Indians Rule Big Tech. Why Can’t India Build?
The UAE shipped six generations of frontier AI models in three years. India, with 140x the population, just shipped its first.
India produces 9.2% of the world’s AI research papers [7]. It ranks second globally in AI skill penetration [7]. Indian-origin executives run Alphabet, Microsoft, IBM, Adobe, Palo Alto Networks, and Snowflake, companies collectively worth over $10 trillion as of this writing [8]. The country graduates 5 million STEM students annually [9]. And yet India’s total private AI investment in 2024 was $643 million, roughly 170 times less than the $109.1 billion invested in the United States [7].
The UAE, with a population of 10 million, has built the world’s most prolific open-source AI program since 2023, with its latest model punching far above its weight class [10]. India, with 1.4 billion people and the world’s largest technical workforce, shipped its first homegrown foundation model at frontier scale on February 18, 2026 [61].
Two companies show how.
In January 2024, Ola founder Bhavish Aggarwal announced Krutrim, India’s first AI unicorn, valued at $1 billion before shipping a product. He invested $230 million of his own money [1]. The pitch was irresistible: India’s massive engineering base, its linguistic complexity, its billion-plus population. Surely the country that runs Silicon Valley could build its own foundation model.
By December 2025, Outlook Business reported Krutrim was “in crisis.” Senior engineers had departed. Layoffs had hollowed the team [2]. The Kruti AI assistant app had 100,000 downloads. ChatGPT had over 110 million users in India [3]. Aggarwal himself acknowledged: “We’re nowhere close to global benchmarks yet” [2]. Fundraising stalled.
Three weeks later, a different Indian AI company told a different story. On February 18, 2026, Sarvam AI (founded by two IIT researchers, funded at $53.8 million [4], built on government-subsidized GPUs at ₹65 per hour [5]) launched Sarvam-105B. It was 105 billion parameters, trained from scratch, designed for India’s official languages. The company claims it outperforms DeepSeek-R1 on certain benchmarks at one-sixth the size [6].
One founder approached AI like a Bangalore unicorn play: invest big, hit valuation milestones, iterate on the product later. The other came out of academic research, built with patience and public compute, and shipped something real. That split is India’s entire AI story in miniature. The fact that even the success story required government life support to exist tells you everything about the structural problem.
This gap is not accidental. I visited India several times during my years at AWS, and what struck me most was the sheer intensity of the tech community: the curiosity, the ambition, the speed at which engineers absorbed new ideas. I came away each time thinking this was a country on the verge of building something enormous. That it has not, not yet, is what makes the structural question worth asking. The questions that only India can answer, applying AI to 22 languages and a billion people outside the English-speaking internet, are the ones that make the structural analysis worth doing.
Not to diminish what India has accomplished, but to understand why the output has not matched the energy. The answer is the predictable result of an economy that was optimized, over three decades, to build for everyone but itself.
The talent pipeline has a leak at the top
India’s AI workforce looks formidable on paper: between 416,000 and 800,000 professionals in AI roles, depending on how broadly you count [11]. The IIT system admits roughly 17,000 undergraduates annually through one of the world’s most competitive entrance exams. Eleven IITs now offer dedicated AI and data science degrees. Stanford’s AI Index ranks India first or second globally in AI skill penetration, with concentration growing 263% since 2016 [7].
A devastating NBER study tracking the top 1,000 scorers on the 2010 IIT entrance exam reveals the leak. Of the top 100 rankers, 62% have left India, overwhelmingly for the United States. Among the top 10, nine out of ten now work abroad [12].
The salary math makes this almost inevitable. An entry-level AI engineer earns $10,000–$18,000 in India versus $70,000–$120,000 in the US, a 7–8x gap [13]. At the senior research scientist level, the differential can exceed 10x when stock compensation is included. Indian nationals received 283,397 H-1B approvals in fiscal year 2024, accounting for 71% of all H-1B visas issued [14]. The Gulf is a growing draw as well: as detailed below, the UAE and Saudi Arabia recruit senior AI researchers with tax-free compensation packages that rival Silicon Valley, without the visa uncertainty. Fewer than 3% of IIT graduates join critical Indian institutions like ISRO or DRDO [12].
The brain drain is not merely quantitative. Carnegie Endowment research from February 2025 found that India has predominantly low- and mid-tier AI talent (implementers and application developers), while top-tier researchers who design novel architectures end up in US labs [15]. The single most consequential example: Niki Parmar, from Pune, co-authored “Attention Is All You Need,” the 2017 paper that created the Transformer architecture underpinning every major language model. She now works at Anthropic in San Francisco [67]. Only 0.08% of Indian engineers pursue AI PhDs, compared to 4.2% in China [15]. India’s talent flow index stands at -0.18 (a net exporter of talent), while the US sits at +6.29, and even China maintains a +0.38 [7].
