Everyone is buying the body. Nobody owns the brain.

For a decade the promise was that software would democratise opportunity. A laptop and a wifi connection, and a kid in Naraina could compete with a kid in Palo Alto. That promise is quietly inverting. As code gets commoditised (as the model writes the boilerplate, the test, the migration), the differentiator stops being can you code and becomes can you afford the tool that codes. And the tool is no longer free.

The economics here are genuinely strange. On one hand, raw intelligence has never been cheaper: Stanford's AI Index found the inference cost of running a GPT-3.5-class model fell roughly 280-fold between late 2022 and late 2024, with hardware costs dropping ~30% a year and energy efficiency improving ~40% a year. On the other hand, the frontier (the tools that actually keep you competitive) has settled into a subscription. Claude Pro, ChatGPT Plus, Cursor: the going rate to stay in the game is about $20 a month, and the genuinely capable tiers run $100–200.

Twenty dollars sounds trivial. It is trivial, but only if you earn in dollars. Run it as a share of income and the divide snaps into focus.

This is the first way India enters 2030 unprepared, and it is the most intimate one. It is not a data-centre statistic. It is a talented, motivated student who is now one paywall away from falling behind a peer who never had to think about the charge. Generative-AI adoption, the AI Index notes, correlates strongly with GDP per capita. The tools diffuse fastest exactly where they were already least needed for equality. A technology sold as the great leveller is, at the individual level, behaving like a sorting hat.

When the floor to compete becomes a recurring charge in a foreign currency, "access" stops being a slogan and starts being a class.

Give the IndiaAI Mission its due: it is real, fast, and unusually cheap for builders. As of early 2026 the national common-compute pool crossed roughly 38,000 GPUs, offered to startups and researchers at about ₹65–150 per GPU-hour (some 42% below market), with a target of 100,000 public GPUs by December 2026 and a path past 200,000 once Reliance, Tata and the hyperscalers are counted. Sarvam and Krutrim shipped the first sovereign Indian LLMs on that compute. This is genuine progress, and it deserves to be said plainly.

But scale and ownership are different things, and the gap on both is stark. Stanford's composite AI-capacity index for 2026 puts the US at 78.6, China at 36.95, and India at 21.59. The investment picture is starker still: US private AI investment hit $285.9 billion in 2025 (more than 23× China's $12.4 billion), while India's flagship national commitment, the entire IndiaAI Mission, is about $1.25 billion. That is not a typo. India's whole sovereign-AI budget is a rounding error against a single quarter of one American hyperscaler's capex.

And the compute India does run, it does not own. Every chip in that 38,000-GPU pool is an NVIDIA H100, H200 or Blackwell, and not one is domestically designed or fabricated. India imports roughly 20,000–25,000 high-end GPUs a year, about $2 billion worth. The Tata–PSMC fab at Dholera and the Micron plant at Sanand are real, but they target memory and packaging, not AI training silicon; fabrication of leading-edge accelerators is a decade away. The whole edifice rests on continued access to someone else's chips.

Which is the quiet emergency. The same US Commerce Department authority that cut China off from H100s in 2022 placed India in a "Tier-2" export category in 2025. A 2026 interim trade framework softened it, but a trade agreement is not a permanent guarantee. India has built its sovereign-AI ambition on a supply line that a foreign policy memo could throttle. That is not sovereignty. That is tenancy with a nice lease.

And then the wall no subsidy can climb: power

Chips are only half the bottleneck; the grid is the other half, and it's the half India can't fast-track. Training and serving frontier AI has quietly become an energy industry. Global AI data-centre power crossed 29.6 gigawatts in 2025, roughly enough to run New York State at peak. India's own stated ambition implies a 6–9 GW data-centre build-out over five to seven years (per Yotta's Sunil Gupta); Reliance alone is erecting a 1-gigawatt facility in Gujarat, expandable to 2 GW, at an estimated $20–30 billion. You can subsidise a GPU-hour down to ₹65. You cannot subsidise a gigawatt of reliable, cooled, watered power into a grid that still browns out, and India's electricity is costlier and less dependable than the American or Chinese baseline. The compute race is, underneath, a power race, and power is the input India is least able to conjure on command.

There is a rhythm to this field, and once you see it you can't unsee it. Roughly every four years, the ground moves: a single result resets what's possible and everyone spends the next four years catching up to it. Track it and the metronome is almost eerie.

2012: AlexNet wins ImageNet and deep learning becomes inevitable. 2016–17: AlphaGo beats Lee Sedol, and "Attention Is All You Need" introduces the Transformer that everything since is built on. 2020: GPT-3 proves that scale alone buys capability. 2022: ChatGPT turns it into a consumer phenomenon. 2024–25: reasoning models and test-time compute open a new axis. By the pulse's own logic, the next reset lands around 2028, comfortably inside the 2030 horizon.

Now overlay India onto that timeline, honestly. At which of those resets did an Indian lab, company, or university produce the breakthrough? The answer is none. India's role at every pulse has been the same: a magnificent adopter. The world's IT back-office in the 2010s; the world's largest pool of AI talent by the 2020s, much of it employed building application layers on top of other people's foundations, or building those foundations inside US labs after emigrating. India's first sovereign frontier-grade models, Sarvam and Krutrim, arrived only in 2025–26, a full cycle or two after the paradigms they're built on.

This is not a morality tale about effort; Indian researchers are everywhere at the frontier, often with a US zip code. It is a structural observation. A country that is always one pulse behind doesn't just lag in prestige. It lags in option value: it builds on foundations whose direction, pricing, and availability are set elsewhere, and it inherits the downside when those foundations wobble. Which brings us to the wobble.

