Free resource in French · Word + Excel Kit
Sovereignty Kit: Reducing your AI dependencies
Who does your generative AI project depend on, and how far is that acceptable? In 2 hours, map your dependencies, define the level of sovereignty your project needs (from data residency to full control), choose the model and hosting based on evidence, and maintain the ability to switch providers. No flags, no jargon.
The AI Sovereignty Kit for AI Project Sponsors, in brief
- What it is
- A decision-making method for the person leading a generative AI project: map out who the project depends on (computing, hosting, model, platform, application, data and skills), qualify the level of sovereignty it needs (S1 residence, S2 European operation, S3 European control, S4 mastery), choose the model and hosting based on evidence rather than a flag, then contract and prepare the exit, with a decision sheet to be signed.
- For who
- The sponsor of an AI project, not the lawyer or the architect: business directors, sponsors, AI project managers, transformation directors, program managers of Copilot, ChatGPT, Claude, Gemini or Mistral, in groups, banks, insurers, public institutions and industrial companies; and their interlocutors CIO, CISO, DPO, legal and purchasing departments, who will find the questions that will be asked of them.
- Format
- 48-page Word document (the four-step process, the post-decision phase and seven annexes) and 11-tab Excel workbook: instructions for use, mapping, needs, options, licenses, in-house testing, costs (including exit costs), clauses, reversibility, decision, Data Act.
- What it contains
- The six dependency layers and the dependency score; the twelve questions that define the required level and the table of what the law already imposes (SREN law, cloud doctrine at the center, HDS, Data Act, AI Act, NIS2, DORA); the six families of model and hosting options with their achievable level, their limits and their cost order; reading a license in ten questions; the in-house testing protocol in seven steps; the economics of choice (API, dedicated, internal, break-even point, exit cost); the weighted decision grid; the twenty-four contractual clauses with their standard wording; the ten points of the Data Act to check in each cloud contract; the six rules of reversibility architecture and the table of what is portable; the exit plan and exercise; the decision sheet; ten reassessment triggers and seven indicators. In the appendices: the note to the management committee, the thirty-point checklist, fourteen frequently asked questions, a glossary of twenty-seven terms, the dated overview of open models, licenses and hosting providers (October 2026), dated sources.
- What is it based on?
- European texts (Data Act applicable since September 2025 and its clauses on changing providers, AI Act and Omnibus AI Regulation of July 2026, proposed Cloud and AI Development Act of June 2026, Commission's cloud sovereignty framework of October 2025); French texts (SREN law and its decree of April 14, 2026, cloud doctrine at the center, SecNumCloud 3.2 framework and ANSSI catalogue of September 2026, Arcep guidelines of July 2026); the work of the Senate (hearing of Microsoft France, June 2025) and the National Assembly (committee of inquiry on digital dependencies, July 2026); figures from Synergy Research, Cigref, Gartner, EY and Hexatrust; the designation of critical ICT providers under DORA (November 2025) and the status of the transposition of NIS2; publisher announcements (Mistral, Aleph Alpha, OpenAI, Anthropic, Google, Meta, Alibaba, DeepSeek), lifecycle and pricing pages reviewed in October 2026, and AI projects we support in large organizations.
- Price
- Free, in exchange for a professional email address. No unsolicited emails.
- Update
- Version 2, updated October 26, 2026: Data Act made operational (ten points to check and a new tab), clauses increased to twenty-four (including the exit strategy required by DORA), exit cost quantified, and updated offers (Mistral Large 4, Mistral regional inference, NIS2 status). Reviewed every six months, and at each milestone: January 12, 2027 (end of cloud provider switching fees, Data Act), Court of Justice decision on the Data Privacy Framework (Latombe appeal), adoption of the Cloud and AI Development Act, EUCS scheme.
- What it doesn't replace
- The legal analysis from your legal department or your DPO, the risk analysis from your CISO (see the Security and Data kit) nor the advice from your buyers: the kit informs the sponsor's decision, it does not delegate it and does not replace it.