This is not a one-way loss. India’s diaspora sends back over $125 billion in annual remittances, the largest inflow of any country. Indian-origin executives at Microsoft, Google, and Amazon are part of the reason those companies are now committing tens of billions to Indian infrastructure. Diaspora angel investors fund a disproportionate share of Indian deep-tech startups. These return flows are real. They do not, however, substitute for having the researchers in-country doing the foundational work. Remittances do not train models.
India does not have a talent shortage. It has a talent retention crisis, and it is losing precisely the people who build foundational things. The salary gap is the obvious explanation. The deeper one is career optionality: an AI researcher at Google DeepMind has a credible path to co-founding a startup, joining a venture firm, or leading a competitor’s lab. An AI researcher at TCS has a clear path to senior management. India’s ecosystem does not offer the career graph that retains ambitious people, and that connects directly to what happens next.
The services DNA problem
The brain drain matters, but it is a symptom of something deeper: India’s IT industry was built to deliver software services, not to invent.
India’s IT services industry generates $283 billion in annual revenue, and TCS, Infosys, Wipro, and HCLTech are its anchors, collectively employing over 1.4 million people and accounting for roughly $74 billion of that total [16]. They shaped an entire generation’s career aspirations around efficient execution, predictable returns, and client delivery, not open-ended research with uncertain outcomes. This is not a criticism; it built the modern Indian middle class. It also created an ecosystem with a specific gravity that pulls everything toward services.
None of India’s major IT companies is building foundational AI models from scratch. TCS has built four small language models of 2.5 billion parameters each, all fine-tuned on Meta’s Llama [17]. At the AI Impact Summit itself, the event designed to showcase Indian AI sovereignty, TCS announced a partnership with OpenAI, and Infosys announced one with Anthropic. Both formalized their roles as integration layers for American frontier models [18]. All four IT giants deployed 50,000+ Microsoft Copilot licenses each in late 2025 [19]. They are deployers and integrators, not builders.
Nandan Nilekani, Infosys chairman, made this explicit at a Meta summit: “Let the big boys in the Valley do it. We will use it to create synthetic data, build small language models quickly” [20]. This is not surrender. It is a coherent bet that frontier models will commoditize and the value will accrue to whoever applies them best at population scale, though it is worth noting that Nilekani runs a $19 billion services company whose business model depends on that bet being correct. It also leaves India permanently dependent on someone else’s infrastructure layer, and it only works as long as the models remain open and the pricing stays favorable.
Sarvam’s founders made the opposite bet. Vivek Raghavan spent 12 years as a volunteer on Aadhaar’s biometric systems and co-founded AI4Bharat before deciding a company was the right vehicle. Pratyush Kumar left faculty positions and research roles at IBM and Microsoft. Both took below-market compensation to build from scratch in India. As Raghavan told Outlook Business on February 18, a country that lacks ambition will only build small things [64].
They are the exception. In my experience sitting on tech company boards, the reason exceptions stay exceptional is that the incentive structure actively punishes their choices. Every engineer who joins Sarvam at Indian wages is forgoing a career path that would pay five to ten times more and offer a richer ecosystem of next moves.
Venture capital follows the same logic. Indian VCs prefer predictable returns, application-layer bets over foundational research. India saw only 74 new AI startups in 2024, versus 1,073 in the US, 116 in the UK, and 98 in China [7]. This undercounts India’s applied AI activity (hundreds of companies deploy AI in fintech fraud detection, TB screening, agricultural yield prediction), but it reinforces rather than undermines the diagnosis: the energy is almost entirely in the application layer. The largest Indian AI funding round, Sarvam AI’s $41 million Series A, would not rank in the top 500 globally [4]. As Mohandas Pai of Aarin Capital put it: “Who will give $200 million to a startup in India to build an LLM?” [21].
India’s total R&D spending has been stuck at 0.65% of GDP for decades, below that of every major Asian peer: South Korea at 4.9%, Japan at 3.3%, and China at 2.4% [22]. The services DNA is not just a cultural preference. It is an economic equilibrium that the entire incentive structure reinforces.
Infrastructure: the global context changes the story
The human capital flows toward services. The physical infrastructure tells a similar story.
India’s data center capacity reached 1.5 GW by end-2025, approximately 1.2% of global capacity, compared to the US at 53.7 GW [23]. The country generates about 20% of the world’s data but stores only a fraction domestically. Building a data center requires navigating 30+ separate clearances across central and state agencies [24]. US-based Colt purchased land in Mumbai in 2018 and took six years to bring just 22 MW of a planned 100 MW facility online [25]. Power grid losses run at 16.6%, more than double the global average [24]. AWS reported its worst global PUE at its Hyderabad facility (1.50 versus a best-practice target of 1.1–1.2) [26]. There is no national data center policy; the 2020 draft remains a draft in 2026 [24].