Here is the cruellest irony of the lot, and the most immediate, because it isn't a 2030 abstraction, it's already on the 2026 payroll. India's modern prosperity rests on one brilliant bet: become the place the world outsources its knowledge work to. IT services, business-process outsourcing, call centres, the global code shop, a roughly quarter-trillion-dollar export industry employing millions, and the engine that lifted a whole generation into the middle class. India won that game decisively.

AI's first and sharpest productivity gains land on exactly those tasks. Stanford's AI Index puts measured gains at around 26% in software development, ~15% in customer support, and up to 50% in marketing output, the precise activities India staffs for the planet. The leading edge already shows in the labour data: US employment for software developers aged 22–25 has fallen nearly 20% since 2024. The bottom rungs (junior dev, QA, L1 support, data entry) are the most automatable, and they are exactly the rungs India's young workforce was supposed to climb.

The demographic dividend assumes the jobs at the bottom of the ladder still exist when the workforce arrives. AI is sawing off the bottom rungs.

This is the demographic dividend turning, quietly, into a demographic liability. India adds millions of graduates to the labour market every year on the premise that the service-economy ladder will absorb them; the country's own Chief Economic Adviser has begun warning publicly that AI threatens entry-level service jobs at precisely the moment that cohort arrives. And the final twist: the talent good enough to build the technology doing the automating frequently leaves to build it elsewhere. The frontier labs of San Francisco are staffed, in no small part, by the diaspora of the very workforce this displaces.

The point isn't that the jobs vanish overnight; it's that India is uniquely exposed because it specialised, successfully, in the layer AI automates first. Preparing for 2030 isn't only about building models, it's about having somewhere for a hundred million young people to go when the entry rungs disappear. That conversation has barely started.

Walk into any 2026 keynote and you'll hear it: vision is solved, language is solved, audio is solved, and the next frontier is physical AI: humanoids, embodiment, robots in the world. The capital agrees. Robotics investment surged ~300% in Q4 2025; Figure closed a billion-dollar round at a ~$39B valuation. The narrative is that the robot is the iPhone of the 2030s.

Here is the problem with that narrative, and it's the load-bearing argument of this whole essay: a humanoid robot cannot be commoditised, because it is a capital good, not a consumer good. Look at the actual numbers behind the hype and the timeline tells on itself. Goldman Sachs's base case is roughly 250,000 humanoid shipments in 2030, almost all industrial. Units cost between $16,000 (a bare Unitree G1) and $150,000; Tesla's aspirational price is $20,000–30,000 at scale. Morgan Stanley's blockbuster $5-trillion-by-2050 forecast still only puts ~80 million humanoids in actual homes by 2050, with consumer adoption not accelerating until the late 2030s.

So the bullish robot story, followed to its conclusion, says: through 2030 the body sells to factories and warehouses in the rich and the Chinese-manufacturing worlds, at prices no consumer pays, in volumes measured in the hundreds of thousands. That is a real and important market. It is also, definitionally, not a thing that gets distributed to a billion people. The body stays scarce.

Everyone is racing to fund the robot. Almost nobody is asking what runs inside it, and that is where the value, and the commoditisation, actually live.

What does get commoditised is the layer underneath the robot: the substrate. The cheap, fast inference. The small models that run on the edge. The world models and simulators that let a robot (or a game, or an agent) learn physics without a body. That substrate deflates relentlessly (remember the 280× cost collapse), it scales to everyone, and it is the thing that has to work before the robot is worth anything. The robot is the visible noun. The substrate is the verb. India, and a lot of capital, is betting on the noun.

Run the logic forward. The body stays scarce and enterprise-only. The substrate (cheap inference, small models, world models, agents) deflates and reaches everyone. So what is the consumer form that the substrate takes? Not a robot in your kitchen. It is the screen you already own. And the highest-bandwidth consumer use of interactive AI on a screen is gaming: model-generated worlds, NPCs with an LLM for a brain, speech in and speech out, the whole immersive loop, distributed to a billion phones at near-zero marginal cost. Gaming is to consumer physical-AI what the browser was to the internet: the place the abstract capability becomes a product people actually touch.

The market agrees. Global gaming runs from about $298 billion in 2024 to ~$505 billion by 2030. India is the demographic jackpot of that story, a gamer base growing from roughly 433 million to 657 million by 2030, 659 million smartphones, UPI making microtransactions frictionless, 5G collapsing cloud-gaming latency. On paper, India is built to win the consumer-AI layer.

And there is the catch, in one line: over 60% of India's top mobile titles are foreign-made, and India is about 5.2% of global gaming revenue while being a far larger share of its players. The pattern is identical to the chip story and the model story. India shows up as the world's most valuable market and almost never as the maker. It consumes the layer it is best positioned, demographically and linguistically, to build.

This is the essay's one genuinely hopeful turn, because the AI-native gaming stack is the rare frontier where India's actual strengths line up. The consumer layer rides open-weight models and a creator/modding economy, which is exactly where the cheapest compute on earth (India's subsidised GPU pool) and the world's deepest pool of application talent could compound. The NPC-brain layer wants small, fast, multilingual models, and India's Indic-language and voice data is a moat almost no one else has. The substrate India should be racing to own is not the humanoid. It is the cheap inference, the small Indic models, the world-model and voice stack that makes a generated, fully-voiced, Hinglish game world run on a $200 phone. That is a layer that commoditises, scales to a billion people, and plays to what India already is.