AI Vey is a podcast about AI in India, for India, and from India. Most weeks that means resisting the gravitational pull of the American news cycle. This week it meant something harder: sitting with a guest who thinks the entire Indian AI conversation — models, GPUs, sovereignty, the lot — is organised around the wrong noun. His claim is that the question is not what we can build. It is whether what already exists will reach a district court in Maharashtra, and what has to be true of a society for that to happen.
Pranay Kotasthane is deputy director of the Takshashila Institution and chairs its High-Tech Geopolitics Programme. Before he was a policy person he spent seven years designing chips at Texas Instruments, which is why he can tell you what an EDA licence costs and why that matters. He co-wrote When the Chips Are Down with Abhiram Manchi, the first book to look at semiconductor geopolitics from an Indian vantage point; he co-writes the newsletter Anticipating the Unintended with RSJ; and he co-hosts the Hindi-Urdu podcast Puliyabaazi. He also ships software. That last part is why he belongs on this show: most people who write about AI policy in India do not open a terminal, and most people who open a terminal do not write about policy.
Some of what you get in the hour: the argument that the absence of state capacity is a reason to use AI rather than a reason to despair of it, worked through the specific case of court scheduling. A correction to the founding myth of Indian software — the industry did not grow because the government stayed away, it grew because software services were filed under the Shops and Establishments Act instead of the industrial and labour acts. And a warning about buying GPUs at national scale that ends on the phrase “what will we do of that bricked chip?” Along the way, a tour of what a 40-person think tank actually builds when it decides its only durable advantage is disproportionate use of technology — including a bot that trawls the internal channels and produces an Eisenhower matrix every Monday.
Highlights
* The frame: success in AI is a societal problem, not a technological one — Pranay opens on a line from Michael Mazarr’s RAND work on the societal foundations of national competitiveness, which looks back at earlier technological revolutions and finds that the role of diffusion is consistently underplayed. He then runs the Indian question through Samaj, Sarkar, Bazaar rather than through chips and compute. “We don’t need an AI strategy, necessarily.”
* State capacity is the argument for AI, not against it — Ashish asks the obvious sceptic’s question: whatever you design for the centre will break at the local level. Pranay inverts it. Not having people is precisely why you reach for the tool. His example is court scheduling — judges spend enormous time maintaining rosters rather than deciding cases, which XKDR Forum has documented, and roster-building is an operations research problem. The alternative is recruiting a new cadre of court managers, which stalls exactly the way police vacancies stall.
* Produce, finance, regulate — governments do three things, so ask the question three times. Produce AI: no. Finance it: yes, and more usefully as a procurer than a funder — nobody builds AI-integrated autonomous weapon systems in India unless the government places volume orders up front. Regulate it: India, unusually, is getting this right. No AI-specific regulator, no ministry of AI, no per-model licensing, in deliberate contrast to the EU.
* The Shops and Establishments Act, and the myth it corrects — the best two minutes in the episode. Navin asks whether Indian software grew because the government stayed out, and whether it should stay out again. Pranay’s answer is that it never stayed out. It made a choice: software services were classified under shops and establishments rather than the industrial acts, with their hiring, firing and PF regimes. Regulation existed; it was just the lighter kind. The live question for AI is which side of that line it lands on.
* A Purvapaksha of his own co-author — Ashish asks Pranay to steel-man RSJ’s bearish case on Indian IT services: with AI and a solid engineering team, enterprises bring the work in-house, and demand for Indian services falls over three or four years. Pranay’s counter is that legacy companies cannot move to AI workflows unaided, and that the leaders of the services and SaaS eras need not be the leaders of the AI era. “Why should the TCS of the AI era be the TCS of the software services era? It’s good if there is churn.” What he does concede: the intake numbers from engineering colleges will not hold.
* Richard Baldwin’s asymmetry — manufacturing supply chains are shortening under geopolitical pressure, but services supply chains can lengthen because of AI. Navin pins down what “lengthen” means — literally distance in metres — and the worked example is surgery: today the doctor flies to the patient and requalifies in that jurisdiction; with good enough robotic and AI-assisted surgery, the doctor stays in India. Navin brings the week’s news to the same point: an Anthropic and Goldman Sachs enterprise AI tie-up read by some as the end of Infosys, and by others as the beginning of Infosys++, because enterprises need more hand-holding than a model vendor wants to provide.