- Go Further
- Security and data kit, sponsor's GDPR kit, AI Act kit from the sponsor, frugal AI kit, responsible AI kit, AI Systems Governance Kit.
The scene had become commonplace by 2026. The pilot was a success, the committee wanted to roll it out more broadly, and someone asked, "What if the vendor withdraws the model, doubles its prices, or gets bought out?" No one had the answer, because no one had listed who the project depended on. Six months later, a model was withdrawn with only two months' notice, a subscription had its quotas halved at the same price, a license changed, and the organization discovered that its prompts were hosted by a third-party provider, its index couldn't be transferred, and its fine-tuning capabilities were only available from the publisher.
This kit empowers the client to fulfill their role without needing to become a lawyer or architect: a four-step process—mapping, qualifying, selecting, and contracting—that takes two hours with the project manager and culminates in a written decision. It doesn't advocate for a purely European or purely American approach; it advocates for the level of sovereignty you require, demonstrated by a test of your documents, and the flexibility to change your mind.
1. The model is no longer the subject; dependence is.
The differences between the best models are now measured within their confidence intervals, and a publisher's mid-range model can reach the previous top-of-the-line model within six months for a fifth of the price. What creates the difference is what's added around the model; what creates the risk is what's beyond our control. The facts below are those on which the kit is based, along with their source and date.
| Date | Fact or figure | What this changes for a sponsor |
|---|---|---|
| 24th July 2025 | Synergy Research: European providers hold 15% of the European cloud market (29% in 2017); Amazon, Microsoft, and Google hold 70%. Asterès for Cigref (May 2025): €264 billion per year for US software and cloud services, representing 83% of large enterprise spending. | The dependency is structural: it is not about avoiding it but about choosing it and limiting it. |
| 10th June 2025 | Before the Senate inquiry committee, the director of public affairs for Microsoft France, when asked about the guarantee that French data would never be transmitted to American authorities without France's agreement, replied: "No, I cannot guarantee it." | Data localization does not extinguish extraterritorial rights. That's what the provider says. |
| September 12, 2025, then January 12, 2027 | The Data Act applies: notice period for changing cloud providers of no more than two months, 30-day transition, data recovery, reduced fees then prohibited from January 12, 2027. In France, transfer fees have been set at zero since the decree of November 17, 2025. | Exiting from the cloud is becoming a right. But you still need to have something portable to use. |
| November 18, 2025 | European financial supervisory authorities designate 19 critical ICT providers under DORA, including AWS, Google Cloud and Microsoft. | For a bank or insurer, the concentration on these providers becomes a subject of supervision: the exit strategy is no longer optional. |
| 14th January 2026 | AWS opens its European Sovereign Cloud in Brandenburg, operated by German entities; analysts note that the parent company remains subject to the CLOUD Act, an additional cost of around 15% and a reduced model catalog at launch. | The "sovereign" offers from hyperscalers address operational matters, not legal ones. |
| April 14, 2026 | Implementing decree for Article 31 of the SREN law: sensitive data belonging to the State, its operators, and six public interest groups (including the Health Data Hub) will be moved to a SecNumCloud or equivalent service, with migration taking place within 18 months. On April 23, the Health Data Hub, hosted by Microsoft since 2019, awarded its hosting contract to Scaleway. | For the public, the required level is set by law; and a migration from the basics takes years. |
| 3 June 2026 | The Commission proposes the Cloud and AI Development Act: four levels of assurance for the public sector cloud, including two requiring European ownership and control, and a criterion of added value of the Union in public procurement. | Europe grades sovereignty into levels; the kit adopts the same logic, S1 to S4. |
