How Mumbai is fast emerging as India's AI capital

MUMBAI

It is a fantastic time to learn, said Prof Raghavan B. Sunoj. So fantastic, even machines were learning.

We were at IIT Bombay’s KReSIT building―its grand atrium, vaulted ceiling and Penrose-like stairs mirroring the unusual name. Odd, yet functional.

I first met Sunoj in his chemistry department office, at one end of the institute’s “infinity corridor”―a half-kilometre passage linking academic departments and symbolising IIT Bombay’s transdisciplinary ethos. Since KReSIT, home to the computer science and engineering (CSE) department, lay outside this corridor, he led the way. He wanted to introduce me to someone.

Lean and energetic in his fifties, Sunoj has spent 21 years at IIT Bombay. His commitment to students―staying up late in hostels and mess halls to help them―earned him the nickname “the Night Professor”. In 2023, it also won him the National Teacher Award from President Droupadi Murmu. A computational chemist, he used mathematical models to solve chemical problems. Six years ago, curiosity led him to artificial intelligence.

“I found many parallels between my field and AI,” he said. “I love mathematics, and as a computational chemist, I had a bird’s eye view of diverse concepts―maths, statistics, probability, linear algebra―all converging in AI. Fascinating!”

As we walked through KReSIT, Sunoj recounted how he learned AI through Stanford’s online courses. Stanford mathematician John McCarthy had coined the term ‘artificial intelligence’ in the 1950s, transforming an obscure discipline―“automata studies”―into a serious research area. The courses made Sunoj fluent enough to engage in deep discussions with his KReSIT colleagues over tea. “When you talk to priests,” he said, “you should at least know the Bible.”

His expertise grew markedly. Last year, he was appointed Google Cloud Chair at IIT Bombay―the only such position in India―for his work in using generative AI to predict complex chemical reactions. When I asked him about the perks, he was matter-of-fact: “There was a notional top-up to my salary, but more importantly, we are collaborating with Google on AI applications in molecular biology and drug discovery. I have molecular machine-learning expertise that Google, as such, may not have.”

We reached the third floor―home to the institute’s rapidly growing AI wing, the Centre for Machine Intelligence and Data Science (C-MInDS). Sunoj, now part of C-MInDS, said the centre initially borrowed faculty from other departments―its current head, Prof D. Manjunath, came from electrical engineering. “Now we are directly hiring faculty with primary affiliation to C-MInDS,” he said. “You will meet one today.”

Navigating KReSIT’s labyrinthine stairs and corridors can be disorienting. Sunoj paused, checked a door number, seemed to calculate something mentally, and then took off in a new direction. He finally found the room he was looking for.

Inside was Arpit Agarwal, who had joined C-MInDS in September after stints at Harvard, Columbia, Google and Meta. Originally from Agra, Agarwal embodied the paradox of many Indians in their thirties―both cynical and hopeful. He had left a prestigious but, as he put it, “limiting” AI research role in New York for something more fulfilling.

After Sunoj introduced us and left, I asked why.

“That’s exactly what my friends at Meta and Google ask: ‘Why leave industry, where the real action is?” Agarwal said. “I tell them―if all AI talent is concentrated in companies like Google and Meta, who will study their impact on society?”

For decades, academia led AI development. But in recent years, industry has taken over. A 2023 study by the Massachusetts Institute of Technology found that seven of ten AI PhDs chose private industry over academia. Two decades ago, the number was just two. The reasons are clear: higher salaries, access to superior computing power, and larger data sets. As a result, industry AI models are now, on average, 29 times larger than those in academia. The biggest AI models each year now emerge from the private sector 96 per cent of the time.

Agarwal, too, joined industry after his postdoc at Columbia, working at Meta AI in New York. “Meta has several AI labs,” he said. “One of the largest is in New York, and I was part of Yann LeCun’s team there.”

LeCun, an AI pioneer and longtime New York University professor, joined Meta as its chief AI scientist when Mark Zuckerberg sought to pivot Facebook into an AI giant. LeCun had two conditions: that he keep teaching at NYU, and Meta’s AI models remain open-source, enabling academia-industry collaboration.

So why leave New York for Mumbai?

