How Countries Build Sovereign AI: Capabilities, Partners, And Dependence

A government announces national AI. A laboratory releases a model. A company offers a regional cloud. Each can change what people are able to do, but the country name leaves the important questions open. What capability is being built, who can use it,

and which outside relationships still matter? We will compare documented choices in India, Singapore, the United Arab Emirates and France to answer those questions. Lessons 8 and 9 followed public support into usable capacity and traced dependencies through a service. We can now apply

those ideas to real institutions. Keep a public objective in view: enabling people to develop and use suitable AI in their own setting. The examples have different dates and purposes. Identify the ministry, university laboratory or private provider making each claim, and the decisions it can actually make.

Start with the function that matters. A research community may need affordable training time. Speakers of an underrepresented language may need models that handle their language and context. A public institution may need control over where a service operates or a way to continue after a supplier interruption.

These objectives lead to different investments. Ask what changes for researchers, users and operators before treating a larger model or a nearby building as success. India illustrates a public ecosystem approach. In a parliamentary answer dated 20 March 2026, its electronics and IT ministry described subsidized compute supplied through 14 empanelled

service providers and 12 teams shortlisted in phase 1 for foundational-model development. It also reported model launches at the 2026 IndiaAI Impact Summit. Shared access can let research groups use capacity without each financing a complete facility. The service conditions, staff and usable allocation still determine

what a particular team can accomplish. There is also a developer record to inspect. On 6 March 2026, Sarvam said its 30B and 105B models had been trained from scratch, learning their parameters through a new training process, in India using IndiaAI Mission compute. Its acknowledgements identify collaboration

with NVIDIA on software libraries. A public repository lists the 30B model's weight files and an Apache 2.0 licence declaration. This connects public support to a claimed training process and an inspectable release, with hardware and software supply relationships still visible. A teacher using this ecosystem needs an appropriate model, permitted teaching material,

integration and a way to check answers. Researchers need dependable access and people able to operate training. Subsidies use public resources that could serve other purposes. Examine accepted teaching exercises, usable research allocations and continuing support costs to judge who benefits.

Another country may concentrate its development resources by adapting an existing foundation. What capability can that route build? Singapore's government linked its National Multimodal LLM Programme to regional skills, datasets and cultural-language fit in a 24 February 2026 parliamentary reply. A specific historical checkpoint shows 1 development route:

AI Singapore's model card dated 14 April 2025 for Llama-SEA-LION-v3-8B-IT. It describes continued training and instruction tuning for Southeast Asian languages on a Llama 3.1 foundation, with Singapore National Research Foundation funding. The release retains the Llama 3.1 Community License. Trace this checkpoint's exact base and terms;

later family versions have changed. Training from scratch develops capability across a new model's training process. Adaptation can concentrate effort on language, data and use conditions while reusing earlier development. Evaluate each through actual language tasks, permitted data and results checked by appropriate speakers. Inspect rights attached to the

base model and resulting release. A later change of base model changes that relationship. This also raises a practical question for another team: what can it actually obtain and change? The UAE's strategy adopted in 2019 sets 2031 objectives for research, talent, data infrastructure

and AI governance supporting services and wellbeing. Its research ecosystem includes the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence, headquartered in Abu Dhabi with research hubs in Paris and Silicon Valley. IFM's 3 September 2026 K2 Horizon announcement identifies a largest model with

375B learned numerical settings, called parameters, and about 23B active per token, a small text-processing unit. It declares Apache 2.0 for models and code, with separate dataset licences and construction recipes where redistribution is restricted. The public K2 repository lists weight shards. Its linked training-code project has an

Apache licence, and a linked corpus contains downloadable data under its own licence. These artifacts give another team concrete material to inspect and potentially adapt. To assess complete reproduction, investigate the matched training inputs, configurations and process. Some inputs may be described through recipes. Check the particular artifact, version and

permission needed for the proposed work; released examples alone leave complete matched access unverified. The active parameter count describes the part used in a processing step. It does not turn the whole model into a 23B-parameter file or provide an electricity measurement. Deployment must account for suitable equipment,

software, memory and skilled operations. A country can gain freedom to adapt weights while relying on imported processors or other organizations' tools. That leaves a physical question: who will supply the infrastructure needed to operate the model? The UAE also illustrates an infrastructure partnership.