| 8th July 2026 | National Assembly inquiry commission: 79% of public software purchases go to non-European publishers, 1.5 billion euros per year; recommendation of a "sovereignty clause" in public contracts. | The sovereignty clause is entering the vocabulary of buyers: you might as well write it yourself. |
| 27th July 2026 | Regulation Omnibus IA (EU) 2026/1744 enters into force: high risk of Annex III postponed to 2 December 2027, article 4 rewritten as an obligation of means; nothing changes for general purpose models, whose obligations have applied since August 2025 with sanction powers since August 2026. | The licensing of a model has become a regulatory object. |
| 11th August 2026 | Mistral opens regional endpoints (inference to choice in Europe or the United States) and announces a European computing capacity that could reach 1 GW by 2030. | A European publisher offers inference residency; however, its terms specify that subcontractors may operate outside the region: request the list. |
| July to October 2026 | Mistral doubles the price of Small 4 (July 6); DeepSeek multiplies its prices by 4.5 with one week's notice (August 13); Nvidia announces the acquisition of Hugging Face for $12.9 billion (September 3); OpenAI halves the quotas of its $200 subscription starting October 30 (announcement of September 29). | The price decrease is not monotonous, the distribution of open models is concentrated: contracts and retained weights must anticipate this. |
| 3 October 2026 | Aleph Alpha releases Kolibri-1: 78 billion parameters, including 3.46 billion active ones, Apache 2.0, German and English, executable on two A100 boards. Mistral Small 4 (March 2026), Gemma 4 (April 2026), gpt-oss (August 2025) and Qwen 3.8 are also under Apache 2.0. | Open models that can be used in business exist; all that remains is to test them on your documents, in French. |
| October 6th 2026 | Mistral presents Mistral Large 4 in public preview (approximately 1 trillion parameters, 52 billion active) and announces the publication of its weights, with possible deployment in a private cloud or on-premises. On the same day, the French law transposing NIS2 (the "Resilience" law) is once again removed from the National Assembly's agenda; it had still not been enacted as of October 26, 2026. | High-end European software is once again available for download: check the license on the day the weights are published. And anticipating NIS2 in contracts avoids the need to renegotiate them. |
2. The four-step process, from mapping to the signed decision
The kit follows a single thread, from assessment to signing. Each step has its own chapter in the Word document and its own tabs in the Excel workbook, which calculates the dependency score, the required level, the eligibility of options, the three-year cost, and the break-even point for the dedicated server. The yellow cells are to be filled in; the gray ones are pre-calculated.
| Stage | What we do | What comes out of it |
|---|---|---|
| 1. Map | List the dependencies across six layers and ask four questions for each: who controls the provider, where does the service run (data, inference, support), is it replaceable and how quickly, and what impact would it have if it stopped? | A score per dependency, the list of critical dependencies, the project overview (tab "1 Mapping") |
| 2. To qualify | Answer twelve questions about the data, the process, the regulatory framework, the time horizon, and the capacity for change; verify what the law already requires. | The required level, S1 to S4, the observed level and the gap (tab "2 Need") |
| 3. Choose | Compare the options that meet the required level based on four criteria: the license (ten questions), the in-house test on 50 to 100 real-life cases, the total cost over three years, and the weighted grid. | One option selected and its alternatives tested (tabs "3 Options", "4 Licenses", "5 In-house test", "6 Costs") |
| 4. Formalize the contract and prepare for the exit | Build the reversibility architecture, require the twenty-four clauses, verify the ten points of the Data Act, write the exit plan, run the exercise, sign the decision sheet | A contract that protects, a proven exit, a signed decision (tabs "7 Clauses", "8 Reversibility", "9 Decision", "10 Data Act") |
3. The six layers of dependency in an AI project
You can't decide on a dependency you can't see. The toolkit categorizes them into six layers, from the most technical to the most underestimated. The last layer, data and skills, is the one that most often blocks a migration: what is neither documented nor owned by your teams cannot be transferred.