The West is hitting a wall that money cannot fix. In the United States, data center interconnection queues have stretched from under 2 years to over 8 years, opponents have thwarted roughly $100 billion in projects in a single quarter of 2025, and moratorium bills are pending in 5 states [53][54][55]. PJM ratepayers absorbed an extra $9.4 billion in electricity bills during summer 2025 alone [56]. Ireland imposed a de facto moratorium on grid connections for Dublin data centers, stranding €5.8 billion in projects [57]. Amsterdam and Germany have their own restrictions [58]. This is no longer an engineering problem in the West. It is a political crisis.
India has none of these political constraints, not yet. No organized community backlash at scale, though early friction over water and power diversion near data center campuses in Navi Mumbai and Chennai suggests the politics may eventually follow the megawatts [59]. For now, states are actively competing for investment: offering land at 25–50% below market price, waiving the 5–10% stamp duty on property transactions, and promising single-window clearance to replace the 30+ separate approvals that normally slow construction to a crawl [59]. Vacancy rates are compressed to 4.3% [59]. Construction costs of $6.60–7 per watt are among the lowest globally [23]. A Deloitte report published in February 2026 projects India will attract $200 billion in data center investment by 2030, scaling from 1.5 GW to 8–10 GW in four years [60].
The distinction matters. India’s data center barriers are engineering problems that capital can solve: grid upgrades, cooling systems, and clearance streamlining. The US and EU barriers are political and social resistance problems that capital cannot solve, and that are getting worse as ratepayer anger grows. When hyperscalers look at where to deploy the next hundred billion, India is one of the few major markets where construction is still welcomed.
This is precisely where the “building for everyone but themselves” thesis bites hardest. The $200 billion wave is overwhelmingly driven by hyperscalers building for their global customers, not by Indian AI builders. The 20-year tax exemption explicitly incentivizes foreign companies to process worldwide workloads through Indian soil [49]. Sarvam trained its 105B model on IndiaAI Mission GPUs, government-subsidized compute that would not be provided by the market. The infrastructure is being solved. The question is: for whom?
The mirror: How the Gulf built sovereign AI with less
This is where the comparison gets uncomfortable, because a country with a fraction of India’s talent and none of its domestic market built what India has not.
To be fair, no large democracy has matched the Gulf’s pace of sovereign AI development. Japan has spent over $2 billion on national AI initiatives and produced modest results. Germany has no national foundation model. Brazil has nothing. India’s gap is distinctive not because of comparisons with Gulf monarchies, but because of comparisons with its own talent base.
The UAE’s Technology Innovation Institute was founded in May 2020 with a core team of around 25 researchers [29]. It shipped six model generations in under three years: from Falcon 40B topping the open-source leaderboard in 2023 [30], to a 180B model matching Google’s PaLM 2 Large [30], to a multimodal model [31], to efficient small models that beat Meta’s Llama 3.1 and Alibaba’s Qwen 2.5 [32], to a genuine architectural innovation (a hybrid Transformer-state space design whose 34B flagship outperformed models twice its size) [33], to a 7B reasoning model in January 2026 that outperformed NVIDIA’s Nemotron 47B, a model seven times larger [10]. Each generation showed the team had learned from the last. No Indian institution has demonstrated anything close to this iteration speed.
Saudi Arabia followed a parallel track. SDAIA developed the original ALLaM models with IBM in 2023-24 [34]. Then the Kingdom formed HUMAIN, a $100 billion PIF-backed national AI company that consolidated Saudi Arabia’s previously fragmented AI apparatus by absorbing 95 SDAIA employees and combining Aramco Digital and the National Center for AI talent [35]. In August 2025, it launched ALLaM 34B: a foundation model trained from scratch by 40 PhD researchers working in the Kingdom, with a consumer app (HUMAIN Chat) to rival ChatGPT in Arabic [36].
Three advantages explain the Gulf’s speed.
First, concentrated capital with no deliberation overhead. The GCC’s combined sovereign wealth funds exceed $4 trillion [37]. A single ruler can direct billions without parliamentary debate, inter-ministerial coordination, or coalition politics. India’s entire IndiaAI Mission budget is $1.25 billion over five years [5]. The UAE’s MGX fund announced a $30 billion commitment to AI infrastructure in a single announcement [38]. India’s democracy, however, is not the binding constraint its defenders claim: India can do concentrated state-driven technical programs when it chooses to (Chandrayaan-3 cost $75 million [62]). It simply has not chosen to treat AI as one.
