovereign-ai-vs-byok
@misc{mindmap2026sovereignaivsbyok, author = {Goenka, Saurabh}, title = {Sovereign AI vs Bring Your Own Key: Why BYOK Doesn't Close the Compliance Gap}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/sovereign-ai-vs-byok} } - Decision framework
- Saurabh Goenka
- Published June 2026
- 9 min read
Open-Weights vs Closed API: The Architecture Decision That Defines 2026
On enterprise workloads, the choice between open-weights models and closed APIs is now decided by regulatory posture, vendor concentration risk and unit economics at the customer's volume, not by capability. Includes where closed APIs still win and a one-paragraph decision framework.
Cite this
Goenka, S. (2026). Open-Weights vs Closed API: The Architecture Decision That Defines 2026. MindMap Digital. https://www.mindmapdigital.ai/blog/open-weights-vs-api-2026
@misc{mindmap2026openweightsvsclosedapi, author = {Goenka, Saurabh}, title = {Open-Weights vs Closed API: The Architecture Decision That Defines 2026}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/open-weights-vs-api-2026} } - Cost model
- MindMap Engineering
- Published June 2026
- 11 min read
What It Actually Costs to Run Llama 3.3 70B On-Prem
Whether on-premise inference is cheaper than a cloud API depends on monthly token volume: the cloud wins at low volume and on-premise wins at high volume. The piece sets out what a fully loaded on-premise cost per million tokens includes and works through three volume tiers.
Cite this
MindMap Engineering. (2026). What It Actually Costs to Run Llama 3.3 70B On-Prem. MindMap Digital. https://www.mindmapdigital.ai/blog/llama-on-prem-economics
@misc{mindmap2026onpremeconomics, author = {{MindMap Engineering}}, title = {What It Actually Costs to Run Llama 3.3 70B On-Prem}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/llama-on-prem-economics} } - Comparison
- MindMap Engineering
- Published June 2026
- 10 min read
Choosing Between Llama 3.3, Qwen 2.5 and Mistral for Enterprise: The 2026 Cut
The capability differences between the main open-weights model families are now small enough that the choice is operational and licensing-driven, not performance-driven. The piece gives the cut applied on each deployment and the questions to ask before choosing.
Cite this
MindMap Engineering. (2026). Choosing Between Llama 3.3, Qwen 2.5 and Mistral for Enterprise: The 2026 Cut. MindMap Digital. https://www.mindmapdigital.ai/blog/llama-qwen-mistral-comparison
@misc{mindmap2026openweightsmodelcomparis, author = {{MindMap Engineering}}, title = {Choosing Between Llama 3.3, Qwen 2.5 and Mistral for Enterprise: The 2026 Cut}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/llama-qwen-mistral-comparison} } - Engineering patterns
- MindMap Engineering
- Published April 2026
- 12 min read
RAG on Your Own Servers: Architecture Patterns for Air-Gapped Enterprises
Retrieval-augmented generation inside an air-gapped bank or hospital fails in ways cloud tutorials do not cover. Seven patterns: embedding model selection, vector database by scale, chunking, hybrid search, inference serving, evaluation, and re-ranking with caching.
Cite this
MindMap Engineering. (2026). RAG on Your Own Servers: Architecture Patterns for Air-Gapped Enterprises. MindMap Digital. https://www.mindmapdigital.ai/blog/rag-on-your-servers
@misc{mindmap2026airgappedragpatterns, author = {{MindMap Engineering}}, title = {RAG on Your Own Servers: Architecture Patterns for Air-Gapped Enterprises}, howpublished = {MindMap Digital}, year = {2026}, month = {4}, url = {https://www.mindmapdigital.ai/blog/rag-on-your-servers} } - Regulatory map
- Saurabh Goenka
- Published June 2026
- 10 min read
AI Compliance for Tier-1 Banks: RBI, SAMA, EU AI Act in One Picture
For a bank serving customers across India, the Gulf and the EU, three AI regimes (the RBI Master Direction, the SAMA Cyber Resilience Framework and the EU AI Act) converge on one requirement: model lifecycle artefacts under the bank's exclusive control. The piece maps all three onto one reference stack.