“I always wanted to be in academia,” Agarwal said. “It offers the freedom to study AI’s societal impact, not just build models. Also, while IIT Madras had a head-start in AI research, Mumbai now has the strongest AI talent concentration across academia and industry.”

Mumbai is quietly emerging as India’s AI capital. It has distinct advantages―financial muscle (home to 39 Nifty 100 companies, with spending capacity of $590 billion), technological infrastructure (over half of India’s data centres), and a vast talent pool (20 lakh technology professionals). It also has a growing AI startup ecosystem, which in 2022 had birthed India’s first AI unicorn (Fractal Analytics), and is expected to attract $450 billion in FDI.

At the heart of Mumbai’s AI transformation is IIT Bombay. It has steadily climbed global rankings―breaking into the top 100 in industry reputation in the 2025 QS World University rankings. A remarkable achievement, as all except three in the top 100 are decades or centuries older than IIT Bombay, now 67.

IIT Bombay’s AI journey began in the 1960s with a Soviet-built Minsk II computer―bulky and outdated. But its very obsolescence was a boon. Unlike IIT Kanpur, which received a high-tech IBM computer that even IBM India engineers were not allowed to tinker with, IIT Bombay students and faculty could freely experiment.

Today, a state-of-the-art supercomputing facility drives AI research at C-MInDS.

Agarwal’s focus at C-MInDS, like at Meta, is on recommendation systems―the AI behind Netflix suggestions, Spotify playlists, and Amazon’s “Customers who bought this also bought” feature. While these systems keep users engaged, their unintended consequences are now under scrutiny.

Several US states are suing Meta, accusing it of fuelling a teen mental health crisis by making Instagram addictive. Amazon is facing allegations that its algorithms recommended sodium nitrite―a food preservative lethal in high doses―to depressed teenagers, spiking cases of people buying the compound to harm themselves.

“The problem,” Agarwal said, “is misaligned objectives.”

He described a movie where an AI, tasked with “saving earth”, starts eliminating humans. “The AI objective misaligned with human interests,” he said. “Industry AI could pose similar risks. That is why independent research in academia is crucial.”

The current AI boom was born in university cities after decades of stagnation.

Before 2012, there was an “AI winter”―a period of reduced funding and interest that began in the 1980s. Companies abandoned research, governments cut funding, and scientists avoided the term “artificial intelligence” to avoid being seen as dreamers.

Then everything changed.

In 2012, a Toronto team led by British researcher Geoffrey Hinton trained an AI model to recognise images. A descendant of George Boole (Boolean algebra) and George Everest (Mt Everest), Hinton was a psychologist, neuroscientist, and an intellectual rebel: he had left the US for Canada in the 1980s, disillusioned by Ronald Reagan’s policies and military dominance of AI research.

He approached AI the way babies learn. When a baby sees a cat, neurons in her brain fire up in a distinct pattern, which becomes the brain’s recognition of a cat. If the neurons fail to reproduce the pattern when the baby is again shown a cat, she misidentifies it―perhaps as a dog. But her parents correct her, refining her neural network until she gets the pattern―and the answer―right.

Hinton’s team trained a neural network―a mathematical model mimicking the brain―like a baby. It was fed pictures from a vast database called ImageNet, until the neural network could recognise not just cats and dogs, but also ants, elephants, great white sharks, and thousands of objects.

“This was a major breakthrough,” said Agarwal. For decades, computer vision improvements came in tiny increments―just 0.1 per cent. Hinton’s model delivered a stunning 20 per cent accuracy boost. Their nine-page paper on their neural network’s “deep learning” revitalised AI. “Suddenly, AI was back,” said Agarwal. “The data had improved, the computing power had improved.”

Though sceptics questioned whether the model would work in real-world settings, Hinton built consensus with support from LeCun, whose New York team had developed a similar AI model. Recognising the commercial potential, Google spent $44 million to acquire a company Hinton’s team had formed, while allowing them to maintain university ties. LeCun joined Facebook, and more researchers moved from academia to industry.