Second, a “talent recruitment” model that simply attracts what India organically produces and then loses. TII recruited Professor Mérouane Debbah from CentraleSupélec. MBZUAI recruited Eric Xing from CMU as president and poached faculty from Berkeley and Carnegie Mellon. KAUST attracted Jürgen Schmidhuber [39]. Senior AI scientist salaries in Saudi Arabia average $420,000, tax-free [40]. The UAE has acquired 23.1 million H100 GPU equivalents, compared with India’s 1.2 million, nearly 20 times more compute for a fraction of the population [41].
Third, regulatory simplicity and energy abundance. Saudi Arabia’s industrial electricity costs $0.048 per kWh, 30–50% cheaper than India’s rates [42]. Both the UAE and Saudi Arabia offer free zones with 100% foreign ownership, streamlined approvals, and abundant state-owned land. A data center that takes 2–6 years to build in India can be operational in the Gulf in a fraction of that time.
The GCC model has real vulnerabilities, though they are specific, not the clichés that Western commentators usually reach for. The Gulf’s AI outputs are genuine: the H1’s hybrid architecture was a real research contribution, not a checkbook exercise. What is fragile is the institution. TII’s research culture was built by specific people, Debbah and the core team he recruited, and whether it survives their eventual departure is an open question. The tax-free salaries that attract world-class researchers today can be outbid by Saudi Arabia tomorrow, or undercut by a shift in Emirati priorities next year.
The most existential external risk is one that the Gulf cannot control: US chip export controls classify both countries as Tier 2, and the restrictions are tightening. G42 was forced to divest from all Chinese partnerships to maintain access to NVIDIA GPUs [43]. If Washington further restricts advanced GPU exports (which is actively under discussion), the Gulf’s entire compute advantage evaporates overnight. India, as a Tier 1 strategic ally, has more durable access to the hardware that enables AI development.
The Gulf’s models also face a market problem India does not: the UAE and Saudi Arabia lack a domestic population large enough to test or commercialize AI at scale. Falcon’s benchmarks are measured on English-language evaluations; ALLaM serves an Arabic-speaking internet population roughly one-fifth the size of India’s. The Gulf built models without a market. India has the market without the models.
That critique cuts both ways. India has organic talent and a domestic market, but it has not built on them. The Gulf lacks both, and it has. The question is not which model is more sustainable. It is why India’s natural advantages have not translated into output.
The summit spending spree
The India AI Impact Summit (February 16–20, 2026) produced an avalanche of investment pledges. The headline numbers are staggering:
Amazon announced $35 billion in new investment by 2030, with $16.4 billion specifically for AWS infrastructure [44]. Microsoft committed $17.5 billion over four years, its largest Asian investment, anchored by a hyperscale data center in Hyderabad, set to go live mid-2026 [45]. Google pledged $15 billion over five years, including a gigawatt-scale AI hub in Visakhapatnam with Adani [46]. Reliance Industries announced ₹10 lakh crore (roughly $110 billion) over seven years for AI and data infrastructure [47]. OpenAI and Tata are building HyperVault, starting at 100 MW and aiming for 1 GW [48]. L&T and NVIDIA announced a gigawatt-scale sovereign AI factory [48].
Investors should apply heavy discounting. A commitment contingent on regulatory conditions being met over seven years is a letter of intent, not a capital commitment, and should be valued accordingly. The government’s own policy moves are real (a 20-year tax exemption, data localization rules that will force companies to store Indian user data on Indian soil [63], GPU subsidies scaling to 38,000 units [5]).
Reliance has a pattern. Its “Made in India” 5G stack was announced with similar fanfare at the 2022 AGM, but the network launched on Nokia and Ericsson equipment, and 5G utilization sat at just 15% two years later [65]. The $110 billion figure, spread over seven years with no published breakdown by segment, should be evaluated against that track record. Much of the infrastructure being built will serve global customers: Amazon, Microsoft, and Google are building capacity for their worldwide cloud businesses, not specifically for Indian AI development. India’s own public AI funding remains dwarfed by foreign private investment, raising questions about who will own the resulting capability.
What actually changes the equation
The honest assessment is that India is accelerating from a very low base, with structural headwinds that remain considerable. The tax exemption, the GPU subsidies, and Sarvam’s proof of concept are real. The strongest bull case, though, is the DeepSeek effect, and it deserves direct engagement. If DeepSeek proved you can build frontier-competitive models for a reported $5.6 million in compute, rather than the $100 million+ previously assumed, then India’s capital gap matters dramatically less than it did a year ago. The barrier drops from “you need a billion-dollar war chest” to “you need a few million and brilliant researchers.” India has the researchers, if it can retain them. The DeepSeek precedent has energized Indian founders; as one industry leader observed, “DeepSeek is probably the best thing that happened to India. It gave us a kick in the backside” [50].