| Layer | What it contains | Typical dependency |
|---|---|---|
| 1. Calcul | The chips and servers that run the models (mainly Nvidia; European gigafactories currently in the tendering process) | Invisible to the client, but it is she who explains the quotas and the prices |
| 2. Accommodation | The cloud or the servers where the models run and where the data resides | The applicable law and the actual location of the inference |
| 3. Model | The language model, its weights, its license, its lifecycle | End of life, price change, licensing, availability in a region |
| 4. Platform and tooling | Gateways, agent orchestration, document retrieval, vector database, connectors | Proprietary formats and non-transferable connectors |
| 5. Application | The tool that users see: assistant, business SaaS with AI, agent | The model is chosen by the publisher, not by you: a second-order dependency |
| 6. Data and skills | Your documents, prompts, test sets, indexes, adapters; the people who know how to evaluate and utilize | Fine-tuning that can't be exported, indexes that need recalculating, prompts from a service provider, lack of expertise |
4. The four levels of sovereignty, and why the required level is not the maximum level
The toolkit adopts a flag-free definition: the sovereignty of an AI project is its ability to make decisions, and to change its mind, without being prevented by a third party. This sovereignty is measured in four cumulative levels, aligned with the assurance levels of the Cloud and AI Development Act and the SecNumCloud framework. No option is sovereign in itself: an open model on a US cloud is technologically controlled and legally exposed; a proprietary model on a qualified cloud is legally protected and technologically licensed.
| Level | Requirement | What this guarantees | That does not guarantee |
|---|---|---|---|
| S1 Residence | Data stored and processed (inference) within the European Union, no training on your data, known subcontractors | Basic GDPR compliance, no routine transfers outside the EU | Immunity to extraterritorial law, continuity in the event of a political or commercial decision by the supplier |
| S2 European Operation | S1+ operation and support from the EU, encryption with keys you control, recognized security certification | Routine access impossible without you, traceability of interventions | That an injunction addressed to the parent company may not succeed |
| S3 European Control | S2 + supplier whose capital and control are European (SecNumCloud 3.2 criteria or equivalent) | Legal protection: no third state can impose access or a prohibition | Technological mastery if the model's weights are closed |
| S4 Mastery | S3+ open weights under unrestricted license, stored on your premises, infrastructure operated by you or a European third party, teams capable of evaluating and exploiting | The ability to continue alone | State of the art at any time |
The required level is determined by data flow and process: a project that is S1 for 90% of its uses and S3 for a health data flow is an S3 project, or it can be broken down so that only that flow meets the demanding level. Over-requiring costs months and pushes users towards unregulated tools; under-requiring exposes them to a forced migration that can last for years. The "What the law already requires" table in the toolkit (SREN law, cloud doctrine at the center, HDS, Data Act, AI Act, NIS2, DORA) sets the minimum level before any discussion.
5. Choose based on evidence: license, in-house test, full cost
The toolkit describes six families of options, with the level each can reach, its strengths, limitations, and cost range: an API from a US vendor in a European region (S1), a "sovereign" offering from a hyperscaler (S2, S3 within a qualified scope), an API from a European vendor (S2 to S3), an open model served by a European cloud (S2 to S3), an open model on a dedicated server on a qualified cloud (S3), and an open model on your own servers (S4). Combinations are the norm in 2026: proprietary high-end solutions for demanding tasks, an open European model for volume, all behind a gateway that routes according to the task.
