In May 2024, India’s Union Cabinet approved the IndiaAI Mission with a budget of Rs 10,372 crore – the country’s most ambitious attempt to build sovereign artificial intelligence infrastructure. At its core was a target that quickly became a headline: 10,000 Graphics Processing Units (GPUs) made available to researchers, startups, and public institutions at affordable rates. The ambition is real. But so is the comparison that follows it everywhere: China is reported to have more than 100,000 advanced GPUs in state-backed AI compute clusters, with some estimates placing the number far higher once private hyperscalers are counted.
Does that gap matter? Is India even playing the same game? And what does it mean for a country of 1.4 billion people to build AI infrastructure that serves its own needs rather than simply replicating what worked in Silicon Valley or Shenzhen? This article takes stock of where India stands, what the IndiaAI Mission is actually building, where the honest gaps lie, and why the path forward for India – and the Global South more broadly – may not require closing the GPU count gap at all.
The GPU Race: What the Numbers Actually Mean
GPU count is a proxy metric, not a complete picture. A GPU – Graphics Processing Unit – was originally designed to render video game graphics. Its ability to perform thousands of mathematical operations simultaneously made it the hardware of choice for training large neural networks, the foundation of modern AI. NVIDIA’s H100 and A100 chips have become the oil barrels of the AI era: nations and corporations measure their AI potential partly by how many they can access.
China’s reported figure of 100,000+ GPUs in national compute clusters refers primarily to state-backed infrastructure built through entities like the National Computing Center network and large private companies such as Baidu, Alibaba, Tencent, and Huawei. The Chinese government designated AI as a core strategic industry in its Made in China 2025 plan and has since invested hundreds of billions of yuan across hardware, software, and talent. US export controls on NVIDIA chips introduced in 2022 and tightened in 2023 have complicated China’s access to the most advanced chips, pushing it to develop domestic alternatives like Huawei’s Ascend 910B series.
India’s 10,000 GPU figure is the public procurement target announced under IndiaAI Mission Phase 1. These GPUs are intended for a shared compute infrastructure accessible to startups, academic institutions, and government research bodies through a pay-per-use marketplace. The chips being procured include NVIDIA H100s and comparable alternatives. When fully deployed, this represents a meaningful foundation – but the gap with China’s reported capacity is not a rounding error. This mirrors a broader pattern of strategic technology bets India is making, similar to the Rs 6,000-crore Quantum Mission where India is building capabilities in parallel rather than in sequence.
| Metric | India (IndiaAI Mission) | China (Est.) |
|---|---|---|
| Public compute GPUs (announced) | ~10,000 | 100,000+ |
| AI investment (national, 2023-24) | ~$1.25B (Rs 10,372 cr) | ~$15B+ (state-backed) |
| AI startup ecosystem | Top 5 globally by volume | Top 2 globally |
| Language model focus | 22 scheduled languages | Mandarin + multilingual |
| Chip export restrictions | Not targeted | US-restricted since 2022 |
The honest answer is that India and China are not building the same thing for the same purpose. China’s AI infrastructure is designed to power large-scale general-purpose foundation models, military and surveillance applications, and global product competition. India’s stated mission is different: build affordable compute for 22 scheduled languages, enable startups that serve Indian users, and keep sensitive government data from flowing to foreign cloud providers.
What IndiaAI Mission Is Actually Building
The IndiaAI Mission, approved in March 2024 and operationalised through MeitY (Ministry of Electronics and Information Technology), has seven distinct pillars. Understanding them separately matters because the GPU headline captures only one.
- IndiaAI Compute Capacity: The 10,000 GPU public compute infrastructure, procured through a transparent bidding process and made available as a shared national resource.
- IndiaAI Innovation Centre: A centre for developing indigenous large multimodal models (LMMs) – AI systems that can process text, voice, and images in Indian languages.
- IndiaAI Datasets Platform: A curated, consent-based repository of Indian data – agricultural, healthcare, judicial, urban – to train AI models that actually reflect Indian realities.
