Synopsis
India is striving to build its own artificial intelligence capabilities, aiming for self-reliance and global export. However, the nation faces significant challenges including underinvestment in computing power and a late start in advanced AI model development. This ambition is crucial for technological independence as global powers leverage AI for geopolitical advantage.Listen to this article in summarized format
The moment, part curiosity, part performance, signaled something larger. India, long cast as the world’s back office for code and customer service, was ready to chart its own course in artificial intelligence, in its own languages, for its own 1.45 billion people.
But India’s ambition to build and export a sovereign AI template across the world is colliding with structural constraints: years of underinvestment in compute capacity, a late start in building the most advanced AI models, deep reliance on foreign cloud providers and a venture capital culture wary of the vast, risky bets that define the global AI race.
At stake is more than technological prestige. As the US and China transform artificial intelligence into a form of geopolitical power — controlling chips, cloud access and the models themselves — countries like India risk becoming dependent on systems they do not own and may not always be able to control.
The world’s most populous nation and a leading digital economy, India is becoming a test case for whether a middle-power country can build AI on its own terms. Its successes — or failures — could offer a blueprint for much of the Global South, where governments face similar trade-offs between technological independence and reliance on foreign platforms.
“Every country wants sovereign AI,” said Abhishek Singh, the bureaucrat who helped shape India's early AI self sufficiency policy as the first chief executive of the IndiaAI Mission. “It's as strategic as atomic energy or space technology.”
Artificial intelligence is fast becoming foundational infrastructure, akin to electricity or telecommunications, but with far greater implications for national security and economic power. Across the world, governments are recalibrating to adjust priorities accordingly. China has built a state-influenced AI ecosystem spanning chips, cloud and applications, while aggressively rolling out low-cost AI models and services to businesses and consumers at home, and exporting its models as far as Africa. The US, led by companies such as OpenAI, Microsoft Corp. and Alphabet Inc.’s Google, continues to dominate both research and deployment, disseminating its bots, tools and models into much of the global economy.
India is attempting something more complex: carving out a third path, building an indigenous ecosystem spanning public computing infrastructure, homegrown multilingual models and applications. It’s all built on the country’s deep engineering talent, which powers a $315 billion outsourcing industry and may now offer an alternative to the gravitational pull of the US and China.
“India can’t and won’t out-OpenAI but that shouldn't be the strategy either,” said Ian Bremmer, president of geopolitical advisory, Eurasia Group. “The real question is whether India can occupy the bottlenecks and scarce complements that will become more valuable as intelligence gets cheaper: energy, manufacturing depth, privileged data, the ability to deploy AI at scale in the physical world.”
Beyond market share, concerns also touch on a familiar question in India’s economic history: who ultimately captures value as Indian users generate the data that trains global AI systems? India generates nearly 20% of the world’s data but has less than 5% of its AI computing power. In an update of the 2000s-era slur “software coolies,” drawing on a colonial-era term for exploited manual labor, critics warn that Indians risk becoming “unpaid interns” in the global AI economy. Some industry executives frame the risk more starkly, as a modern version of an old pattern.
“The fear is that for India, if neglected, AI could be a repeat of the East India Co. model where the British colonized India, ransacked cotton farms and sold back cotton fabric,” said Sunil Gupta, chief executive of Yotta Data Services. “We can’t allow foreign firms to use our datasets and our skills to build their IP and sell the AI back to Indians.”
Officials have grown wary in quieter ways, too. When Amazon.com Inc. offered to host India’s national archives at no cost, framing it as a CSR initiative, according to people familiar with the matter, officials recoiled at the symbolism, and the strategic exposure.
The tensions are visible in the rise, and stumbles, of Sarvam AI, the Bangalore startup behind the eyeglasses donned by Modi. Building large language models in more than 22 Indian languages, Sarvam has been cast as a potential cornerstone of India’s AI future, with applications spanning government services, enterprise software and even defense.
In the weeks following Modi’s high-profile visit to Sarvam’s expo booth, the company ran into a colder reception from investors. It knocked on the doors of some of the world’s most prominent backers — SoftBank Group Corp., Prosus NV, Andreessen Horowitz LLC and General Atlantic LP — and came away empty-handed. Accel Ltd. walked away late in the process, people with knowledge of the discussions said, citing concerns about the company’s late entry into the space and its business model.
Sarvam is now stitching together about $300 million at a valuation of $1.5 billion, said the people, raising a respectable sum but one that pales beside the tens of billions commanded by Western peers. Investors include Nvidia, which declined to lead the round but participated.