The counter matters too. DeepSeek’s $5.6 million figure covers only cloud rental hours; SemiAnalysis estimates the actual infrastructure investment at roughly $1.3 billion across approximately 50,000 Hopper GPUs [66]. More importantly, DeepSeek’s team included senior researchers with years of frontier experience at top Chinese labs. India does not have that level of domestic bench depth because of brain drain. Even the DeepSeek path, which is the most favorable scenario for resource-constrained builders, requires solving the talent retention problem first. Efficiency breakthroughs lower the capital barrier. They do not eliminate the human one.
The underlying realities documented above remain stubborn regardless. The capital gap ($643M vs $109.1B) cannot be closed with tax breaks or government programs. It requires a fundamental shift in the risk appetite of Indian capital markets for deep tech [7]. The brain drain continues; the Trump administration’s $100,000 H-1B supplemental fee, enacted in September 2025 and now under legal challenge, could paradoxically help by making it more cost-effective to invest in Indian talent domestically than to sponsor them in the US, but this is speculative [51]. R&D spending remains stuck at levels below every major Asian peer, with no announced policy to change it [22].
The verdict, and the problem only India can solve
The most likely outcome is that India becomes the world’s most important AI deployment market while remaining a minor player in model development over the next five to seven years. The structural forces are that strong. The services equilibrium is self-reinforcing: it shapes the career graph, the capital allocation, the institutional priorities, and the talent pipeline simultaneously. No single policy, not the tax exemption, not the GPU subsidies, not the summit pledges, changes that equilibrium.
Sarvam is an exception that required government compute, academic founders, and a linguistic problem that nobody else would solve. Replicating those conditions at scale requires changes to the incentive structure that no current policy achieves. A concrete test: if India has not produced a second foundation model at the 100B+ parameter class by 2028, the services equilibrium will have proven self-reinforcing.
India’s 22 official languages and roughly 1,600 dialects mean an AI product needs to support 5–7 languages just to reach 80% of the population. Indian languages collectively represent less than 1% of internet content, creating severe data scarcity for model training [27]. Digital literacy hovers around 37%, and 630 million people remain offline [28].
Aadhaar covers 1.4 billion people. UPI processes 20 billion monthly transactions [52]. The application of AI to TB diagnosis across 100,000 clinics, to agricultural yield prediction for 150 million smallholder farmers, to governance at the scale of India’s complexity: this represents a problem that no one else in the world will solve. OpenAI is not building for Marathi-speaking farmers in Vidarbha. Google is not optimizing TB screening for district hospitals in Jharkhand. Only homegrown builders will.
This is important work. No other country has the combination of population, digital rails, and linguistic diversity to prove that AI works for a billion people who do not speak English. It is also a different ambition than building frontier models, and it means the country that produces more AI talent than any country except the United States and China will continue to be the world’s most productive talent exporter, training its best minds, watching them build the future somewhere else, and importing the results.
The country that landed on the Moon’s south pole for less than the budget of a Hollywood film [62] has not decided AI deserves the same focus. That is a choice, not a constraint, which means it is reversible.
Sarvam’s Vivek Raghavan, who built Aadhaar before co-founding a company to build AI for every Indian language, put it most directly: “If you are not ambitious as a country, then we will just do small things” [64]. His 105B model, shipped on February 18 on subsidized compute, is the first crack in the old equilibrium. Whether it widens into a real shift, or whether Sarvam remains a lonely exception, depends on whether India decides that building for itself is worth the structural cost of changing the incentives that have made building for everyone else so profitable.
Notes
[1] Outlook Business, December 2025. Aggarwal’s personal investment figure from company disclosures and Indian media reporting.
[2] Outlook Business, “Krutrim in Crisis,” December 2025. Senior departures, layoffs, and Aggarwal’s benchmark acknowledgment.
[3] ChatGPT India weekly active users from Similarweb and industry estimates, Q4 2025. Kruti app downloads from Google Play Store.
[4] Tracxn, Sarvam AI company profile, 2026. Total funding of $53.8M across seed and Series A ($41M) rounds.
[5] India AI Impact Summit announcements, February 2026. IndiaAI Mission GPU deployment figures and subsidized access rates from Ministry of Electronics and IT.
[6] Sarvam AI launch presentation, February 18, 2026. TechCrunch, “Indian AI lab Sarvam’s new models are a major bet on the viability of open source AI”; Business Standard, “Why Sarvam’s new 105B model marks a shift in India’s sovereign AI ambitions”.