- IndiaAI Application Development Initiative: Funding and sandboxes for startups building AI applications in priority sectors: agriculture, health, education, and governance.
- IndiaAI FutureSkills: A curriculum initiative to train 1 million+ AI professionals by 2025, anchored to IITs, NITs, and state technical universities.
- IndiaAI Startup Financing: Grant and equity support for deep-tech AI startups, particularly those working in regional languages or underserved domains.
- Safe and Trusted AI: Governance frameworks, audit mechanisms, and safety guidelines to ensure that AI deployed in India does not amplify bias or harm.
What makes this framework distinctive is the Datasets Platform. Most Western and Chinese AI systems are trained primarily on English or Mandarin internet data. Indian-language internet data is sparse – there are more speakers of Bengali, Tamil, Telugu, and Marathi than of French, but the training data available for these languages is a fraction of what exists for French. The IndiaAI Datasets Platform intends to change this by systematically digitising and licensing Indian-language text, court records in regional languages, agricultural advisories, and healthcare records (de-identified).
The Estonia Model: Small Nation, Big Digital Ambition
Estonia offers India the clearest proof that size of compute investment is not the determinant of AI outcomes. Estonia – a nation of 1.3 million people – built the most digitally advanced public administration in the world not by matching larger nations on hardware spend, but by making deliberate choices about what its digital infrastructure needed to do. Estonia’s X-Road data exchange layer, e-Residency program, and AI-assisted public services run on a fraction of the compute that the US or China deploys. Yet Estonian citizens can access 99% of government services digitally, and the country’s per-capita productivity in digital services rivals any large economy.
The Estonian lesson for India is this: sovereign digital infrastructure built with a clear mission – serve citizens, protect data, reduce friction – outperforms infrastructure built to match a competitor’s headline numbers. India’s IndiaAI Mission, with its emphasis on shared compute for specific Indian use cases rather than frontier model training, is making the same strategic bet Estonia made two decades earlier. The difference in scale between 1.3 million Estonians and 1.4 billion Indians makes India’s ambition more complex – but the logic holds.
Bhashini and the Language Sovereignty Argument
India’s most distinctive contribution to global AI infrastructure is not GPU count. It is Bhashini, the National Language Translation Mission. Launched under MeitY in 2022, Bhashini is an open-source AI platform that provides translation, transcription, and voice interfaces for 22 scheduled languages. By mid-2025, Bhashini had logged over 600 million API calls – used by government portals, private apps, and civic technology tools across India.
The Bhashini model family is publicly available, free to integrate, and continuously improving as more data is contributed through crowd-sourced campaigns. Its voice-to-voice translation capability – where a farmer in Tamil Nadu can speak in Tamil and have the query answered in Hindi or English – is not replicated by any commercially available system at this scale and at zero marginal cost to the user.
This is the language sovereignty argument: India does not need to build a ChatGPT. India needs to build AI infrastructure that makes a Marathi-speaking nurse in Nagpur as capable as an English-speaking software engineer in Bengaluru. The GPU count required for that mission is far smaller than the GPU count required to train a GPT-4 scale model. And that mission – democratising AI access across language lines – is something neither the US nor China has solved.
Krutrim and the Foundation Model Question
In December 2023, Ola founder Bhavish Aggarwal launched Krutrim – India’s first large language model built from the ground up, trained on Indian-language data. Krutrim means “artificial” in Sanskrit. The model, and the subsequent Krutrim Cloud infrastructure, represented India’s first serious private-sector attempt to build a full-stack AI company: chips, cloud, and foundation models.
Krutrim’s approach is instructive. Rather than trying to match GPT-4 on English benchmarks, Krutrim focused on multilingual performance across 22 Indian languages and optimised for Indian knowledge domains – UPSC-style reasoning, Indian legal frameworks, regional agricultural advice. It demonstrated that a smaller, purpose-built Indian model could outperform larger general models on Indian-specific tasks.