| The evidence | What the kit makes you do | What you must remember |
|---|---|---|
| License | Ten closed-ended questions regarding the license of an open model (commercial use, thresholds, geographical exclusions, redistribution, attributions, patents, training data), and a reference table dated October 2026 | Open weights does not mean open source. Most of the major open models from 2026 are based on Apache 2.0 or MIT; exceptions (Llama 4 and its thresholds, exclusion of EU entities for multimodal models; proprietary licenses for very large Chinese models) should be reviewed before making a choice. And regardless of the license, download and save the weights of the version you choose. |
| The in-house test | A seven-step protocol: 50 to 100 real-life cases, a static environment, candidates from each family, anonymized responses, two human judges and one automated judge from another family, cost per successful task, test repeated every quarter | Public rankings are biased by publishers (study "The Leaderboard Illusion", 2025) and a judge model favors its own group (2026). The only useful metric is yours, and it forms the basis of a routing strategy. |
| The full cost | A three-year comparison: API, dedicated server, internal server, considering your storage volumes, engineering, reversibility, and price risk; the break-even point for dedicated servers in tokens per month | An H100 server will cost between €2,000 and €2,500 per month from European cloud providers in October 2026; the threshold for API usage is between 200 million and 2.5 billion tokens per month, depending on the assumptions. Below this threshold, the API wins; above it, or as soon as confidentiality or version stability demands it, dedicated servers win. And budgeting based on a promotional price is a mistake. |
| The weighted grid | Up to five options across eight criteria (quality in your cases, legal sovereignty, reversibility, cost, timeframe, continuity, security, environmental impact), with automatic elimination of options below the required level | The grid does not decide for you: it makes the decision explainable to the committee, the CISO and the purchasing department. |
6. Formalizing the agreement and preparing for the exit: twenty-four clauses, six rules, one exercise
A contract doesn't create reversibility, it simply acknowledges it. The toolkit therefore begins with six architectural rules: an OpenAI-compatible API gateway between applications and models, versioned prompts and instructions on your end, connectors to the open MCP protocol, preserved source documents (not just indexes), no non-exportable fine-tuning, and a second provider tested on critical data flows. It then provides a table of what is portable (open prompts, test suites, weights, adapters, exported logs) and what is not (fine-tuning with a closed vendor, vector indexes that recalculate, proprietary features, and "latest" aliases that change without your input).
Then come the twenty-four clauses that matter, each with its standard wording: data and inference localization, no training, subcontractors, declared applicable law, contesting government access requests and standstill orders (commitment made by Microsoft for European governments in 2026, therefore negotiable), customer-controlled keys, twelve-month end-of-model notice (compared to 60 days to 6 months in standard policies), six-month price notice, guaranteed quotas, explicit versions, portability and free exit assistance (Data Act, CCAG TIC), tested reversibility, change of control, final sovereignty clause, audit, termination for sovereignty reasons, incident transparency, and, new in version 2, the full Data Act clauses, jurisdictional transparency, data exit fees (zero from January 12, 2027) and the exit strategy required by DORA in finance. The Data Act is subject to a ten-point checklist to be verified contract by contract (two-month notice period, 30-day transition period extendable, exportable data, data recovery, erasure, fees, prior notification, jurisdictional transparency, interoperability), with its two limitations: it does not protect against extraterritorial law, and the comprehensive digital framework that would amend it has not been adopted. The cost of phasing out (export, switchover, dual operation) is calculated and compared to the annual cost to determine if the threat of phasing out is credible. The reversibility plan comprises seven sections and two pages; the phasing exercise, conducted before production launch and then annually, yields three figures: switchover time, quality gap, and roadblocks. The decision form is signed by the commissioning party, with input from the IT department, the CISO, the DPO, the legal department, and the procurement department.
7. What the kit contains
The Word document has 48 pages: the user manual, with what the kit does not replace; part 1, why this kit (a useful definition of sovereignty on three levels, the facts dated 2025 and 2026, why it is the client's responsibility, seven misconceptions); step 1, mapping (six layers, four questions, dependency score, annotated example of an insurer, invisible dependencies); step 2, qualifying (four levels, twelve questions, what the law already requires, six project families); step 3, choosing (six families of options, license reading, residency and inference, sovereign offers from hyperscalers and Mistral, in-house test, economics of choice, weighted grid); step 4, contracting and preparing for exit (six architectural rules, portable or not, twenty-four clauses, the ten points of the Data Act, reversibility plan, exit exercise, decision sheet); the post-decision phase (ten triggers, seven indicators, quarterly review, monitoring). Seven annexes complete the guide: the note to the management committee, the thirty-point checklist, fourteen frequently asked questions, the glossary of twenty-seven terms, the dated overview of open models, licenses, European hosting providers, American publishers in Europe and SecNumCloud providers (October 2026), dated sources and a page to go further.