[7] Stanford Institute for Human-Centered AI, “AI Index Report 2025”. India data on research output (9.2%), skill penetration (#2), startup formation (74 vs 1,073 US), private investment ($643M vs $109.1B US, Chapter 4 investment figures), and talent flow indices.
[8] Market capitalizations as of February 2026. Indian-origin CEO positions from Forbes, Bloomberg, and company disclosures.
[9] UNESCO and Indian Ministry of Education STEM graduate data, 2024-25 academic year.
[10] TII press release and FinancialContent, “Falcon-H1R 7B: The Hybrid Model Outperforming Behemoths 7x Its Size,” January 13, 2026. AIME 2025 score of 83.1%, NVIDIA Nemotron 47B comparison, LiveCodeBench v6 score of 68.6%. See also TII official announcement and arXiv:2601.02346.
[11] NASSCOM and EY estimates of India’s AI workforce, 2025. Range reflects different definitional scopes (narrow AI research vs broader AI-adjacent roles).
[12] NBER Working Paper, “Top Talent, Elite Colleges, and Migration: Evidence from the Indian Institutes of Technology”. 62% emigration rate among top 100, 90% among top 10, ISRO/DRDO employment rates.
[13] Glassdoor, LinkedIn Salary Insights, and TeamLease Digital compensation surveys for AI/ML roles in India versus the United States, 2025.
[14] USCIS H-1B Employer Data Hub, Fiscal Year 2024. Indian nationals: 283,397 of 398,708 total approvals (71.1%).
[15] Carnegie Endowment for International Peace, “The Missing Pieces in India’s AI Puzzle: Talent, Data, and R&D,” February 2025. Tier analysis of talent distribution, PhD pursuit rates (0.08% India vs 4.2% China).
[16] Industry revenue of $283 billion (including hardware) from NASSCOM Strategic Review 2025, FY2025 estimate. Combined revenue of TCS ($30.2B), Infosys ($19.3B), HCLTech ($13.8B), and Wipro ($10.5B) from FY2025 annual reports.
[17] Analytics India Magazine, “Why TCS, Wipro & The Ilk Fancy Agentic AI Over Foundational Models,” 2025. TCS four LLMs at 2.5B parameters each, all Llama fine-tunes.
[18] TCS-OpenAI and Infosys-Anthropic partnerships announced at AI Impact Summit, February 2026. Reported in Financial Times, “India’s AI ambitions hit limits at global summit,” February 22, 2026. 460 generative AI projects from Infosys Q3 FY2026 investor presentation.
[19] Source Asia, “Cognizant, Infosys, TCS and Wipro emerge as Frontier Firms with Microsoft,” 2025. Copilot deployment figures exceeding 50,000 licenses per company.
[20] Nandan Nilekani remarks at Meta AI summit, 2025. Widely reported across Indian business press.
[21] Mohandas Pai, Aarin Capital, quoted at Indian venture capital industry event, 2025.
[22] UNESCO Institute for Statistics; India Department of Science and Technology R&D expenditure data, historical series. India at 0.65% vs South Korea 4.9% vs Japan 3.3% vs China 2.4% of GDP. OECD Science, Technology and Innovation Outlook for peer comparison data.
[23] CNBC, “India’s data center gold rush,” 2025. CBRE and JLL India data center market reports. US capacity from Cushman & Wakefield global data center tracker.
[24] Observer Research Foundation, “Building Sovereign Data Centre Infrastructure in India,” 2025. 30+ clearances, power grid losses at 16.6%, regulatory fragmentation analysis.
[25] Colt Data Centre Services India, Mumbai facility timeline. Land acquisition 2018, 22 MW operational 2024 per industry reporting.
[26] AWS sustainability disclosures; PUE data compiled by Greenpeace India and independent data center analysts for Hyderabad region. 40°C+ temperature data from India Meteorological Department.
[27] AI4Bharat and Microsoft Research India, Indian language internet content studies. Less than 1% of global web content in Indian languages despite 1.4B population.
[28] IAMAI-Kantar “Internet in India” Report 2025. Digital literacy at ~37% per UNESCO estimates. 630M offline from TRAI subscriber data.
[29] TII founding details from ATRC official communications; team of 25 from Dr. Hakim Hacid, Chief Researcher, in media interviews (Capacity, AI for Good Summit).
[30] Hugging Face Open LLM Leaderboard historical rankings, March and September 2023. PaLM 2 comparison from TII benchmarking data.
[31] TII, Falcon 2 11B release announcement, May 2024. Vision-to-language capabilities. Hugging Face independent verification of Llama 3 8B comparison.
[32] TII press release, “Falcon 3: World’s Most Powerful Small AI Models,” December 17, 2024. VentureBeat, The Decoder, and MarkTechPost coverage. 14T tokens, sub-13B leaderboard position verified by Hugging Face.
