AICFI is being formed to flesh out India’s AI stack, from the energy beneath the silicon to the agents above it. It brings together the companies and institutions that will build it, and keeps the commons they share: the open specifications, tests and resources no single company would build alone.
A GPU on its own does nothing useful. Software runs the model, decides which machine runs it, and decides which model to use at all; applications sit on top. Beneath the plant, power is orchestrated, traded with the grid and bought. Each layer is a real business, with its own customers, and companies exist that do nothing but one of them.
India is building the silicon. Most of the stack around it is open.
The layers that run and coordinate models are open source. What is scarce is people who can operate them well, and that is a hiring and training problem.
The interface every layer will have to meet is being written now, in the open. A standard being written is one you can shape.
The Western stack was assembled for one model, one tenant and chat-shaped traffic. The parts that matter most for agents were built last, or not at all.
Government figures: MeitY in Parliament, March and August 2026; Ministry of Power, July 2026.
Three of these layers are open source and can be adopted here today. The two that route work, the gateway and the model router, decide where work lands and what it is worth, and they have almost no Indian participant. Beside them sit the foundation models the stack serves.
The applications and agents that do the work. The only layer that knows what a request is for.
Built withAgent frameworks, harnesses, applications
Picks a model and a provider for each request, and so reads every request in the clear.
Built withOpen routers such as vLLM Semantic Router; commercial routing platforms
Model-aware load balancing, priority and fairness. It is hardening into the formal boundary between applications and engines.
Built withKubernetes Gateway API Inference Extension, Envoy AI Gateway, llm-d
Spreads one model’s serving across many machines: splits prompt processing from generation, tiers the cache, and places work where its cache already is.
Built withNVIDIA Dynamo, given away free
Choose a replica within one fleet.
Built withvLLM production-stack router, sgl-router
Runs the model on each replica: continuous batching, paged attention, prefix caching.
Built withvLLM, SGLang, TensorRT-LLM
What the stack serves. A model is run at layer 2 and chosen at layer 6, and depends on data, language resources and tokenizers, and on evaluation that tells a buyer what it can do.
Built byIndian foundation-model labs, publicly funded and private; open-weight models from abroad
This is the hardest layer to build and the one India is building. Tenancy, sitting just above it, decides whether that capacity is sold as machines or as a service.
Turns a fleet of GPUs into a cloud: provisioning machines, an isolated cluster per tenant, sandboxing between workloads, metered billing.
Built withInfrastructure-tenancy platforms; hyperscalers’ in-house stacks
The plant that produces tokens. It needs land, power, capital and time, which the software layers do not.
Built withDatacentre operators, colocation, the IndiaAI compute programme
In the West, separate software businesses now sit beneath the plant. In India, this is where the binding constraint is. The numbering runs down from the plant so that layers 1 to 7 keep their names.
Runs power, cooling and compute together inside a fixed power envelope.
Built withFacility-orchestration software, early-stage
The facility as a flexible load on the grid: shifting work in time, and between sites, when the grid is constrained.
Built withGrid-flexibility platforms
Where the power comes from and how it is bought: generation, long-term contracts, captive plants, storage.
Built withUtilities and independent power producers
Five concerns run through the whole stack. Each is handled within a layer and breaks at the boundary between two organisations, which makes each one open ground for a shared specification.
Tracing and evaluation tools are converging on the OpenTelemetry conventions for generative AI. Evaluation in Indian languages and Indian conditions is still thin.
Each guardrail enforces policy inside one layer. Nothing carries a policy across layers.
Not just which request, but on whose authority. There is no settled standard anywhere.
The telemetry exists at the bottom of the stack. No shared contract carries GPU time, batch position or real cache reuse back up across an organisational boundary.
Telemetry and security at the level of the GPU itself. It matters to anyone selling shared GPUs: tenants on the same silicon have been shown to learn what a neighbour is serving.
Alongside them, AICFI improves access to computing and data, maintains the shared software, data and language resources the stack depends on, and supports the organisations building India’s foundation models.
Open specifications for how the parts of the stack connect, written with the companies that will implement them, and tests anyone can run to show an implementation conforms.
Benchmarks and evaluation methods for AI systems, models and agents included, in Indian languages and Indian conditions. Certification against criteria published in advance, decided by an independent board and open to members and non-members on the same terms.
AICFI’s own researchers and engineers, on the shared problems no single company will staff: evaluation suites, Indic datasets, tokenizer efficiency and deployment practice. What they build is published openly.
Research and assessment of AI policy and public programmes, resting on published evidence, and advice to government when asked. It does not represent members’ commercial interests.
AICFI maps the stack, brings the builders together, specifies what sits between them, and measures and certifies what they build. Members and others build the products.
Companies and academic institutions that join at incorporation and shape the work from the start.
Operators, vendors, model labs, applied AI companies, services firms, start-ups and research institutions, joining as the work grows.
Academic institutions join at nil or nominal dues. Researchers sit on the Technical Council and the Research Advisory Board, and work with AICFI’s Lab as staff, fellows and visitors, free to publish what they find.
Individuals take part in working groups in a personal capacity, with no dues. The tests and outputs are open to everyone.
People with experience in government and public policy advise AICFI as individuals. Government bodies are not members.
Companies and academic institutions interested in joining AICFI as charter members are welcome to get in touch.