A commons for India’s AI stack.

Some of the stack will not build itself.

Sovereign AI capability is not one model or one company. It is a stack, from the energy and grid beneath the silicon, through facilities, serving engines, orchestration, gateways and routing, up to agents. Datacentres are being built in India and GPUs deployed. Much of the stack above them is still open, and where Indian participation exists, it is thin.

Some of that will resolve on its own: inference engines are open source and improving fast. Other parts are unlikely to arrive by themselves:

• Interfaces between layers run by different organisations. Inside one company they are plumbing. Across many operators and routers they exist only if someone specifies them.
• The routing and gateway layers, which decide where work lands, have almost no Indian participant.
• Language infrastructure with a maintenance commitment. Indian languages cost on average about eight times as many tokens as English, a cost paid in GPU time on every request.
• Measurement that outsiders believe. Self-reported, self-translated benchmarks do not let Indian models prove what they can do.
• Access, not hardware. Using public compute is limited by process and planning capacity more than by silicon.
• Deployment practice. Indian AI already runs at a scale few countries attempt, and the practice for doing it safely and measurably is not written down.

Each is a shared cost that competitors would rather split than each pay in full. That is what a foundation is for.

Four functions, all of them open.

1. Standards and protocols

   Open specifications for how the parts of the stack connect, with tests anyone can run to show an implementation conforms. A specification is final only once two independent implementations pass.

2. Evaluation and certification

   Benchmarks and evaluation methods for AI systems, models and agents included, in Indian languages and conditions. Certification against criteria published in advance, decided by an independent Certification Board.

3. The Lab

   AICFI’s own researchers and engineers, on shared problems no single company will staff: evaluation suites, Indic datasets, tokenizer efficiency and deployment practice.

4. Policy

   Research, assessment and advice on AI policy and public programmes, resting on published evidence. Non-partisan, and not a representative of members’ commercial interests.

Across all four, AICFI supports the organisations building India’s foundation models: open evaluation in Indian languages, maintained data and language resources, tokenizer efficiency, compute access, deployment practice, and a forum for the people building them. The support is open to every model builder on the same terms.

Defined by how it works, not by a fixed list of projects.

The domain is the whole stack. The discipline is that every piece of work advances the same way: proposed in a working group, resourced on evidence that people use it, and closed in public when they do not.

• Convene. Working groups per problem, open to member nominees and to individuals in a personal capacity.
• Specify. Where an interface between organisations is missing, publish an open specification and its conformance tests.
• Steward. Take on the maintenance shared artifacts need: versioned releases, provenance, a defect process.
• Measure. Build open methodology, and publish results with the method that produced them.
• Certify. Assess systems against published criteria, and grant, suspend or withdraw certification in public.
• Research. Staff shared problems in the Lab, and publish what it finds.
• Assess. Where a public decision turns on evidence about the stack, publish the research.
• Enable access. Help researchers, start-ups and institutions use public compute and data.
• Promote. Make Indian work at every layer visible to buyers and to each other.

AICFI’s outputs are specifications, test suites, evaluations, certifications, research, policy assessments and shared artifacts. Products belong to members and anyone else who builds them. Members compete on products and cooperate on the substrate beneath them.

Members fund it. They do not grade their own work.

1. Board of directors

   A founder director, the Executive Director and independent directors. Members are not directors: the organisations whose systems AICFI tests do not govern the body testing them.

2. Technical Council

   Member nominees and independent technical members. Holds final authority over the content of specifications, test suites, evaluation methodology and certification criteria, independently of the board.

3. Certification Board

   Decides individual certifications. No one with a commercial interest in a system under certification sits on it or directs it, and every decision is published with reasons.

4. Research Advisory Board

   Senior researchers advising on the Lab’s agenda and reviewing the quality of its work.

5. Government Advisory Council

   People with experience in government and public policy, advising as individuals. Government bodies are not members.

Interests are disclosed, not hidden. Directors, Council members and officers declare their commercial interests in a public register, updated within 30 days of a change. An undisclosed conflict is disqualifying; a disclosed and constrained one is how standards work has always been done.

A not-for-profit company.

AICFI is being formed as a company limited by guarantee under Section 8 of the Companies Act, 2013, for the promotion of science, research and education. Its income and property are applied only to its objects, and nothing is distributed to members.