[33] TII, “Falcon-H1: A Family of Hybrid-Head Language Models,” May 20-21, 2025. Technical blog post and arXiv:2507.22448. 0.5B-34B parameter range, 256K context, 18 languages. BusinessWire launch announcement. AWS Bedrock Marketplace availability confirmed September 2025.
[34] IBM-SDAIA joint announcement at Think Boston 2024, May 2024. ALLaM 13B on watsonx. Subsequently available on Microsoft Azure, September 2024.
[35] Arab News, Bloomberg, and Middle East AI News reporting on HUMAIN formation under PIF, SDAIA employee transfer (95 staff), and organizational merger with Aramco Digital and National Center for AI, 2025.
[36] HUMAIN CEO Tareq Amin interviews with Asharq Al-Awsat and Arab News, August 2025. ALLaM 34B: 40 PhD researchers, trained from scratch on in-Kingdom proprietary data, HUMAIN Chat consumer application. AI World Journal, “$100 Billion HUMAIN AI Company.”
[37] Sovereign Wealth Fund Institute global rankings, Q4 2025. ADIA ~$1T, PIF ~$925B, Mubadala ~$302B, plus QIA, KIA, and others.
[38] MGX fund: Mubadala, G42, BlackRock, and Microsoft partnership announced 2024. $30B AI infrastructure deployment.
[39] TII recruitment of Prof. Mérouane Debbah (CentraleSupélec/Huawei); MBZUAI faculty from UC Berkeley and CMU; KAUST recruitment of Jürgen Schmidhuber. Per institutional announcements and press coverage.
[40] Hays Middle East and Robert Half GCC salary surveys for senior AI researchers, 2025. Tax-free status per UAE Federal Decree-Law No. 47/2022 (corporate) and absence of personal income tax.
[41] Introl, “India’s GPU Infrastructure Landscape: A Comprehensive Survey,” 2025. UAE compute capacity from ATRC disclosures and industry tracking (23.1M H100 equivalents vs India’s ~1.2M).
[42] Saudi Electricity Company published industrial tariff schedules. India comparison from Central Electricity Authority industrial rate data.
[43] Reuters, Financial Times, and Bloomberg reporting on G42 divestiture of Chinese partnerships (including ties to Huawei) under US government pressure, 2024. Tier 2 classification per US Commerce Department chip export control framework.
[44] Amazon India investment announcement, AI Impact Summit, February 2026. $35B new + $40B prior. $16.4B AWS infrastructure breakdown per company filing.
[45] Microsoft India announcement, AI Impact Summit, February 2026. $17.5B over four years. Hyderabad hyperscale data center mid-2026 timeline from Satya Nadella keynote.
[46] Google-Adani partnership, AI Impact Summit, February 2026. $15B over five years. Visakhapatnam gigawatt-scale AI hub.
[47] Reliance Industries, Mukesh Ambani keynote, AI Impact Summit, February 2026. ₹10 lakh crore (~$110B) over seven years.
[48] Upstox News, “From TCS, Infosys to Qualcomm, NVIDIA: Who announced what amid India AI Impact Summit,” February 2026. OpenAI-Tata HyperVault (100 MW → 1 GW) and L&T-NVIDIA sovereign AI factory details.
[49] CNBC, “India tax breaks for hyperscalers as part of AI bet,” February 2, 2026. Union Budget 2026-27: 20-year tax exemption through 2047, elimination of 35% permanent establishment tax.
[50] Indian AI industry leader quoted at AI Impact Summit sideline event, February 2026, on DeepSeek’s catalytic effect on Indian AI ambitions.
[51] Presidential Proclamation, “Restriction on Entry of Certain Nonimmigrant Workers,” signed September 19, 2025, effective September 21, 2025. $100,000 supplemental fee on new H-1B petitions. Under legal challenge: Chamber of Commerce v. DHS (D.D.C.) and Global Nurse Force v. Trump (N.D. Cal.).
[52] NPCI monthly UPI transaction data, January 2026 (~20B transactions/month). UIDAI Aadhaar enrollment statistics (1.4B+ registered).
[53] RMI, “PJM’s Speed to Power Problem and How to Fix It,” November 2025. Interconnection queue timeline from <2 years (2008) to 8+ years (2025). Virginia 7-year delays and 300+ data centers from Acres Land Values and industry reporting. PJM serves 67 million people across 13 states per PJM Interconnection official data.
[54] The American Prospect, “Demands for Data Center Moratoriums Surge,” December 2025. Data Center Watch/10a Labs reporting: Q2 2025 opponents thwarted ~$100B in projects, 20 projects fell through in one quarter. Includes Chandler, Arizona rejection.
