Where India Actually Stands in the Global Artificial Intelligence Race

Understanding AI in India

If you have fifteen minutes and want to actually understand where India sits in the global AI race - this is that write-up.

Most people following AI in India and the global AI race still understand it in fragments. You know Nvidia matters. You've heard of DeepSeek. You've seen the IndiaAI Mission GPU announcements. You've read that GCCs are hiring for AI roles. What's missing isn't more information but the structure that connects all of it. Once you see the structure, every individual headline you read afterward slots into a place, and the picture stops being noise.

So let's build the structure first, then place India inside it.

The AI industry isn't one thing. It's six layers.

Every AI product you've ever used, such as a chatbot, a coding assistant, an enterprise tool, sits on top of a stack most people never see. Six layers, from the ground up:

The six layers of the AI industry, from energy and hardware at the bottom to applications at the top.

To understand where India stands in AI, you have to look at all six layers and not just AI models.

Energy and power - the foundation of AI

Nothing above this layer runs without it, and 2026 has made that uncomfortably literal. AI data centres are turning electricity supply, grid connections and reliable power into strategic constraints.. The constraint is increasingly not simply whether someone can buy a GPU, but whether there is enough power and infrastructure to run it at scale.

That is changing how the market values energy. Data-centre operators increasingly care about the availability, reliability and location of power, and the economics of AI infrastructure increasingly overlap with the economics of the electricity system. Microsoft is paying Constellation Energy billions of dollars over 20 years to restart a dead nuclear plant purely to feed AI data centers. Over 2,300 gigawatts of power projects are stuck in US connection queues, some for up to eight years, and sites near a nuclear plant now command a real estate premium simply for guaranteed power.

AI chips and hardware - where the physical bottleneck sits

GPUs, custom AI chips, memory and advanced manufacturing form the physical layer. Nvidia remains the dominant supplier of accelerated AI compute, with CUDA creating a software ecosystem that makes switching away from Nvidia more difficult than simply buying a different chip.

This is the part of the stack where geography matters enormously. Designing an AI chip is one thing. Fabricating leading-edge semiconductors, producing high-bandwidth memory, advanced packaging and the equipment required to manufacture them is another. Taiwan and South Korea therefore occupy a structurally important and unassailable position in the global AI economy.

Nvidia has become a $5 trillion company and the world's most valuable public company, a remarkable illustration of the financial value being created in the infrastructure layer of the AI boom.

AI infrastructure and cloud compute - turning chips into usable capacity

Data centres, hyperscalers and GPU-rental platforms turn chips into usable computing capacity. This is also where 'sovereign compute' plays out that is countries deciding how much they want to depend on somebody else's cloud and how much domestic capacity they need.

The scale of the buildout is extraordinary. McKinsey estimates nearly $7 trillion of cumulative global data-centre capital expenditure may be required through 2030 to keep pace with compute demand. The infrastructure layer is therefore not simply a supporting service around AI. It is the central investment stories of the technology cycle.

AI data - the training material

The next layer is data. As easily accessible public web data becomes less sufficient for frontier training, the industry is moving toward licensed, vertical and even synthetic data - medical, legal, industrial and other specialised that is harder to obtain and often more valuable per token.

The economics of data are changing at the same time as the legal environment around it. Publishers and platforms are increasingly negotiating licensing arrangements with AI companies, turning data from something many model builders initially treated as a free input into a strategic asset with a price. ews Corp's deal with Meta is reportedly worth up to $50 million a year, while Reddit's licensing arrangements with Google and OpenAI are estimated at roughly $140 million a year combined.

AI foundation models - the layer that gets the headlines

The Intelligence layer

The models are the part of AI most people interact with indirectly. When you ask ChatGPT a question, generate an image, write code or use an AI assistant, a foundation model is doing the underlying reasoning and generation. Large language models (LLMs) such as GPT, Claude, Gemini and DeepSeek are the best-known examples, alongside multimodal models that can work across text, images, audio and video. They are called large language models because they are trained on enormous amounts of text (“large”) to learn the patterns of human language (“language”), using neural-network models with billions of parameters (“models”).

These models are trained at enormous scale and then used as the underlying intelligence for thousands of applications. This is why the model layer has attracted so much attention and capital. But owning a model is only one source of power in the AI stack.

The frontier arms race remains concentrated among a small number of companies in USA and China. At the same time, many governments are increasingly interested in sovereign or domestically adapted models. The motivation is not necessarily to beat the frontier labs at their own game. It is often more practical: reducing dependence on a handful of foreign providers, preserving control over sensitive data, and building models that work better with local languages, institutions and regulatory requirements.

That has produced a genuinely global map of sovereign AI efforts. France has Mistral. The UAE has Falcon. China has built its own large domestic model ecosystem. South Korea is pursuing a substantial national AI programme. India is building indigenous models with a particular emphasis on Indian languages and local use cases.