India may have the talent and the ambition, but does not yet have the ecosystem to finance them at scale. That gap is especially stark in frontier AI, the most advanced corner of the space, where companies often operate without profit for years. “US VCs are willing to fund hundreds of billions of dollars on loss-making AI model builders,” said Sharad Sanghi, chief executive of Neysa Networks Pvt., an Indian AI-focused cloud company backed by Blackstone. But India doesn’t attract that risk-tolerant capital, he said.
The funding gap persists despite the glossy optics of Modi sharing the stage with global AI leaders such as Google Chief Executive Sundar Pichai, OpenAI founder Sam Altman and Anthropic CEO Dario Amodei at the India AI Summit in February. India announced more than $200 billion in AI infrastructure commitments — $110 billion from Reliance Industries Ltd. and tens of billions more from Google, Microsoft and Amazon. Most of that capacity will be built and operated by foreign cloud providers. Even so, the country’s deep pool of engineering talent continues to fuel a growing startup ecosystem. A recent report by research firm Zinnov identified 100 leading AI startups in India, with nearly half at a key inflection point.
But building the most advanced AI systems requires industrial-scale investment, with training runs costing $100 million to $1 billion and the infrastructure behind them running into the hundreds of billions.
India could still build competitive systems with fewer resources, said Sanghi of Neysa Networks, pointing to the country’s space program, which has reached the moon and Mars at a fraction of the cost of Western missions, and far below even the budgets of Hollywood blockbusters The Martian or Gravity. Indian LLM builders are already producing frontier-scale models at remarkably low cost, aided by modest funding, government-backed compute access and cheaper engineering talent.
The constraint is structural. The cloud — the backbone of modern AI — remains overwhelmingly foreign. Core public infrastructure runs on global providers: DigiLocker, India’s digital identity document vault, and DigiYatra, its airport facial-recognition system, are hosted on Amazon Web Services. Programs spanning rural livelihoods, customs and the labor ministry's careers portal all run on Microsoft’s cloud.
“The cloud game was won by foreigners,” said Yotta’s Gupta. In the AI era, the cloud is not merely where data is stored; it is where AI is trained, deployed and monetized, he said.
The risks have already surfaced. Last year, Microsoft blocked Russia-backed Indian refiner Nayara Energy from accessing data stored on its own cloud servers, citing sanctions and triggering a weeklong standoff that rattled Indian policymakers. “Nayara was a reality check,” said Abhishek Singh, a reminder that owning data is not the same as controlling it. Roughly a third of Indian government systems run on foreign cloud providers, according to a person familiar with the matter.
Government programs aim to expand computing capacity and develop domestic chips, but those efforts are years away from closing the gap. A flagship project at IIT Madras, one of the country’s top engineering schools, is working toward building India’s own high-performance, energy-efficient 7-nanometer commercial processor, though it is not expected before 2028. Critics say progress remains slow.
India’s late start in building large language models has also become a handicap. When Sarvam released its first major model last year, trained on France’s Mistral, it drew criticism for relying on foreign technology. Silicon Valley investor Deedy Das, a partner at Menlo Ventures, called it “embarrassing” that a flagship Indian model was built on Mistral and drew just 23 downloads in its first two days.
Efforts have since accelerated, with newer models trained on local data, including one with more than 100 billion parameters, a measure of an AI system’s scale. But the gap with the most advanced systems remains wide. BharatGen, a sovereign LLM builder, has outlined plans for a trillion-parameter model but has yet to receive government-backed GPUs — the specialized chips used to train AI models. Like many Indian AI startups, it lacks the funding to secure compute at scale on its own and is relying on India’s AI Mission, which subsidizes access by paying data-center operators directly.
For smaller startups trying to push the boundaries, the obstacles can feel existential. Jithin V.G., 32, co-founded Bud Ecosystem, which is trying to make AI cheaper to run. The company has developed an architecture that it says could allow large AI models to operate on standard laptops, potentially bringing generative AI within reach of hospitals, government agencies and small businesses. Working with Intel Corp. and Advanced Micro Devices Inc., Bud estimates it could cut the cost of running AI models, or inference, by nearly 90%.
But companies like Bud can take years to mature, and funding is scarce. “India doesn’t back such startups,” said Jithin. “We focus heavily on building for a billion people. If we don't own the AI, how can we control its distribution?”