The boom has benefited India, which now ranks fifth globally in AI development. According to Stanford’s annual AI index, India is behind only the US, China, the EU and the UK. India’s share in AI research, measured through AI journal publications, grew from 1.3 per cent in 2010 to 5.6 per cent in 2021. Stanford says Indian developers now contribute more than 24 per cent of all projects on the code-sharing platform GitHub―above the EU and the UK (17.3 per cent, combined) and the US (14 per cent).

Between 2013 and 2022, India launched 296 AI startups―seventh highest globally, ahead of Japan and Germany. This startup boom seems to be reversing brain drain: before 2000, nearly 40 per cent of IIT Bombay graduates moved overseas. (The KReSIT building was named after one of them―Kanwal Rekhi, the first founder-CEO from India to take a company public on Nasdaq. He founded the Kanwal Rekhi School of Information Technology, which merged with CSE in 2006.) Thanks to startup opportunities, more than 80 per cent IIT Bombay graduates now stay in India, with some like Bhavish Aggarwal of Ola launching AI unicorns like Krutrim.

The power shift in AI research became obvious in 2017 when Google published Attention is All You Need, a paper detailing a breakthrough in language processing.

Google researchers had been trying to teach machines language, breaking it into basic components such as words, phrases and punctuation. They fed these components into neural networks, hoping they would ‘digest’ the relationships between words and ideas. But language’s inherent complexity challenged AI.

As philosopher-historian Yuval Noah Harari wrote in Sapiens, human language is “amazingly supple”―“We can connect a limited number of sounds and signs to produce an infinite number of sentences, each with a distinct meaning.” Unlike humans who naturally understand context in sentences like ‘The dog could not cross the road because it is blind’ and ‘The dog could not cross the road because it is wide’, machines struggled to determine how the meaning of ‘it’ in the two sentences change. Google’s paper introduced a solution―a new neural network, called the “transformer”, that “pays attention” to key parts of a sentence to decode meaning.

The transformer tracked not just words, but also the relationships between them. They trained on raw internet data―petabytes of English text from websites―turning context into a probability game. The more data they processed, the better they became at grasping context, giving rise to large language models (LLMs) capable of generating human-like text.

From a baby recognising images, AI became a toddler who talked.

LLMs―such as OpenAI’s ChatGPT and Meta’s Llama―have been delivering groundbreaking results across fields. AlphaFold, an LLM-like model developed by DeepMind (which Hinton helped Google acquire) solved a 60-year-old science challenge―predicting protein structures. The achievement earned its creators the 2024 Nobel Prize in Chemistry. That same year, Hinton received the Nobel Prize in Physics for his AI contributions.

LLMs are also reshaping geopolitics. Reports emerged last November that researchers linked to the Chinese military had developed ChatBIT, an AI model optimised from Meta’s Llama. ChatBIT could rapidly process battlefield data―terrain adversary movements, real-time threats―giving China an AI-driven advantage in warfare.

Israel’s Gaza offensive has already demonstrated AI applications in sifting through intelligence data, tracking militants and selecting kills.

“When the Israelis bomb Gaza,” said Harari in an interview to THE WEEK, “it is increasingly AI that tells them, ‘You should bomb this building, and that person.’ So much authority is being given to AIs―it’s a kind of arms race now.”

With AI advancing faster than expected, the UK hosted the first AI Safety Summit at Bletchley Park in November 2023, where 29 governments and industry leaders discussed regulation.

“At Bletchley, we championed two principles,” Rajeev Chandrasekhar, who represented India as minister of state for IT, told THE WEEK. “AI must be inclusive―it should not deepen the divide between haves and have-nots. And, the world needs a shared understanding of its risks and safeguards.

A month later, India hosted its own summit, where delegates endorsed the New Delhi declaration, reaffirming AI safety commitments. “At Bletchley and in New Delhi,” Chandrasekhar said, “we stood for AI that is inclusive, accessible to the Global South, and never weaponised.”

Last year, recognising AI’s strategic importance, India committed Rs10,372 crore for the IndiaAI Mission to build AI infrastructure, develop indigenous models and support startups. “We had asked for Rs15,000 crore,” said Chandrasekhar. “The money sanctioned is the first tranche for our AI infrastructure.”

A key initiative is creating Indian datasets to reduce dependence on western models. Early this month, India launched AI Kosh, a library of non-personal data that can help companies build foundational LLMs.