Cite this
Goenka, S. (2026). AI Compliance for Tier-1 Banks: RBI, SAMA, EU AI Act in One Picture. MindMap Digital. https://www.mindmapdigital.ai/blog/bfsi-ai-compliance
@misc{mindmap2026bankaicompliancemap, author = {Goenka, Saurabh}, title = {AI Compliance for Tier-1 Banks: RBI, SAMA, EU AI Act in One Picture}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/bfsi-ai-compliance} } - Analysis
- Saurabh Goenka
- Published June 2026
- 9 min read
Why HIPAA Timelines Are Quietly Pushing Healthcare AI On-Prem
HIPAA permits cloud LLM use under a Business Associate Agreement, but the review can take long enough that US health systems deploy on-premise first. The piece explains why the review is slow and describes the architecture used instead.
Cite this
Goenka, S. (2026). Why HIPAA Timelines Are Quietly Pushing Healthcare AI On-Prem. MindMap Digital. https://www.mindmapdigital.ai/blog/hipaa-pushing-healthcare-on-prem
@misc{mindmap2026hipaaandonprem, author = {Goenka, Saurabh}, title = {Why HIPAA Timelines Are Quietly Pushing Healthcare AI On-Prem}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/hipaa-pushing-healthcare-on-prem} } - Deployment notes
- MindMap Engineering
- Published June 2026
- 10 min read
Sovereign Chatbots for Citizen Services: What 12-Language, Air-Gapped Production Looks Like
What it takes to run a citizen-service assistant fully air-gapped, in twelve languages, with every interaction logged for audit: the multilingual layer, refusing out-of-policy requests, accessibility as a design constraint, and the operating metrics that matter.
Cite this
MindMap Engineering. (2026). Sovereign Chatbots for Citizen Services: What 12-Language, Air-Gapped Production Looks Like. MindMap Digital. https://www.mindmapdigital.ai/blog/sovereign-chatbots-citizen-services
@misc{mindmap2026citizenservicesassistant, author = {{MindMap Engineering}}, title = {Sovereign Chatbots for Citizen Services: What 12-Language, Air-Gapped Production Looks Like}, howpublished = {MindMap Digital}, year = {2026}, month = {6}, url = {https://www.mindmapdigital.ai/blog/sovereign-chatbots-citizen-services} }
Playbooks and whitepapers
Downloadable PDFs, free, sent by email.
- Playbook (PDF)
- MindMap Digital
- 8 pages · free, sent by email
Sovereign AI Playbook
A practical guide to the architecture regulated enterprises use to deploy production generative AI without sending data outside their network perimeter. Six chapters: definition, why now, reference architecture, the regulators, deployment and a decision framework.
Cite this
MindMap Digital. (n.d.). Sovereign AI Playbook. https://www.mindmapdigital.ai/playbooks/sovereign-ai
@techreport{mindmapsovereignaiplaybook, author = {{MindMap Digital}}, title = {Sovereign AI Playbook}, institution = {MindMap Digital}, url = {https://www.mindmapdigital.ai/playbooks/sovereign-ai} } - Whitepaper (PDF)
- MindMap Digital
- 18 pages · v1.1 · free, sent by email
The EU AI Act for Enterprise
What the EU AI Act requires of enterprises with AI in scope: the four risk tiers, the Articles 9–15 obligations, provider versus deployer, penalties, the 2 December 2027 deadline for Annex III high-risk systems, and a 12-point engineering checklist.
Cite this
MindMap Digital. (n.d.). The EU AI Act for Enterprise. https://www.mindmapdigital.ai/playbooks/eu-ai-act
@techreport{mindmapeuaiactwhitepaper, author = {{MindMap Digital}}, title = {The EU AI Act for Enterprise}, institution = {MindMap Digital}, url = {https://www.mindmapdigital.ai/playbooks/eu-ai-act} }
Tools and calculators
Interactive pages that turn the guidance into an answer for your own case.
- Interactive tool
- MindMap Digital
- Free · two minutes
Sovereign AI Readiness Checker
Five questions that map a workload to a reference architecture: air-gapped, in-country private cloud, hybrid core, sovereign edge or cloud-acceptable.
- Interactive tool
- MindMap Digital
- Free · interactive
Cloud LLM vs On-Prem AI: TCO Calculator
Shows where a monthly token volume sits on the cloud-versus-on-premise cost curve, so the cost argument can be tested on your own numbers.
Guides and reference
Evergreen pages the other entries link back to.
- Guide
- MindMap Digital
Sovereign AI: definition, architecture and deployment
The pillar guide: what sovereign AI is, the reference architecture, how a deployment runs and which regulations it answers.