The Excel workbook "Tools of the kit" has eleven tabs, in the order of the process: "0 Instructions"; "1 Mapping" (one line per dependency, calculated score and reading); "2 Need" (the twelve questions in drop-down lists, the required level, the recommended level, the gap with the observed level); "3 Options" (up to five options, eight weighted criteria, elimination below the required level, ranking); "4 Licenses" (ten questions for three models, verdict, reference table); "5 In-house test" (one hundred cases, three models, three judges, acceptance rate, cost per successful task); "6 Costs" (three years, API, dedicated, internal, break-even point, exit cost and its weight in the annual budget); "7 Clauses" (the twenty-four clauses: required, obtained, gap, action); “8 Reversibility” (components, portability, effort, last test, the three figures of the exercise, the plan in seven sections); “9 Decision” (the pre-filled sheet from the other tabs, Data Act compliance included); “10 Data Act” (the ten points for three contracts, with counters and reading per contract).
Download the AI Sovereignty Kit for AI Project Sponsors (Word + Excel, free)
48 pages, the four-step process and seven appendices, plus the 11-tab tool workbook. Written by NextStart.AI based on European and French texts in force in October 2026 (Data Act, AI Act and Omnibus IA, SREN law, SecNumCloud), the work of the Senate and the National Assembly, publishers' announcements and price pages, and the AI projects we support in large organizations.
- The six levels of dependency, the dependency score, and the twelve questions that determine the required level, S1 to S4
- Understanding a license in ten questions and a dated overview of open models, hosting providers, and sovereign offerings
- The in-house testing protocol, the three-year cost comparison with exit costs, and the weighted decision grid
- The twenty-four clauses with their standard wording, the ten points of the Data Act, the reversibility architecture, the exit exercise, and the decision form to be signed
8. Frequently Asked Questions
Should we ban American suppliers?
No. The toolkit doesn't rank suppliers by flag but by level achieved and evidence provided. An S1 project is well served by an American vendor in the European region; an S3 project is not, regardless of the model's quality. Forrester predicted in October 2025 that no European company would completely abandon American hyperscalers by 2026: the challenge is to choose and limit dependence.
Our data is with Microsoft, within the EU Data Boundary. Are we sovereign?
You are at S1 level, and S2 if you control the keys and have secured the contractual commitments for 2026. You are not at S3 level: Microsoft remains subject to US law, as its representative confirmed before the Senate in June 2025. This applies to both Copilot Chat and Microsoft 365 Copilot (often called "Copilot Premium"). For most office uses, S1 or S2 is sufficient; for highly sensitive data, it is not. In the financial sector, Microsoft has been listed as a critical ICT provider supervised under DORA since November 2025, which does not exempt you from developing your own exit strategy.
Is an open model necessarily inferior?
Not necessarily, and not everywhere. For everyday office tasks, the open-source models of 2026 perform on par with mid-range proprietary solutions for a fraction of the price; however, for long-form reasoning, complex code, or certain tasks involving formal administrative French, high-end proprietary solutions sometimes maintain an advantage. Our in-house testing determines the best approach on a case-by-case basis and allows for a hybrid strategy.
What is the value of a "sovereign" offer from a hyperscaler?
This is evidenced by its scope. AWS European Sovereign Cloud governs local operations, not the parent company's jurisdiction; S3NS is SecNumCloud certified for infrastructure but not for generative AI as of October 2026; Bleu is not yet certified. Ask each provider about their S level certification and demand proof: certification, capital, contract.
Is dedicated hosting always more expensive?
Up to a certain volume, yes: 2026 analyses place the threshold between 200 million and 2.5 billion tokens per month, depending on engineering and pricing assumptions. Beyond that, or as soon as confidentiality, latency, or version stability require it, dedicated servers win out. The "6 Costs" tab calculates your threshold based on your volumes.