India is shaping its strategy by observing global regulatory trends. “We cannot halt AI progress or eliminate all risks,” said Agarwal. “The first step is identifying the most pressing risks for society. The European Union is particularly active in this space.”

The EU’S Artificial Intelligence Act, passed in August 2024, imposes strict regulations. It has established the European Artificial Intelligence Board, which will ban high-risk AI applications, impose transparency requirements on companies, and prohibit social-scoring systems which evaluate individuals based on personal data.

Supporters argue that it protects privacy and public interest, but critics warn of the “Brussels effect”―strict regulations stifling innovation, creating funding bottlenecks, and slowing economic growth. EU startups have struggled, with data laws delaying access to OpenAI and Meta’s models, while US and Indian users adopt them earlier.

In contrast, the US has favoured a hands-off approach, letting companies set their own guardrails―a “Washington effect” that prioritises innovation and competition.

“Neither model suits India,” said Rajan Luthra, a distinguished fellow at the Observer Research Foundation in Mumbai. “The EU’s stringent rules have hurt its startups―just compare the number of unicorns in Europe and India. The US approach, meanwhile, risks underestimating AI’s dangers.”

Luthra and Tehilla Shwartz Altshuler of the Israel Democracy Institute recently wrote a report proposing a “Mumbai-Tel Aviv effect”―a alternative regulatory model for developing economies that balances innovation and oversight. The report―titled Mumbai and Tel Aviv Effect: An Alternative to the Bandwagon Effect of Brussels and Washington in Global AI Regulations―lists a few similarities. Like Israel, India has avoided formal AI legislation. Both countries are wary of overregulation stifling AI’s potential in health care, particularly diagnostics, drug discovery and risk assessment.

“India and Israel share similar AI ambitions: harnessing AI for economic growth and social progress,” Luthra said. “Israel’s startup ecosystem is impressive on a per-capita basis, while India ranks third globally. A shared focus on responsible AI could define a new regulatory path.”

Geopolitical tensions further shape AI policy. At the AI Summit in Paris―the second after Bletchley and the first co-hosted by India―US Vice President J.D. Vance underscored America’s stance. “I am not here to talk about AI safety,” he said in a blunt 15-minute speech. “I am here to talk about AI opportunity.” A former venture capitalist, Vance criticised the EU’s “excessive regulations” and said America wanted to accelerate AI development. “To restrict AI development now,” he said, “would mean paralysing one of the most transformative technologies in generations.”

The US’s aggressive push, and its refusal to sign a pledge to make AI safe, have unsettled researchers. “We can’t look to the US for leadership anymore,” said Arjun Bhagoji, a researcher who left the University of Chicago for C-MInDS in February. “India must craft sensible, actionable regulations through consultation with academia, industry and civil society.”

Certainly, India and Israel cannot follow China’s path. Despite rapid AI progress, Chinese firms face accusations of violating international laws. In January, Deepseek’s R1 model gained attention for its affordability, but critics alleged unlawful access to US tools and misrepresentation of costs. DeepSeek claimed R1 cost under $5.6 million, while OpenAI spent $100 million on its latest ChatGPT model―raising suspicions about undisclosed computing resources.

As stakes rise, Harari is urging countries to collaborate to remain competitive. “Even a huge country like India cannot compete alone,” he told THE WEEK. “The US and China are far ahead. Countries must join forces to level the playing field.”

Luthra, citing the Mumbai-Tel Aviv model, sees collaboration as vital for AI governance. “India has many smart policymakers, but too few AI experts,” he said. “Good regulation demands more than just intelligence―it requires deep technical understanding.”

The rapid adoption of AI demands regulation. In 2022, Maharashtra passed the Universal AI University Act, establishing India’s first AI-dedicated university in Karjat near Mumbai―the world’s third, after institutions in France and the UAE.

Founded in 2009 as a business school by former Thomson Reuters executive Tarundeep S. Anand, the university now offers “AI-embedded” courses across disciplines. Anand recognised AI’s potential during his time at Reuters, where “hedge funds consumed truckloads of market data and used algorithms to predict outcomes”. “By 2005,” he told THE WEEK, “algorithmic trading was the norm in Chicago.” A meeting with two Stanford professors solidified his conviction: “I thought, this is really transformative.”

