Foundation model

Indian Economy glossary

Also called: Large language model, LLM · Topic: Factors of Production, Entrepreneurship and Startups · NCERT: Beyond NCERT

Meaning

A foundation model is a very large AI model trained once on a huge amount of general data, such as text, code, images or speech. It can then be adapted (fine-tuned, i.e. trained a little more on special data) to do many different tasks. A large language model (LLM) is the best-known type. It works mainly with language, and that includes code and reasoning in words.

Why it matters: a foundation model works like general-purpose capital. It costs a lot to build once, but then many firms can use it to make many products. That is why countries treat owning one as a question of technological and strategic autonomy.

Explanation

How it works: parameters, tokens and compute

  • Parameters are the learned weights (the internal settings the model adjusts while it learns). They decide the size of the model.
  • Tokens are small units of text, such as a word or part of a word. The number of tokens shows how much data the model was trained on.
  • Compute is the computing power used for training. It comes from GPUs (graphics processing units), which are special chips for AI work.
  • More parameters and more tokens → more GPU-hours needed → higher cost.

  • So building a foundation model is capital-intensive: it uses a lot of machines and few workers.

  • Worked example (EKA-1):

  • Parameters = 26.3 billion. Training tokens ≈ 7.5 trillion.
  • Tokens per parameter = 7,500 billion ÷ 26.3 billion ≈ 285 tokens for each parameter.
  • Exam lesson: the billion figure is the size of the model and the trillion figure is the size of the data. Do not swap them.

Types

  • Large Language Models (LLMs): work with text. They answer questions, write, translate and code.
  • Large Multimodal Models (LMMs): work with more than one kind of input, such as text, speech and images.
  • Small Language Models (SLMs): lighter models that can run on-device (on a phone or a local computer) without the cloud. Example: Dhīr runs on GPUs, TPUs and mobile phones, and it works offline.
  • Sparse models: only part of the model works on each task. This cuts running cost. EKA-1 is an open, sparse hybrid model.
  • Open models: the weights are released, so others can use and adapt them.

Where it fits among the factors of production

  • Technology as the enabler: technology shifts the production function (the link between inputs and output). With it, a firm gets more output from the same inputs.
  • Like physical capital: the model is a produced means of production. It is also left behind as newer, better models arrive, much as a machine depreciates.
  • Human capital is still needed: engineers build it, and skilled workers turn it into useful products.
  • Entrepreneurship: startups bear the risk of building it.
  • Frank Knight: profit is the reward for bearing uncertainty that cannot be insured.
  • Joseph Schumpeter: new technology replaces old ways of working, a process he called creative destruction.

What makes it cheaper or costlier to build

  • Cost and access of GPUs: subsidised compute lowers the entry barrier for startups.
  • Data availability: a shared data platform cuts the cost for each firm.
  • Model design: small and sparse models are cheaper to run for every query.
  • Import dependence: GPUs are made mainly in one country. A supply cut would stop new training.

In India

  • IndiaAI Mission (MeitY):
  • Approved by the Union Cabinet on 7 March 2024.
  • Outlay of ₹10,371.92 crore over 5 years.

  • Vision: "Making AI in India and Making AI Work for India".

  • Seven pillars: Compute Capacity; Innovation Centre (Foundation Models); Datasets Platform (AIKosh); Application Development Initiative; FutureSkills; Startup Financing; Safe & Trusted AI.

  • Policy model:
  • The state gives subsidised public AI compute and pays for related research costs.

  • Private Indian startups build the models. In economic terms, this is a public–private split: the state provides the costly shared input and the entrepreneurs do the innovating.

  • Selection:

  • The call for proposals (30 Jan 2025) drew 506 applications.
  • Phase 1 startups: Sarvam AI, Soket AI, Gnani AI, Gan AI.
  • 20 proposals were identified for support: 12 LMMs and 8 SLMs.

  • 12 startups were selected across phases 1–2 by Feb 2026.

  • Models released:

  • EKA-1 (Soket): 26.3 billion parameters; covers all 22 scheduled languages plus English; made for BFSI (banking, financial services, insurance) and cybersecurity.

  • Dhīr 2B / 5B (ZenteiQ): lightweight; covers 10 Indian languages plus English; made for defence, government and low-connectivity areas.

  • Compute:

  • Common compute rose from 34,000 GPUs to 45,000+ GPUs by June 2026.

  • 93.18 lakh GPU hours were sanctioned across 237 projects (Lok Sabha, 5 August 2026). That is about 39,000 hours per project on average.

  • AIKosh: the platform for datasets and models. It had 7,541 datasets and 273 models as of Feb 2026.

  • Hardware gap: under the National Supercomputing Mission, India has started designing its own processors based on RISC-V (a free, open rulebook for designing chips).

Don't confuse with

  • LLM vs foundation model: an LLM is one type of foundation model and works with language. A foundation model can also handle speech or images, as an LMM does.
  • Parameters vs tokens: parameters (26.3 billion for EKA-1) measure model size. Tokens (~7.5 trillion) measure training data.
  • GPUs sanctioned vs GPUs used: sanctioned GPU hours (93.18 lakh) are hours allotted. They are not hours actually used. GPUs onboarded show capacity, not use.
  • Model autonomy vs chip autonomy: India owns the model weights, but the GPUs it trains on are imported. "Strategic autonomy" so far covers the model and the data, not the chip.

Prelims Hooks

  • Nodal ministry for the IndiaAI Mission: MeitY, not DST or NITI Aayog. Cabinet approval: 7 March 2024.
  • Innovation Centre (Foundation Models) is one of the mission's seven pillars. AIKosh is the Datasets Platform pillar.
  • EKA-1 is by Soket and covers 22 scheduled languages (Eighth Schedule). Dhīr 2B/5B is by ZenteiQ and covers 10 Indian languages.
  • EKA-1: 26.3 billion parameters, trained on ~7.5 trillion tokens. It is an open, sparse hybrid model.
  • Dhīr runs on GPUs, TPUs and mobile phones and is offline-capable, so data stays on the device.
  • Trap: "Foundation model and LLM mean exactly the same thing." Partly false. An LLM is one kind of foundation model.

Mains Points

  • Capital-intensive technology in a labour-abundant economy:
  • Foundation models need costly GPUs, and only a few workers.
  • They can raise productivity in government, industry, research and education. But automation and AI can replace routine jobs.

  • The answer is to link AI with human capital: skilling through FutureSkills and SWAYAM, so that the demographic dividend is not wasted.

  • State as provider of shared inputs:

  • Subsidised compute lowers entry costs for startups. This is like the state building a road that many firms then use.
  • Risk: if sanctioned GPU hours go unused, the subsidy produces no model.

  • Remedy: publish GPU hours actually used, per model. The OECD blueprint judges compute by capacity, effectiveness and resilience.

  • Sovereign but small: the trade-off:

  • Owning the weights removes dependence on foreign owners for licences and updates.
  • Small, sparse models suit Indian conditions: phones, offline use and low cost per query.
  • But imported GPUs and the missing independent benchmarks limit real autonomy. Building chips through the India Semiconductor Mission and NSM's RISC-V work is the long-term fix.

Read more