India's progress is real, but it is important to describe it accurately. Sarvam AI's 105B model, for example, performs strongly on Indian-language benchmarks and is competitive with a range of open and closed models on several general benchmarks. That is a meaningful achievement but it does not, however, make India a frontier-model leader. The more useful description is that India is building strong regional-language capability while still operating below the frontier on the broadest general-purpose systems. More on this later on the section on India…

AI applications - where the trillion-dollar investment has to earn its return

Everything a person or business actually touches i.e.  chatbots, agents, copilots and AI-enabled workflows  sits here. This is where the enormous investment in chips, data centres and models ultimately has to translate into economic value.

The scale of the investment is already extraordinary. Gartner estimates that the world will spend $2.7 trillion on AI in 2026, with more than 45% going into infrastructure. To earn a 25% return on the $6.5- 7 trillion of investment by hyperscalers, AI would need roughly $8.2 trillion of cumulative revenue. That is the part of the AI story that is still unproven. The industry has demonstrated extraordinary demand for computing capacity and enormous willingness to invest ahead of demand (not unlike other technological revolutions in the past but at a much higher scale). It has not yet demonstrated an equivalent scale of economic returns from the applications running on top of that capacity.

Enterprise adoption is moving quickly, but value capture remains uneven. McKinsey's 2026 research finds that 37% of organisations report some EBIT impact from AI, but only 6% qualify as high performers generating significant value. That is why the application layer matters. It is where the AI investment cycle eventually has to close the loop - from capital expenditure, to applications, to productivity, revenue and cash flow.

Coming back to the six layers of AI, an easy way to understand is to look at the oil industry as an analogy. AI has an upstream and a downstream, just like oil -  energy and chips are the wells, infrastructure and data are the pipelines, models are the refineries, and applications are the petrol station.

Right now, the entire industry is fighting over who owns the refinery, while forgetting that refineries only matter once someone's actually buying petrol.

Who is doing what in the global AI race?

Strip away the noise and the global AI map reduces to three real centres of gravity, each strong in a different layer.

The global AI map is better understood as different centres of gravity across the stack, rather than a single race with one winner.

The United States sets the frontier at the model layer and holds enormous application-layer reach into global enterprise and consumer markets. It also has the deepest hyperscale cloud ecosystem, even though much of the physical semiconductor manufacturing on which that ecosystem depends sits elsewhere. Taiwan and South Korea manufacture and supply the physical layer everyone else depends on. TSMC fabricates many of the world's leading chips, while South Korea's SK Hynix is a major supplier of the high-bandwidth memory that makes modern AI hardware possible. Their power is upstream, at the point every other layer has to pass through.

China is the most structurally interesting story. Restricted access to the most advanced Nvidia hardware has pushed Chinese firms to build a parallel hardware and software ecosystem. Huawei is developing domestic AI accelerators and systems, while Chinese model companies such as DeepSeek have demonstrated that competitive models can be built under severe compute constraints.

INDIA IS DIFFERENT. Its certainly has a clear advantage in the sheer capacity to build and reshape technology for the world's enterprises. Its taking a lot of initiatives on each layer and that makes India's position worth examining separately layer by layer as we do in the next section. This is where most commentary either oversells the story or undersells it.

Here's the honest version, one layer at a time.

AI foundation models in India - real progress, but honestly bounded

Sarvam AI, Krutrim and BharatGen are part of a broader push to build indigenous foundation models and language technologies. The IndiaAI Mission said in February 2026 that twelve teams had been shortlisted to develop indigenous foundational models or large language models.

Sarvam's 105B model is a notable example. Its published benchmark results show strong performance across several Indian-language evaluations and competitive performance across a range of general benchmarks.That's a real, non-trivial achievement. But India's position should still be described accurately. On broad frontier capability, Indian models remain behind the leading US or Chinese systems. India has built strong regional-language capability and is building the foundations for more capable domestic models. That is a sensible strategic position.

AI data in India - a real, underused asset

India's multilingual population is one of the most interesting potential data assets in the world for anyone building beyond English-only AI. Hindi alone is not the story. India contains a dense combination of languages, dialects, accents, scripts, cultural contexts and domain-specific knowledge. The opportunity is not simply to collect more data. It is to organise, clean, label and make useful India's multilingual corpus at scale. If that happens, India could possess an important structural advantage for models and applications designed for its own population and for other multilingual markets.

AI infrastructure and compute in India - building fast, from a low base

India is building compute capacity quickly. Under the IndiaAI Mission, more than 38,000 GPUs had been onboarded for common compute by March 2026, with the government subsequently announcing an additional 20,000 GPUs.The capacity is being made available to startups, researchers and academia at subsidised rates. The pace is genuinely fast. But it starts from a small base relative to the US and China, and the critical hardware remains imported. India is building the infrastructure layer faster than it is building the underlying semiconductor supply chain.

AI applications in India - genuinely leading, and not a new story

India hosts a large and increasingly sophisticated Global Capability Centre ecosystem, alongside one of the world's largest technology-services industries. What's changed in the last two years isn't the existence of this advantage but rather it’s what these centres now do.

GCCs have moved beyond back-office labour arbitrage toward owning technology, product, analytics, cybersecurity and increasingly AI mandates for their global parent companies. Deloitte's 2026 India research found that Indian enterprises were reporting higher levels of at-scale AI adoption across several functions than the global average, with particularly high deployment in product development, strategy and operations, and marketing and sales.Add to this a genuinely under-appreciated category — Indian enterprise conversational and voice AI companies such as Uniphore, Yellow.ai, Observe.ai and Haptik, which built export businesses in this space years before 'sovereign AI' became a policy phrase.