One of the university’s innovations is an AI-powered psychology programme. Students input a patient’s social, financial and educational background into an AI model trained on demographic and psychological data. When they add symptoms like depression or anxiety, AI predicts possible remedies. “The more the data, the better the prediction,” said Anand. “By year three, the model might be accurate four out of 10 times. In five years, with one lakh patients, maybe eight or even 10 out of 10.”

But AI’s power comes with risks. “In the wrong hands, the consequences could be devastating,” Anand said.

At C-MInDS, Agarwal highlighted the emerging danger of AI-generated videos that are hypnotically addictive. “These videos are trained on content that hooks children,” he said. “AI optimises visuals―weird shapes, colours, sequences―hacking into a child’s neurological system. They simply cannot look away.”

This underscores the need for regulation. “We need guardrails―rules that ensure that AI is used responsibly by governments, businesses and society,” said Anand.

According to Bhagoji, history offers a sobering parallel. “When cars were first invented,” he said, “they had no seat belts, no airbags, not even brakes. As one of my colleagues noted, it took decades for safety regulations to catch up―despite people dying.”

Can we afford the same mistake with AI?

Venu Gopalakrishnan, CEO of IT firm Litmus7, which is developing an AI ecosystem for IndiaAI, has a deeper concern. He shared an unsettling incident: a customer service bot his company created for a retailer invented a fictional home pickup service to placate an irate customer. “There was no such service; the AI simply made it up,” he said. “Imagine this kind of fabrication happening in not just customer service, but in health care, finance or national security.”

The problem, he said, is that AI is racing to become artificial general intelligence. “AGI is the real AI, when machines become truly intelligent,” he said, “unlike today’s AI, which cannot really think on its own―it just mimics thinking, and therefore is not truly intelligent.”

According to Venu, when AGI arrives―possibly within two years―it will not just understand context, as LLMs currently do, but will start creating it. “They will absorb all the knowledge created by eight billion people, learn in real time, and recreate themselves constantly to apply the learning in real time,” he said. “That is when regulation becomes not just important, but absolutely essential for our survival.”

A Mumbai legend tells the story of Haji Ali Bukhari, an Uzbek merchant who renounced his wealth for sainthood. His white-domed shrine on an Arabian Sea islet remains an iconic landmark.

As the legend has it, Haji Ali once met a weeping woman who had spent her last coins on cooking oil, only to spill it after tripping on a stone. She wept fearing her husband’s wrath. Moved, he pressed a finger into the earth, and oil gushed forth, refilling her vessel. But remorse followed―he had wounded the earth.

Today, Mumbai thrives on a different kind of oil: data.

“Maharashtra holds 60 per cent of India’s data centre market,” Chief Minister Devendra Fadnavis said at Davos recently. “Data’s value now surpasses oil.”

If oil once powered machines, data now fuels machine intelligence. With 90 crore internet users averaging 24GB monthly, India generates 20 per cent of global data but hosts less than 3 per cent of data centres. AI is driving unprecedented growth―India added a record 230MW of capacity in 2024, and 2025 is projected to exceed that figure.

Mumbai leads this surge, commanding 55 per cent of India’s data centre market. Its edge lies in robust telecom infrastructure, a stable power grid, and the highest number of undersea fibre optic cable landing points in the country. “Three crucial undersea cable projects landing in Mumbai could be completed in 2025,” notes the services firm Cushman & Wakefield, “positioning India’s financial capital as a regional data hub.”

AI has triggered a data centre gold rush. Companies are racing to ‘scale’―training models with ever-increasing data and computational power. Since 2012, AI training demands have skyrocketed―an OpenAI estimate said its 2018 models required 3,00,000 times the compute than Hinton’s ImageNet model.

Tech giants are hyperscaling―expanding massive data centre networks to meet AI’s insatiable appetite. Last year, Amazon, Alphabet, Meta and Microsoft spent $200 billion on hyperscaling; this year, that figure is expected to exceed $300 billion. At this pace, their four-year investment could surpass $1 trillion―a quarter of India’s GDP.