[55] Scioto Post, “Moratoriums on AI Data Centers Spark Legal and Political Battles Nationwide,” February 2026. 19+ Michigan municipalities with moratoriums. Virginia, Georgia, Maryland, Minnesota, New York moratorium bills from Built In, “States Push Data Center Moratoriums as AI Growth Surges,” 2026. 230-member environmental coalition letter to Congress from Food & Water Watch, December 2025 (NPR, ENR reporting).
[56] NRDC, “Building Data Centers Without Breaking PJM,” October 2025, and Utility Dive, “Solving PJM’s data center problem,” December 2025. $9.4B extra electricity bills summer 2025. $100B+ projected additional costs through 2033. Capacity prices jumped 1,000%+ in recent auctions.
[57] Irish CRU Large Energy Users Connection Policy final decision, December 15, 2025. De facto moratorium from November 2021 direction (CRU/21/124). €5.8B stranded projects from Enlit World, “Data centre assets left stranded by power constraints.” Data centers at 22% of Ireland’s electricity 2024 from CRU data. New rules: dispatchable generation matching import capacity, 80% renewable sourcing. Analysis from Pinsent Masons, William Fry, Philip Lee LLP, and Arthur Cox.
[58] Amsterdam moratorium from Beyond Fossil Fuels/IEA reporting, December 2024. Germany PUE 1.3 by 2030 and 100% renewable by 2027 from German Energy Efficiency Act (Energieeffizienzgesetz). European Commission new measures from DatacenterDynamics, “European Commission to launch new measures to stem data center energy consumption,” June 2025.
[59] JLL India Data Centre Market Dynamics Report H1 2025: vacancy at 4.3%, 97.9 MW net take-up (+48% YoY). State incentives (land subsidies 25-50%, stamp duty exemptions, single-window clearances) from MarkNtel Advisors India Data Center Market Report 2025 and individual state data center policies (Maharashtra, Karnataka, Telangana, Tamil Nadu, UP, Rajasthan, Haryana).
[60] Deloitte Touche Tohmatsu report, February 19, 2026. $200B India data center investment by 2030 (part of $800B APAC total). 8-10 GW capacity projection. Reported by Business Standard, “India may attract $200 bn in data centre investment by 2030.”
[61] India has produced important Indic-language NLP work — AI4Bharat‘s IndicTrans and IndicBERT translation and language models (IIT Madras), Wadhwani AI‘s TB screening deployments across 23 states, Jugalbandi’s multilingual chatbot for government services — but none at foundation-model scale or with commercial breakout. Sarvam’s 105B is the first trained-from-scratch model at a scale that invites frontier comparison.
[62] ISRO Chandrayaan-3 mission cost of approximately ₹615 crore (~$75 million) per ISRO chairman S. Somanath, August 2023. Comparison: the film Gravity (2013) had a production budget of approximately $100 million. Chandrayaan-3 achieved the first-ever landing on the Moon’s south pole.
[63] Digital Personal Data Protection Act, 2023 (India). Enacted August 2023, rules still being finalized as of February 2026. Data localization provisions require certain categories of personal data to be stored and processed within India. Implementation timeline and enforcement mechanisms remain under development per Ministry of Electronics and IT consultations.
[64] Vivek Raghavan interview with Outlook Business, “If We’re Not Ambitious, We’ll Only Build Small Things,” February 18, 2026. Also: Business Standard, “How two engineers built Sarvam AI from an idea to a summit showcase,” February 19, 2026. Raghavan’s 12-year Aadhaar volunteer work from BusinessToday and Storyboard18 founder profiles, February 2026.
[65] Reliance Jio “Made in India” 5G stack announced at RIL 45th AGM, August 2022. Network launched October 2022 on Nokia and Ericsson equipment per Light Reading and Telecoms.com reporting. 5G utilization at ~15% per equipment supplier sources, RCR Wireless, October 2024. Jio took “a more measured approach to 5G expansion due to low capacity utilization and delayed monetization.”
[66] SemiAnalysis, “DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts,” January 2025. Estimated total server CapEx of ~$1.3 billion. GPU inventory of approximately 50,000 Hopper GPUs (mix of H800, H100, and H20). DeepSeek’s reported $5.6M covered only cloud rental hours for the final V3 training run, not infrastructure, prior experiments, or personnel costs.
[67] Vaswani, Shazeer, Parmar, et al., “Attention Is All You Need,” NeurIPS 2017. Parmar (née Niki Parmar, Pune) co-founded Adept AI (2022) and Essential AI (2023) before joining Anthropic in late 2024. Per company announcements and Parmar’s public disclosures on X, February 2025.