- Reference
- MindMap Digital
The Model Estate: sovereign model families
Named, domain-tuned model families built on permissively licensed open-weight bases and deployed inside the client's perimeter, with the status of each.
- Reference
- MindMap Digital
EU AI Act Articles, in engineering terms
A plain-language reference for the EU AI Act Articles that bind high-risk AI and GPAI: a one-line summary, the engineering implications and a checklist for each.
- FAQ
- MindMap Digital
Enterprise AI FAQ
Short answers to the questions buyers, architects and analysts ask about sovereign AI, the EU AI Act, agentic AI, RAG and document intelligence.
The terms the research uses
Plain-language definitions from the enterprise AI glossary. Each term has its own page.
- Sovereign AI
- An architecture where customer data never leaves the network perimeter, model weights run on customer-controlled hardware, inference logs stay in the customer's SIEM, and the entire stack can operate air-gapped.
- On-premise AI
- Deployment of AI workloads — model serving, embeddings, RAG, fine-tuning — entirely on hardware physically located in customer facilities.
- Air-gapped
- A deployment with zero network connection to the public internet — components cannot reach external services, and external services cannot reach them.
- Bring Your Own Key (BYOK)
- A pattern where data at rest in a cloud service is encrypted with customer-controlled keys, but compute and data-in-use still run on vendor infrastructure.
- Egress control
- Network policies that block outbound traffic from a deployment so no component can call external services, even by accident or compromise.
- Open-weights model
- An LLM whose model weights are publicly downloadable under a permissive licence, allowing on-premise inference, fine-tuning, and full audit of the model artefact.
Sovereign AI research: common questions
What is sovereign AI?
Sovereign AI is an architecture where (1) customer data never leaves the customer's network perimeter, (2) model weights run on hardware the customer controls, (3) inference logs stay in the customer's own SIEM, and (4) the entire stack can operate air-gapped with zero outbound calls. It is not a sovereign cloud region, not a Bring Your Own Key arrangement, and not a hyperscaler deployment in your country's borders.
How is sovereign AI different from BYOK?
Bring Your Own Key gives the customer cryptographic control over data at rest but doesn't address inference, model lifecycle, or the structural dependence on the cloud LLM vendor's continued operation. Sovereign AI closes all four of these gaps. For SAMA, RBI and the EU AI Act, the distinction is the one that matters.
Why do regulated enterprises choose on-premise AI over cloud LLM APIs?
Regulations like the RBI data-localisation circulars, Saudi SAMA framework, UK NHS DSPT, and the EU AI Act require regulated data and the systems processing it to stay under the enterprise's exclusive control. Cloud LLM APIs route prompts through external regions, which counts as a cross-border data transfer that regulators increasingly reject.
What sovereign AI research does MindMap Digital publish?
This page lists 21 pieces: 3 field guides and field notes, 10 analysis and engineering notes, 2 playbooks and whitepapers, 2 interactive tools and 4 guides and reference pages. The lead piece is Sovereign AI Reference Architectures: A 2026 Field Guide. These are practitioner's guides and analysis; none is a survey, a benchmark or a dataset.
How do I cite MindMap Digital research?
Each field guide, analysis and playbook on this page has a "Cite this" reference with the author, year, title and URL, plus a BibTeX entry. Link to the piece's own URL; the URLs and the anchors on this page do not change.
How do I follow new research?
Subscribe to the research feed at https://www.mindmapdigital.ai/research/feed.xml. Every Insights article is also in the RSS feed at https://www.mindmapdigital.ai/blog/rss.xml.
More answers in the enterprise AI FAQ.
For journalists, analysts and researchers
Citing a piece
Open “Cite this” under any field guide, analysis or playbook for an author and date reference and a BibTeX entry. Link to the piece's own URL. The anchors on this page are stable, so you can also link straight to an entry.
The long-form pieces are practitioner's guides and field notes, not studies. They report no survey or benchmark data, so cite them for their analysis and not for statistics.
Following new work
New research appears in the research feed (Atom). Every Insights article is in the Insights RSS feed. The full library across topics is at /research.
Press and interviews
Boilerplate, executive bios and brand assets are in the press kit. Press contact: press@mindmapdigital.ai.
Start with the pillar guide
What sovereign AI is, the reference architecture, and how a deployment runs.