This layer is India's actual, durable advantage. It is not based on having invented the best model. It is based on having millions of engineers and operators who know how to turn technology into working systems for large enterprises.

AI chips and hardware in India - the real gap, named plainly

India has launched a serious semiconductor push. Semicon 2.0, approved in July 2026, carries an outlay of ₹1.275 lakh crore. Twelve semiconductor projects had been approved with investment commitments exceeding ₹1.64 lakh crore, and three facilities had already begun commercial production by September 2026. More than a hundred startups are already working on chip design.

This is a genuine, well-funded start. But it is a start. India does not yet have domestic leading-edge fabrication capability for the advanced AI chips used in its compute buildout. The GPUs running India's AI infrastructure are designed and manufactured elsewhere.

This is the layer where the gap between ambition and current reality is widest, and it is worth naming without spin.

Energy and power in India - an underrated constraint and opportunity

India has a large and expanding power system, and its clean-energy capacity is growing rapidly. But AI adds a new type of demand: concentrated, power-intensive data centres that need reliable electricity and increasingly sophisticated cooling. India's data-centre capacity has already expanded sharply, from about 375 MW in 2020 to around 1,500 MW by 2025, according to the government. As AI workloads grow, access to power, land, cooling and grid connectivity will become increasingly important to where India's AI infrastructure gets built.

India's current position across the six layers of the AI stack. 'Building' does not mean weak; it means the capability is being developed from a relatively low base.

AI startups in India: mostly building on top of AI

India’s AI startup ecosystem reflects the same pattern. Most Indian AI startups are building applications rather than competing to train the world’s largest foundation models. They are applying AI to enterprise software, healthcare, finance, content, customer service, developer tools and other specific workflows.

There are exceptions. Companies such as Sarvam and Krutrim are building models and infrastructure as well as applications. But the broader startup ecosystem is overwhelmingly closer to the application layer than the frontier-model layer. This matters because it shows where Indian entrepreneurial capacity is naturally concentrating: not in recreating the entire AI stack, but in turning rapidly improving AI capabilities into products and businesses.

India is also becoming a major AI market

India's importance in AI is not only on the supply side. It is also becoming one of the world's largest markets for using AI.

On the consumer side, India had more than 100 million weekly ChatGPT users by February 2026, according to OpenAI. Sensor Tower estimated that ChatGPT had around 180 million monthly active users in India in January 2026, followed by Gemini at 118 million.

The broader numbers are even more striking. India accounted for 16% of all global GenAI app downloads in 2025, around 600 million downloads, and Indian users spent roughly 6 billion hours on these applications. Yet India generated less than 1% of global in-app revenue from GenAI apps.

That tells an important story: India is already a huge AI market by users, even if it is not yet a huge AI market by revenue.

The enterprise market is moving quickly too. Deloitte's 2026 India research found that 40% of Indian respondents reported significant or full AI usage, compared with about 28% globally. At-scale adoption was particularly high in product development, strategy and operations, and marketing and sales.

This gives India a second source of importance in the AI economy. It is not only a place where AI applications can be built and integrated; it is a market large enough to influence how those applications are designed, priced and deployed.

So, is India winning the AI race?

That's probably the wrong question. India isn't  and cannot compete with the United States, China, Taiwan and South Korea in exactly the same part of the AI stack. Its position is much more specific.

India is relatively strong in applications and AI delivery, increasingly capable in compute and infrastructure, has significant potential in multilingual data, is building indigenous foundation models, and remains heavily dependent on imported advanced chips and hardware.

The more useful question is therefore not whether India will 'beat' another country in AI. It is which parts of the AI stack India can own, and which parts it will continue to depend on others for. India is not competing for the same prize as the major players.

Having said this, India's focus on compute, foundation models and semiconductors is understandable. These are the layers where technological dependence can become strategic dependence, and building indigenous capability provides resilience. India does not need to replicate every capability of the US or China, but in a geopolitical crisis, access to chips, compute and foundational models cannot be taken for granted. The objective is not complete self-sufficiency, but enough domestic capability to have something to fall back on.

India is therefore pursuing two different objectives at the same time: building enough upstream capability to reduce strategic vulnerability, while using its existing downstream capacity to capture the economic value created by the AI boom.

The Honest Synthesis of AI in India

India may therefore capture less of the shareholder value created by AI than the countries building the dominant models and infrastructure. But it can still capture a significant share of the economic value - through employment, wages, services exports and the expansion of technology work around AI.

That's a slower, less headline-friendly story than 'India's ChatGPT moment.' It also happens to be the more useful way to understand where India actually stands in AI.

Understanding India's position in the global AI race also means understanding where investment opportunities lie, how venture capital firms evaluate technology businesses, and what makes a startup worth backing. The VC Academy is a self-paced program that teaches you how Indian funds evaluate companies through investment cases, term sheets and practical deal analysis.