Does the Data Act allow us to leave whenever we want?
It gives you the right to switch cloud providers with up to two months' notice, a 30-day transition period, 30 days to retrieve your data, certified data erasure, and no fees starting January 12, 2027. It does not provide a replacement model, portable components, or end-of-model notice: these are addressed by the reversibility architecture and the terms of the data kit. Check the ten points of the Data Act in each contract (see the "10 Data Act" tab).
Is Mistral a sovereign option for a business foundation?
Often, this is only true if the chosen offer is precisely qualified. Mistral is a French company, outside the direct scope of the CLOUD Act, but its capital also includes non-European investors. Its regional endpoints, launched on August 11, 2026, maintain inference in Europe, but its terms and conditions stipulate limited transfers to subcontractors outside the region: request the list. To our knowledge, Mistral is not itself SecNumCloud qualified; for S3 level, run its models on a qualified cloud or on your own premises, which is possible thanks to its open weights and private deployment offerings.
What is the difference between this and the Security and Data kit and the GDPR kit?
The Security and Data kit determines whether a project can proceed, go ahead, go ahead with conditions, or no go, based on its risks; the GDPR kit addresses personal data compliance. This kit tackles a question that neither kit asks: who is responsible for the project, what is acceptable, and how to exit the situation? The three kits complement each other.
Who signs the decision form?
The sponsor, because the required level, the reversibility budget, and the acceptance of residual dependencies are business decisions. The IT department, the CISO, the DPO, the legal department, and the procurement department all contribute their input.
9. Going further
This kit is the "dependencies" counterpart to the other kits from the sponsor: the Security and data kit to decide whether to go, go with conditions, or no go in light of the risks, the sponsor's GDPR kit for the compliance of personal data, the AI Act kit from the sponsor for AI literacy, transparency and high risk, the AI and Social Dialogue Kit for the information and consultation of staff representatives, the frugal AI kit for the environmental footprint and the responsible AI kit for bias, transparency, and human oversight. The AI Systems Governance Kit provides the register in which the decision sheet is recorded. Organizations that want to map their dependencies and prepare the decision with their CIO, CISO, procurement, and business stakeholders in half a day can rely on the mission Governing your agents and the AI training in companies.
Read also : AI Project Sponsor Security and Data Kit, GDPR Kit for AI Project Sponsors, Sponsor's AI Act Kit: The 2026-2027 Roadmap et AI systems governance: the 2026 toolkit.
The entire series of free kits. To drive adoption: AI charter template, AI training plan template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, AI-powered time-saving kit, Soft Skills Kit for AI, Adoption Roadmap Kit, AI maturity grid, AI Measurement and ROI Kit, AI Governance Kit, Cowork kit, 50 Copilot M365 prompts, Copilot Studio kit, Deploy Microsoft 365 Copilot kit, Copilot Chat Kit, Copilot Excel kit, Copilot Teams and Outlook kit, Copilot Word and PowerPoint kit, Configure Claude kit, Deploying Claude in the Workplace Kit, Claude Agents Kit, Claude Skills kit, Claude's Marketing Skills, Skills Claude HR, Claude Plugins Kit, Claude Code kit, Claude Design kit et Vibe coding kitFor the AI project sponsor: AI Act kit, GDPR kit, AI and Social Dialogue Kit, frugal AI kit, Security and data kit, Shadow AI kit et responsible AI kitBy profession: AI HR kit, AI marketing kit, AI finance kit, AI kit for lawyers, AI kit for salespeople, AI toolkit for project managers, AI innovator kit, communicator's AI kit, AI Procurement Kit, AI Kit for Researchers et AI kit Customer service.
10. Sources
The figures and positions cited on this page and in the toolkit are based on the following dated sources. Information is current as of October 2026; the publishers' pricing and lifecycle pages are dynamic and should be updated each time they are used.
