Ok so @NandanNilekani and @rvenk said something and everyone is up in arms. Fair. But I see no coherent attempt to lay out what we need to do - just "hey we should not give up." Which I don't think was my takeaway from their article but ymmv. We can agree that "do nothing " isn't an option (and I don't think the article said that either) - but what does "do something" mean here?
TL;DR
A frontier-model push needs capital and talent commitments at a scale that sits beyond any plausible single-year share of Indian GDP - the gap to the US and China isn't proportional, it's structural.
We are missing every leg of the stack that makes frontier labs work: senior researcher cluster, patient capital, indigenous silicon, sovereign compute coordination. None of these has a credible <10-year plan today.
The application and middle layer is where India can credibly win in the next ten years, and is the path that spins up the capital, compute, data and talent flywheel needed to take a real shot at the frontier later.
When the Citrini report dropped a couple of months ago, @sidin and I started kicking around an analysis of what it would take for India to overcome the bleak predictions made in it - "what's the most ideal set of changes India can make to achieve the most ideal outcome in the context of the report" - and I've just pulled this quick and dirty list out of that body of work.
The core problem statement: To ship a frontier model in under ten years, India would have to invest in compute and talent at a scale that sits entirely beyond the envelope of any single year's plausible allocation of national GDP.
Sovereign, private and diaspora capital would have to align in lockstep - the way France did at the Macron summit and China does through state-directed funds.
The bullets below are the core discussion points that imo are missing in the discussion happening on tpot these last couple of days.
Capital - we're spending ~1.6% of what the ecosystem actually needs
IndiaAI Mission outlay: $1.25B over 5 years. (PIB)
Single-year US hyperscaler AI capex 2026: ~$528B. (BofA via Yahoo Finance)
China's 2026 AI investment: $125B, of which ~$48.5B is govt money alone. (Stanford-cited / What's the Big Data, SwarmSignal)
Anthropic's cloud-infrastructure cost alone through 2029: ~$80B. (Sacra)
Mistral (France, $3T GDP) raised $3.4B equity + $830M debt in under three years to attempt the frontier, and is still behind. (CBInsights, CRN on debt round)
Sarvam, our flagship, has raised ~$391M including the pending round. (ET)
Compute - a single private US lab has 26× our entire national GPU pool
IndiaAI national pool (shared across academia, startups, govt): 38,000 GPUs. (PIB)
xAI Colossus, single private lab: 1M+ H100-equivalents (Jan 2026). (SiliconANGLE, CNBC, Business Insider)
OpenAI: target 1M+ GPUs by end-2025; $1.15T infrastructure commitments through 2035. (Tom Tunguz, SiliconANGLE)
Stanford AI Index, H100-equivalents: US 39.7M, China 400,000; India is below the reporting threshold. (AI Index 2026 PDF, comparative analysis)
DeepSeek's "cheap" model required $1.6B in pre-positioned compute capex at parent High-Flyer, a $10B AUM quant fund that had been buying A100s since 2020. (SemiAnalysis-cited / Groundy, Wikipedia on High-Flyer, CO/AI)
Talent - the diaspora is not coming back
Stanford AI Index 2026: India has 50,460 top AI authors (#2 globally). (Indian Express on Stanford)
India also has the largest net outflow (-16.9) of AI research talent in 2025. (Indian Express on Stanford)
10,000+ top Indian AI professionals moved to the US between 2019 and 2024. (Mathrubhumi on Zeki "State of AI Talent 2025")
Zeki's explicit finding: Indian AI professionals already abroad are not coming back in significant numbers. The improvement we are seeing is new graduates choosing to stay; the senior diaspora has not budged. (Mathrubhumi on Zeki)
Zero senior frontier-lab researchers from OpenAI, Anthropic, DeepMind, Meta or xAI have moved to an Indian lab. Every notable 2025 movement (Bansal, Agarwal, Beyer) was intra-US. (TOI on Bansal, ET on MSL departures)
The pay packets they're walking away from: $1M to $100M+ per individual researcher, sometimes $200M. (Wikipedia on MSL, ET on packages)
Multi-generational researcher pool - we don't have one
China graduates 3× the CS undergrads and ~90% more top AI PhDs than the US, and then the US imports the rest of them. (AI-Hive analysis of Stanford AI Index)
Zhipu's founding bench is from Tsinghua KEG with 15+ years of pre-training research predating the LLM era. (Amafi Advisory)
India's R&D spend is 0.64% of GDP against China's 2.68% and the US's 3.45%. (Forbes India on Economic Survey 2025-26, WIPO Global Innovation Index)
That's a 15-20 year pipeline rebuild from where we are today.
Domestic chip stack - we have nothing at scale
China's Huawei Ascend 910B trains half of China's top 70 LLMs. (SwarmSignal)
China has Cambricon, Moore Threads, Huawei and SMIC at 7nm in production now. India's Bodhi (Krutrim) target is 2027-2028, and Krutrim has laid off 200+ employees in 2025. (Outlook Business, ET)
Without indigenous silicon, every frontier-class training run is one US export-control update away from being cancelled.
Ecosystem coupling - the actual bottleneck
Frontier labs need capital, talent, compute, data, a sovereign customer and chip-vendor co-investment, all aligned and moving in lockstep. France assembled this for Mistral - Macron's €109B Paris summit announcement in Feb 2025, banks underwriting $830M of GPU debt, French Armed Forces as anchor customer. (France 2030 profile, CRN on debt round)
The frontier labs that ship had ecosystems sitting underneath them long before the founding pitch. OpenAI had Y Combinator, Sequoia, Microsoft, Stanford and 70 years of US AI research. DeepMind had London plus Google plus UCL and Cambridge. DeepSeek had a $10B AUM quant fund that had been pre-buying GPUs since 2020. Aligned ecosystems break frontiers, and India hasn't aligned one.
Show me a plan that creates a 1,000-strong cluster of senior frontier-trained researchers in India, $50B+ in patient capital, indigenous silicon and sovereign chip-grade compute coordination, and then we have a real conversation, and a shot at the frontier in well under 10 years.
Remember, China's GDP in 2006 was roughly the same as India's is today. On our current path, our future AI researchers are all in junior and middle school right now. Until that stack exists, everything is on a slow burn - 15-20 years. What we need is a viable strategy to shorten that to well under 10 years.
The application and middle layer is where India can credibly win, and where the case for trying is the strongest, because it is a path to spinning up the capital, compute, training data and talent flywheel required to get there sooner.