Yet, AI computing remains prohibitively expensive for most Indian businesses. Hyperscalers pass down infrastructure costs, charging up to $32 per hour for AI computing, stifling innovation and AI research in both academia and industry, while forcing to surrender data without adequate protections.

This threat to political and technological sovereignty drives India’s push to build its own AI ecosystem. On March 6, marking the IndiaAI mission’s first anniversary, IT Minister Ashwini Vaishnaw launched a compute portal to democratise AI access. “The portal will provide access to more than 18,000 GPUs to students, startups, researchers, academia and government departments,” he said.

GPUs (graphics processing units) form AI’s backbone. They perform thousands of calculations in parallel, helping CPUs (central processing unit) handle deep learning. If CPUs are head chefs managing kitchens, GPUs are sous chefs, preparing at scale. With demand soaring, GPUs have become the world’s most sought-after tech commodity.

Nvidia, controlling 70 per cent of the high-end GPU market, became the world’s most valuable company last year. Yet Nvidia still cannot produce chips fast enough to meet demand. Geopolitical tensions complicate matters: the US has imposed export restrictions on advanced AI chips, aimed at China, which are also affecting India. The restrictions limit the sale of high-end chips to India to just 50,000 units―total, not per year.

“Right now, 50,000 is a big number―it takes a lot of money to buy them,” said A.S. Rajagopal, CEO of NxtGen Datacenter and Cloud Technologies. “But India spends just 2 per cent of what the US does on AI. So, if we really get serious, this bottleneck will hurt.”

The global GPU race has been accelerating. In 2022, Meta bought 16,000 GPUs to train Llama. By 2024, Microsoft alone secured 4,85,000, while Meta (2,24,000), Amazon (1,96,000) and Google (1,69,000) followed. Despite restrictions, Chinese giants like ByteDance and Tencent still managed to amass nearly five lakh AI chips.

With global powers investing hundreds of billions, India’s allocation resembles the woman’s last coins spent on oil. “If India can’t get enough GPUs, we will have to buy compute power from players abroad at 10 to 15 times the price,” said Rajagopal. “This is a dangerous situation.”

IndiaAI has tried to address this gap by commissioning companies such as NxtGen to procure GPUs. Of 18,693 planned GPUs, about 10,000 are already operational, offering researchers and startups subsidised access at roughly $1 per hour.

“Initially, they are going to give access to at least 90 institutions,” said Rajagopal. “Something very good will come out of this.”

Yet, like Haji Ali’s remorse about wounding earth, AI exacts environmental costs. Microsoft reported that AI operations increased its water consumption by 34 per cent. AI models demand enormous power, and frequent hardware upgrades generate electronic waste.

“The larger the data, the more expensive the training,” said Prof Sunoj at IIT Bombay. “Keeping AI systems running and cool requires enormous electricity. If you are an environmental purist avoiding flights to cut carbon, you should not use pre-trained models―because their training consumed massive power.”

Governments are responding. India and France recently signed a deal to develop small modular nuclear reactors for powering AI data centres. Meanwhile, companies innovate with cooling technologies like synthetic diamond―the most thermally conducive material―to prevent AI chips from overheating.

Last December, US startup Akash Systems partnered with NxtGen to bring this technology to India. “Akash was using diamond cooling in satellites,” Rajagopal said. “I realised that removing dozens of power-hungry fans inside server racks could cut energy use by 40 per cent―and extend GPU life by keeping them cooler.”

The approach is elegantly simple. “We place a diamond layer between the CPU and cooling unit. Diamond transfers heat five times faster than copper. Cooler CPUs and GPUs last longer, letting us extract every bit of computing power.”

Power supply remains another challenge. Whether India can manage to keep up with the demands of the AI race remains to be seen. “There have been instances where towns or cities near large data centres suffer from electricity shortages,” said Bhagoji. “Power-sucking data centres are concerning as India builds foundational models. We must insulate the rest of the grid from their impact.”

For Rajagopal, technological self-sufficiency is critical to India’s AI sovereignty. “Data sovereignty means keeping our data in India, safe from foreign laws,” he said. “If we truly want to protect this country, we need to control our own AI future.”