OpenAI's 22 May 2025 Stargate UAE announcement named G42, Oracle, NVIDIA, Cisco and SoftBank as partners and described a 1 GW cluster, with an initial 200 MW expected in 2026. G42's 16 October 2025 update described construction under way through Khazna Data Centers. These are dated records of a partnership,

a target and construction progress. A model release and a construction update address different parts of national capability. Pause and explain this claim: because a country has released a model and announced a 200 MW phase for 2026, that phase must now be operating. What do the records support, and

what additional evidence would establish usable operation? Finish your explanation before continuing. The release can support an artifact-availability claim. The announcement supports a plan, and the later construction account supports its reported milestone. Establishing operation needs appropriate commissioning, acceptance and service evidence for the date in question. Current operation remains unestablished by these dated

records; inspect newer commissioning and service evidence. Once a service exists, a further question follows: which operating decisions can its customers control? France's presidency linked national AI leadership to infrastructure development, electricity, grid and site readiness in its 11 February 2025 agenda. A private provider adds another

layer to that public context. Mistral released Small 3.1 under Apache 2.0 in March 2025. Its June 2025 Mistral Compute announcement described an integrated infrastructure offering with NVIDIA systems and an intention to use decarbonized energy. Public policy and these company commitments sit alongside one another; ownership and

contracts identify who actually controls an arrangement. A later Mistral update, dated 11 August 2026, reported regional inference endpoints as generally available, letting customers select Europe or the United States. It also specified limited, safeguarded transfers to subprocessors outside the selected region. The same update discussed

long-term European compute commitments. Inspect each service's current status, processing locations, allowed transfers and operating terms. Keep future infrastructure milestones distinct from the service available to a customer today. Recall the energy lesson. Megawatts describe power capacity; energy use depends on operation over time.

A site also needs an electricity connection, cooling and a workload. A proposal to use lower-carbon electricity does not give us a measured total for a particular service. Compare the useful workload, demand, location, electricity supply and relevant water conditions. Spending on a large facility may be

justified by an objective, but it also commits resources that could support access, skills or a different service. The examples now connect. India's shared compute supports multiple developers. The Singapore checkpoint directs adaptation toward regional languages. The UAE research release makes artifacts inspectable

while a separate infrastructure partnership addresses scale. The France-based provider links model choice to regional operating arrangements. Each changes a different relationship. Public funding, private ownership, model permissions and operational decisions can be distributed among different parties. Keep those parties named

when explaining who can act. Pause for a comparison. How can adapting a foreign base model strengthen a country's AI capability while retaining an upstream dependency? Contrast the Singapore checkpoint with Sarvam's from-scratch training account. Name a useful capability each

approach could develop and a supporting relationship that remains to be checked. Adaptation can develop language expertise, training skills and a model suited to selected uses, while the base model's licence and technical design remain relevant. From-scratch training develops control across a new training process,

alongside requirements for data, computing and operations. Sarvam's own account names an international software contribution. Compare the capability needed for the objective, the rights and resources available, and evidence about the result. The dependency map makes the remaining relationships inspectable.

Partnership can provide equipment, expertise, access and markets. Its terms can also constrain future choices or make switching costly. Ask who can modify the model, approve data use, allocate capacity, change a service and fund maintenance. Include the people whose languages and information

are represented, and the communities supplying electricity, water and public resources. A credible strategy explains acceptable dependencies, an alternative where interruption matters and an observation that would change the plan. Apply the comparison to a fictional technical-college consortium developing an assistant for local-language teaching materials.

Its worksheet describes a locally operated adaptation route and a managed regional-service route. Model rights, source-material permissions, operating readiness and teacher evaluation are separate supplied conditions or unknowns. No trial results are supplied. Choose a required capability, compare the routes against it and identify

a condition that would change your recommendation. Use the blank worksheet to record the objective, the institutions controlling essential inputs and the evidence for each stage. Include an energy or resource question, a meaningful alternative and a revision trigger. Keep proposed tests separate from observed results. Take time for your own

attempt before using the separate key. We can now read a sovereign AI claim as a set of institutional choices. Identify the public purpose, follow what has actually been released or delivered, and trace the remaining relationships through use and maintenance. Compare the trade-offs against the same purpose. Add that explanation to your

mental model of the AI economy: countries can expand capability through different combinations of local development and international cooperation, and the evidence should show which choices people really gain.