- European Commission, Data Act (Regulation (EU) 2023/2854), applicable since September 12, 2025; report of the expert group on standard cloud contractual clauses, April 2, 2025.
- European Union, regulation (EU) 2026/1744 “Omnibus AI”, July 8, 2026, effective July 27, 2026; European Commission, guidelines for general-purpose AI model providers and GPAI code of good practice, July 10, 2025.
- European Commission, proposed Cloud and AI Development Act, June 3, 2026; Cloud Sovereignty Framework v1.2.1 (DG DIGIT), October 20, 2025; proposal for EU Public Procurement Act, September 9, 2026; EuroHPC JU, call for AI gigafactories, July 30, 2026.
- DINUM, "Cloud at the center" doctrine, circulars of 2021 and 2023; law no. 2024-449 (SREN), article 31, decree no. 2026-272 of April 14, 2026 and DAJ technical sheet, May 6, 2026; Arcep, decree of November 17, 2025 (zero transfer fees) and Guidelines July 2, 2026.
- ANSSI, SecNumCloud 3.2 framework and catalogue of qualified services, edition of September 15, 2026; S3NS, PREMI3NS qualification, December 17, 2025; OVHcloud and NumSpot, September 2026 qualifications.
- Senate, hearing of Microsoft France, Commission of Inquiry on Public Procurement, June 10, 2025, and Report of July 9, 2025; National Assembly, Commission of Inquiry on Dependencies and Vulnerabilities in the Digital Sector, Report of July 8, 2026.
- Synergy Research Group, European Cloud Providers' Local Market Share Now Holds Steady at 15%, July 24, 2025; Asterès for Cigref, Technological dependence on American software and cloud services, May 2025; Gartner, survey of 241 CIOs in Western Europe, November 12, 2025; EY and Hexatrust, Digital Sovereignty Barometer, September 30, 2025; France Digitale and EY, 2026 Barometer, September 9, 2026; Forrester, European Predictions 2026, October 8, 2025.
- Microsoft, One year on: progress on our European digital commitmentsApril 29, 2026; AWS, Opening the AWS European Sovereign Cloud, 14 January 2026; CISPE, letter from 25 leaders to the Commission, 17 March 2026; The Register, correction of Microsoft before the British Parliament (International Criminal Court), 18 February 2026.
- Mistral AI, Introducing Mistral 3 (December 2, 2025), Mistral Small 4 (March 16, 2026), Regional inference and Mistral Compute (August 11, 2026) Mistral Large 4 (October 6, 2026) and model lifecycle policy; TLT, designation of critical ICT providers under DORA, November 18, 2025; AlephAlpha, Kolibri-1 sheetOctober 3, 2026; OpenAI, Introducing gpt-oss (August 5, 2025), Deprecations and data residency in Europe; Anthropic, Model deprecations ; Google, Gemma 4 under Apache 2.0 (April 2026); Meta, Llama 4 Community License (April 5, 2025); Open Source Initiative, Open Source AI Definition 1.0, October 28, 2024.
- Techcrunch Nvidia confirms it will buy Hugging Face for $12.9 billionSeptember 3, 2026; Fortune, DeepSeek price increase, August 13, 2026; OpenAI DevDay, September 29, 2026 (The Next Web); OVHcloud, AI Endpoints catalog; Scaleway, Model-as-a-Service award; Clever Cloud, final sovereignty clause, June 16, 2026.
- Singh, Hooker et al., The Leaderboard IllusionarXiv, April 2025; Who Judges the Judge?, arXiv 2603.04033, March 2026; DINUM and Ministry of Culture, compar:IA Anthropic Donating the Model Context Protocol, December 9, 2025.
- General Court of the EU, Case T-553/23 Latombe, 3 September 2025, and appeal C-703/25 P; EDPB, letter to the Commission, July 31, 2026; US Department of Justice, CLOUD Act Resources ; Brennan Center, Section 702 resource page, September 17, 2026.