This mirrors a broader global trend: the era of “one foundation model to rule them all” is giving way to an era of specialised, fine-tuned models deployed at the edge. For India, this is good news. A 70-billion-parameter model fine-tuned on Indian judicial data is more useful to a district court in Rajasthan than a 1-trillion-parameter model trained on English Wikipedia. The compute required to fine-tune is dramatically lower than the compute required to train from scratch.
Data Centres, MeitY, and Sovereign AI
India currently has approximately 900 megawatts of operational data centre capacity, with another 1,500 MW under construction – a total projected to reach 2,400+ MW by 2027 according to JLL India’s data centre report. The majority of this capacity is concentrated in Mumbai, Chennai, Hyderabad, Delhi-NCR, and Bengaluru. These facilities host a mix of cloud provider infrastructure (AWS, Azure, GCP all have India regions), colocation customers, and enterprise private clouds.
The IndiaAI compute infrastructure is being built on top of this existing base, with the government procuring capacity from both private data centre operators and in-house government facilities. MeitY has identified STPI (Software Technology Parks of India) centres as potential colocation nodes for the shared GPU cluster, allowing startups outside major metros to access compute without relocating.
Sovereign AI – the principle that a nation’s AI training data, model weights, and inference infrastructure should not wholly depend on foreign corporations – is the policy logic behind this investment. India has seen what data dependency can look like: the dependence on US cloud providers for government services creates a vulnerability that both MeitY and NITI Aayog have documented. The IndiaAI Mission’s compute infrastructure is explicitly designed as an alternative to cloud compute for sensitive government AI workloads. This same logic of building domestic capacity applies across sectors – just as India’s healthcare infrastructure gap shows the long-term cost of under-investment, over-dependence on foreign AI infrastructure carries strategic risk.
Where India Actually Stands: An Honest Assessment
No accounting of India’s AI infrastructure ambitions is complete without an honest look at the gaps. Three structural challenges stand out.
The Procurement Pace Gap
The IndiaAI Mission announced 10,000 GPUs in March 2024. Actual GPU procurement has moved more slowly than the announcement suggested. By early 2026, only a portion of the announced capacity had been deployed and made accessible to users. Government procurement processes – tender design, L1 bidding, delivery, installation, and hypervisor setup – take time that the private sector does not face. India’s startup community has consistently flagged that wait times for GPU allocation through government portals remain longer than simply purchasing cloud compute from AWS or Azure. This is solvable but requires deliberate process reform.
The Talent Concentration Gap
India graduates more engineering students each year than any country except China – approximately 1.5 million engineers annually. But AI-specific talent – researchers capable of training foundation models, hardware engineers who understand GPU cluster architecture, and ML infrastructure engineers – remains concentrated in a small number of IITs and a handful of private companies. The pipeline feeding into AI research in India is growing but not yet at the scale that China or the US can draw from. The IndiaAI FutureSkills initiative is the right response, but curriculum changes take three to five years to show up in the workforce.
The Power and Cooling Infrastructure Gap
GPU clusters are power-hungry. An NVIDIA H100 chip draws approximately 700 watts under load. A 10,000-GPU cluster requires roughly 70 megawatts of clean, reliable power – the equivalent of powering a small city. India’s power grid, while dramatically improved over the last decade through programs like DDUGJY and Saubhagya, still experiences reliability challenges in second- and third-tier cities. Building AI data centres in locations beyond the big five metro areas requires investment in dedicated power feeds, uninterruptible power supplies, and renewable energy procurement agreements. This is a solvable infrastructure problem, but it requires coordination across MeitY, the Ministry of Power, and state grid operators.
Where the Global South Can Actually Win
The question posed in this article’s title – can 10,000 GPUs match China’s 100,000? – is ultimately the wrong question. A better question: what does India need AI infrastructure to do, and is 10,000 GPUs sufficient to do it?
India’s AI use cases are not the same as China’s or America’s. India needs AI that can diagnose tuberculosis from a chest X-ray taken on a ten-year-old smartphone in a Primary Health Centre in Bihar. India needs AI that can translate a government notice from Hindi to Santali and read it aloud over a low-bandwidth audio connection. India needs AI that can advise a Vidarbha cotton farmer on pest management in Marathi at 6am when no agronomist is awake. These are not problems that require training a GPT-5 equivalent. They require carefully curated datasets, domain-specific fine-tuning, and edge deployment – tasks that are well within the reach of 10,000 shared GPUs used efficiently.