India is not alone in its ambition. Japan, South Korea and the UK are pursuing their own versions of sovereign AI, each strong in parts of the stack but not across the whole. Japan is betting on indigenous models, with companies like SoftBank partnering with NEC, Honda and Sony to build homegrown AI foundation models and cut reliance on fragile supply chains. South Korea is drawing on its long tradition of industrial policy in a bid to become a top-three AI power. The country is leveraging strengths in advanced manufacturing, including memory chips and displays through companies like Samsung Electronics Co. and SK Hynix Inc., while funding a national competition among AI developers vying for billions of dollars in subsidies, a contest dubbed the “AI Squid Games.”
Rising global tensions are also driving a push for self-reliance. “More and more countries are realizing how critical AI will be within every strategically important sector, and want to position themselves to be fully independent of geopolitical turbulence,” said Helen Toner, who heads Georgetown University's Center for Security and Emerging Technology and is a former board member of OpenAI.
India’s linguistic diversity, long a challenge for AI developers, is also an advantage for startups like Sarvam. Models that work across 22 official languages at scale are difficult to replicate, even as companies like Google and Anthropic expand their efforts.
The ability to build AI for a multilingual population is reinforced by the India’s Digital Public Infrastructure, which underpins services like Aadhaar’s unique IDs, UPI’s digital payment interface and DigiLocker, enabling what officials describe as a “presence-less, paperless, cashless” system for more than a billion people. It is among the world’s most ambitious state-built digital architectures, studied by governments from Singapore to the European Union. But even as New Delhi articulates its sovereignty ambitions, foreign AI is becoming more deeply embedded in India's digital life.
Google’s Gemini has partnered with Indian education platforms to help students prepare for the JEE, one of the country’s most competitive engineering entrance tests, once cleared by its own chief executive Sundar Pichai. OpenAI has teamed up with Mukesh Ambani’s JioHotstar, allowing tens of millions of users to navigate content in multiple languages, and has integrated its tools into Indian audio platform Pocket FM’s network of more than 300,000 creators. Models from Anthropic and OpenAI are also embedded in the enterprise offerings of Asia's two largest outsourcers, Infosys and Tata Consultancy Services, deployed in software delivered to customers worldwide.
Pricing has accelerated that shift. Google’s Gemini starts at 199 rupees, or about $2.09 a month, while OpenAI’s ChatGPT Go has been offered free for a year — levels domestic competitors struggle to match. This kind of pricing strategy, long used by tech giants in large emerging markets like India, blankets users with cheap or free services, leaving local firms little room to scale and ultimately discouraging them from building competing offerings.
And a February budget granted a 20-year tax holiday to foreign cloud providers, initially giving firms like Amazon Web Services and Microsoft Azure an advantage over domestic competitors. Following pushback from Indian data center companies, Minister of Electronics and Information Technology Ashwini Vaishnaw later said local operators would receive the same treatment, putting them on equal footing. Amazon, Microsoft and Sarvam did not respond to emails seeking comment.
India also faces a shortage of computing power, with fewer than 80,000 high-end GPUs, because building large-scale data centers is expensive, access to advanced chips remains constrained, and returns on AI infrastructure are still uncertain. The US has roughly a hundred times as many. For startups, that gap translates into delays and workarounds that include renting slivers of capacity where they can find it.
Some analysts say the strategy is deliberate, allowing foreign firms to capture market share and the data that will improve their models over time. But a report by Bernstein described India’s open-door approach as risking the crowding out of domestic players before they reach scale.
The stakes extend well beyond economics.
Unlike oil or semiconductors, AI capabilities cannot be stockpiled. Each query — whether from a citizen accessing a government service or a military analyst reviewing satellite imagery — runs on infrastructure controlled by a handful of firms in two countries. Both Washington and Beijing have demonstrated willingness to weaponize that access.
For India, the economic stakes are high as it seeks to more than double its $4.3 trillion economy by 2030. The systems it builds, and the ones it relies on, will shape how artificial intelligence is used across its economy and institutions for years to come. The moment at the AI summit in February suggested a country ready to shape its own technological future.
“India's advantage is its ability to operationalize AI at population scale -- through public digital infrastructure, local language ecosystems, and real-world deployment,” said PeiChin Tay of the Tony Blair Institute for Global Change. The next phase of AI, she argues, will be defined less by who builds the most powerful models than by who deploys them most effectively. “Even if India cannot match the US or China at the frontier, it could shape how AI is adopted, and who benefits from it at scale.
