Written for Dwarkesh Podcast

What should countries which are not currently in the AI production chain (semis, energy, frontier models, robotics) do in order to not get totally sidestepped by transformative AI? If you're the leader of India or Nigeria, what do you do right now?

The traditional wage arbitrage path that worked for Japan, Korea, Taiwan, and China is being closed by the country that just walked through it. China's 15th Five-Year Plan (March 2026) places humanoid robots and embodied AI, alongside integrated circuits and biomanufacturing, as one of ten prioritized "new industry tracks." The language is action focused: coordinate training grounds, develop embodied models, tackle core components, accelerate deployment. It comes backed by a 1 trillion RMB ($138B) state fund. India has a window, defined not by her or Modi's ambition but by the price curve of a robotic product line, to build something other than a wage-arbitrage manufacturing base, or its in-progress foundations get repriced from underneath it.

The path India is betting on is traditional: Japan, then Korea, then Taiwan, then China, then Vietnam; each successor undercut the predecessor on wages, absorbed the previous generation's capital equipment, and worked its way up the value chain. Modi is betting India is next. Policies like the PLI, Tamil Nadu's iPhone hubs, the 32% global iPhone production target by 2027 are the same play.

This play has a problem, shown by existing unit economics: a Unitree G1 sells for $13,500. A Foxconn assembly worker in Sriperumbudur earns about $2,500 a year. Chinese firms shipped 90% of the world's humanoid units in 2025, and average humanoid prices fell from $85,000 to $25,000 in two years. The newest factory in Foshan rolls one humanoid off the line every thirty minutes. What happens when humanoids get more dexterous and cheaper in the next 24 months? The arbitrage is gone at price. What remains is deployment friction: integration, reliability, dexterity, and the 5 Year Plan is how it'll be funded down.

For iPhones specifically: India produces ~17% of global iPhones, but domestic value addition is only around 20%. Many components are still imported from China, yet final assembly in India costs ~10% more than China despite lower wages, because supply chain and worker productivity haven't caught up. The planned wage arbitrage to manufacturing powerhouse mission is fragile because it exists on a thin assembly layer of a value chain China still owns, the target for humanoid robotics.

So, when it comes to planning ahead and strategizing for AI, we need to be honest about the clock, because at most, there are a few years until humanoid robotics can functionally attend an assembly line; assuming the rate of improvement for the last 5 years, I'm predicting that 2030 is the latest year assembly line economics will be repriced. There are constraints we need to consider:

India does not have one industrial policy but four functioning ones: Tamil Nadu, Karnataka, Gujarat and Telangana, and the tail end of states that are still solving problems from a different decade. States like Bihar and UP aren't going to develop into Tamil Nadu in five years. In my eyes, a realistic AI strategy isn't national reform, but a concentration of bets on the four-ish states that can already execute and have the infrastructure. It's like California, Texas, New York spearheading AI innovation while Missouri, Louisiana, Alabama experience trickle down effects to develop faster.

First: real domestic manufacturing, not final-step assembly. The original PLI scheme expired March 2026, with its successor PLI 2.0 (in discussions at MeitY as I write this) is reportedly designed to do exactly the right thing: shift from rewarding assembly volume to rewarding domestic value addition and component depth (display modules, camera modules, PCBs). The execution will determine whether India enters the 2030s with a real components industry or an inflated assembly base waiting to be automated.

Second: build a domestic humanoid components industry on top of the existing electronics base. This seems counterintuitive but is potentially the most important. India has a window to build its own Unitree, and it exists because of Tamil Nadu's electronics ecosystem. Things like actuators, sensors, controllers, dexterous hands, the exact list named in China's 15th FYP is also a procurement list for what India needs to make.

If India doesn't build domestic humanoid capacity in the next five years, it will be importing the machines that displace its own factory workers from the country that displaced them, a cruel irony. A robotics-and-components fund ($5-10B) over five years, scoped to the four states, with the same export-performance and component-depth conditions as phones. India has the talent and manpower to compete on robotics components against China.

Third: build sovereign inference capacity, not a model. India won't win the model race against Anthropic, OpenAI, or DeepMind this decade and it doesn't need to. The model is a commodity; open-weight Deepseek and Qwen variants close most of the capability gap. The asset is the layer underneath: data center capacity, power, and fine-tuning data. India needs the physical inference infrastructure to serve 1.4 billion people in dozens of languages, running on cheap power, governed by Indian law. There needs to be structured access to court records, medical corpus, and text that Western labs can't license. The combo of open-weight models, Indian compute, Indian data is defensible past 2030 because it isn't a wage game and effectively deploys AI as industrial policy.

The binding constraint cutting across all three is talent. Roughly a third of frontier-lab AI researchers are Indian-origin, but they're in San Francisco and not India. China has Thousand Talents and K-Visa for AI, Taiwan has Hsinchu-Silicon Valley circulation, but India has nothing equivalent at scale. This diaspora is India's largest asset and the cheapest move on this list: equity, tax treatment, research freedom, mission, etc. No strategy or recommendation works if the engineers who can build them stay in Palo Alto.

India has the largest deployment market in the world, the deepest language and data moat outside of Mandarin, a highly educated base and a rapidly modernizing infrastructure. If I were Modi, I would see that the traditional 'ladder' has been erased by AI and robotics, and take steps to build a new one with the PLI 2.0.

As a disclaimer, any information published is up-to-date as of April 26, 2026. I am not an expert in robotic engineering or Indian politics, feel free to contact me at echen1246@gmail.com for any corrections or debate, open dialogue is welcome! Should not be taken as advice.