This is the Global South’s AI advantage: the problems to be solved are different in character, and the solutions do not require frontier compute at frontier scale. A well-designed compute commons – shared, affordable, accessible to researchers in Tier 2 and Tier 3 cities – can produce AI applications with genuine social impact at a fraction of the cost that Western foundation model companies spend on a single training run.
India is not alone in this insight. The African Union’s AI strategy, Indonesia’s Kominfo-led AI roadmap, and Brazil’s MCTI AI plan all articulate the same logic: sovereign compute for sovereign problems, with international cooperation on safety and standards. India, as the world’s most populous democracy and a G20 presidency that made Digital Public Infrastructure a global agenda item, is well-positioned to lead this coalition.
What Citizens Can Do: Five Layers of Action
AI infrastructure is not a passive spectator sport. Every Indian can engage with it – and doing so makes the infrastructure more useful and more accountable.
Personal
Use Bhashini-powered tools in your daily life – several government apps including DigiLocker, Aarogya Setu, and the UMANG superapp have integrated Bhashini for regional language support. Using these features generates data that improves the model for everyone. If you speak a regional language, consider contributing voice data through the Bhashini Data Collection app – 15 minutes of your voice can train a model that helps millions.
RWA / Neighbourhood
Resident Welfare Associations can advocate with their municipal corporation to use AI-powered grievance management tools – several Indian startups (iGot Karmayogi tools, mySociety-inspired platforms) offer these at low cost to civic bodies. Pushing your local ward office to use these tools creates demand signals that accelerate deployment.
Ward / City
City governments that have adopted Smart City Mission infrastructure often have unused data pipelines. Citizens and civil society organisations can file RTI requests asking what data the Smart City SPV collects, how it is used, and whether it is shared with private companies. Pushing for transparent data governance at the city level builds the consent infrastructure that the IndiaAI Datasets Platform depends on.
State / National
Write to your elected representative or state MLA about the IndiaAI Mission’s progress. Ask specifically: how many GPUs are live and accessible? What is the average wait time for startup GPU allocation? How many researchers outside the IITs have accessed compute through the program? These are specific, answerable questions that legislators can raise in the IT standing committee. Constituent pressure on AI policy is rare in India – and therefore disproportionately effective when it happens.
Professional / Institutional
If you work in education, healthcare, agriculture, or the judiciary – sectors with the highest potential AI impact in India – look for the IndiaAI Application Development Initiative grants. Apply, or connect your institution with startups working in your domain. The supply of compute and funding exists; the bottleneck is domain experts willing to collaborate with technologists on problem definition.
Conclusion: The Right Race to Run
India’s 10,000 GPUs cannot match China’s 100,000 on raw compute. That is a fact worth stating plainly. But the more important question is whether India is running the right race – and the evidence suggests it is. The IndiaAI Mission’s emphasis on language diversity, shared public compute, open datasets, and domain-specific applications reflects a clear-eyed understanding of India’s actual AI needs.
The risks are real: procurement delays, talent gaps, and power infrastructure constraints could all slow the mission’s impact. But none of these are structural barriers – they are execution challenges that India has overcome before, in domains ranging from space technology to digital payments.
India did not need to build Boeing to have one of the world’s best satellite programmes. It built ISRO with discipline, frugality, and a clear definition of what problems it needed to solve. The IndiaAI Mission has the same opportunity. Estonia built world-class digital governance for 1.3 million citizens without matching any superpower’s compute budget. India can build world-class AI infrastructure for 1.4 billion citizens on the same principle. The GPU count is not the measure of success. The measure of success is a nurse in Nagpur, a farmer in Vidarbha, and a student in a rural Bihar school, each accessing AI tools in their own language that make their lives meaningfully better. That outcome is within reach.