* “There is nothing like sovereign AI” — the flattest rejection in the episode. You cannot be sovereign in semiconductors, so the idea that you can be sovereign in AI does not survive contact. What is sovereign, he argues, is not AI but particular applications: defence targeting, or a Ministry of External Affairs model reading ten-year risk, where you want something fine-tuned on your own data. Build those on open weights. Sarvam exists; the harder question is who adopts it and why they would.
* What ₹10,000 crore should have bought — a large share of the national AI mission went on NVIDIA and AMD chips. Pranay thinks the intent is right and the prioritisation wrong. He would run a challenge grant aimed at getting out of CUDA — Apache TVM, MLIR, LLVM, ONNX Runtime, the work of making the GPU a switchable layer underneath PyTorch. The reason is not techno-nationalism, it is the export-control ratchet: US diffusion rules, on-chip thresholds, and then “what will we do of that bricked chip?”
* And why the alternative needn’t be Indian — RISC-V started at UC Berkeley, and its foundation moved to Switzerland precisely to sit away from both Washington and Beijing; MeitY is now a sponsor. Pranay’s test is not who builds it. “If it is available to every Indian, then our goal is satisfied.” He adds the corrective he keeps having to make in semiconductor debates: we cannot become Atmanirbhar, and neither can Taiwan.
* Open versus closed, not China versus the United States — and the worry that actually keeps him up — on whether Indian firms should use Chinese open-weight models, Pranay’s answer is that open weights are open to everyone, and it merely happens that many of the leading open models are Chinese. Navin adds the technical half: a backdoor in a model is not a backdoor in a Huawei chip or in shipped software — you cannot steer it the same way, and fine-tuning can wash it out. Pranay draws his line at hardware in critical systems: not in telecom, probably fine in an electric vehicle with audit mechanisms. But the risk he takes seriously is not the bazaar’s, it is society’s — a stack fine-tuned to omit things produces users for whom Tiananmen Square did not happen, and the quotidian example matters less than the structural one: a slow drift toward believing that authoritarianism works, because that is what the information said.
* Substitution versus stimulation, and the skill that now matters — the distinction he drew in the newsletter and sharpens here: substitution is outsourcing your thinking, which for a think tank is fatal; stimulation is widening the range of problems you can attempt. Which leads to the practical claim — with execution cheap, asking the right question has become disproportionately valuable. In Takshashila’s weekly session he makes people start there: name the unique problem in your research domain.
* What a 40-person think tank actually ships — a weekly show-and-tell with two or three people rostered to demo, a Mattermost channel where experiments get posted, and roughly 20 of 40 colleagues who have now built their own sites and trackers. Non-engineers learned Git because they had to. The geospatial lead built a mapped analysis of the Iran war’s effect on India with Claude Code that would otherwise have taken two more people. There is a Claude Code plus ElevenLabs “fourth guest” with a persona that steel-mans a position through an episode; a tool that turns a PDF or EPUB into a 45-minute podcast, with Marathi output via Sarvam, on the theory that most non-fiction books are a paper with history stapled to it; an institutional brain that classifies decisions into decision, playbook, or principle so a joiner in three years can find the provenance nobody wrote down; and the Eisenhower-matrix bot. Navin’s request, on air, is that Pranay write the show-and-tell method up — he calls it homework.
* The diffusion balance sheet — Pranay then scores India against Mazarr’s societal factors, one at a time. National ambition and will: better than before, because Viksit Bharat 2047 is a defined and measurable target — a developed country means about $13,500 per capita. Unified national identity: hard here, though UPI shows adoption can outrun education. Shared opportunity: the real promise, which is why speech-to-text and text-to-speech matter so that this does not stay with “the English-speaking elites like us.” An active and engaged society: fine, median age under 30. Effective institutions: the acute weakness — and note that GST’s data architecture, the vaccination drive and direct benefit transfers worked because technology came with institutional redesign, with UIDAI built new rather than bolted onto an existing department. Intellectual climate: credentialism will matter less and our institutions are stuck on it. Diversity and pluralism: an Indian advantage.
* Choices Choices — coffee with Jensen Huang or Dario Amodei: both hosts guess Jensen, and Pranay agrees, on the grounds that it would be good to have a fight with him. Reborn in the early 1980s onto a TSMC fab floor, or exactly where he is now: he takes now, and corrects the premise — TSMC was founded in 1987, so he would have had to be born in the 1960s. A year inside the India AI Mission or inside an Indian semiconductor startup that ships: Ashish guesses the startup, Navin guesses the mission, and Pranay sides with Navin — “Navin knows me better than Ashish” — because getting government policy 1% right is a non-linear payoff, and there are far fewer people who can do that than there are good chip engineers. Finally, model monogamy for a year, Claude or Kimi: “My stated and revealed preferences are different.” He is experimenting with open models, still shipping on Claude and Codex, and asks to be asked again in a month.
Notable quotes
“I don’t think there is anything like sovereign AI. You cannot become sovereign in AI. I mean, you can’t become sovereign in semiconductors — how will you become sovereign in AI? No country can.”
“The fact that you don’t have state capacity is itself reason enough for you to then use AI.”
“I don’t want us to think it should be something Atmanirbhar, that only Indians will build. If it is available to every Indian, then our goal is satisfied.”
“If all our information is through that, that is the challenge. So how do we preserve cognitive autonomy in the AI age?”
“One year of second best is not as good as one year of best.” — Navin Kabra
“I’m more hopeful about India’s diffusion than India’s innovation.”
Links & resources
Everything below was named on air. Where a URL was not given, we have said so rather than guessed.
* When the Chips Are Down — Pranay’s book on semiconductor geopolitics, with Abhiram Manchi.
* Anticipating the Unintended — the newsletter he co-writes with RSJ.
* Puliyabaazi — his Hindi-Urdu podcast on policy and technology.
* Michael Mazarr, The Societal Foundations of National Competitiveness (RAND, 2022) — the source of the diffusion framework the episode is built on, and of the seven societal characteristics Pranay scores India against.
* Kai-Fu Lee, AI Superpowers — cited for its four-wave framework and the claim that enterprise AI is where adoption is slowest and the money is.
* Richard Baldwin — on manufacturing supply chains shortening while services supply chains lengthen.
* XKDR Forum — cited for work on judicial scheduling and case management.
* RISC-V and the RISC-V Foundation — started at UC Berkeley, now based in Switzerland, with MeitY among the sponsors.
* Apache TVM, MLIR, LLVM, ONNX Runtime, PyTorch — the open-source layer Pranay would fund to make GPUs switchable and break the CUDA dependency.
* Synopsys, Cadence, Mentor Graphics (now Siemens) — the three EDA incumbents, and the reason two Bangalore companies are building AI-based tool flows for mixed-signal design.
* Sarvam-105B — Pranay’s “the Sarvam 105”, cited as the existence proof that India does have a sovereign model, and again for the Marathi output in his book-to-podcast tool. The number is both the name and the parameter count.
* DeepSeek, Moonshot / Kimi, Mistral — the open-weight models under discussion. Model versions were not always clear on air.
* OpenCode — the multi-model router Pranay was pointed to after tweeting about Kimi’s performance.
* Mattermost — Takshashila’s open-source internal communication tool.
* Claude Code and ElevenLabs — the stack behind the geospatial mapping tool and the AI “fourth guest.”
* EkStep and Nandan Nilekani — cited on benefits reaching across the economic spectrum rather than trickling down, with farmer advisory tools in Maharashtra as the example.
* Andrej Karpathy’s LLM wiki — Navin’s recommendation, and his argument that every person, not just every company, should keep one.
* Should the US Sell Advanced GPUs to China? An Indian Perspective — Takshashila, 29 April 2026. The paper on more competition being net-net good that frames the first Choices Choices question.
* TheProfesseer — theprofesseer.com. The litigation-analytics engine collecting records from various courts so a litigant can judge an advocate’s track record, which matters because lawyers cannot advertise.
Connect with Pranay Kotasthane
* Takshashila Institution — takshashila.org.in
* Website — pranaykotas.com
* Twitter — @pranaykotas
* Anticipating the Unintended — the newsletter, with RSJ
* Puliyabaazi — the Hindi-Urdu podcast
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit podcast.aivey.in
