the wrong panic

Vishal Dhupar, who runs Nvidia's India business, spoke at our AI summit at Masters' Union last week, and he narrated a small story about Jensen Huang, who founded Nvidia.

Before AI, Nvidia was a graphics chip company.

They made the processors that drew graphics inside video games, and there was no particular reason to believe the company would ever do anything else. 

In 2006, Jensen made a decision that nearly broke Nvidia. 

He spent years and billions of dollars building a piece of software called CUDA, which is a set of instructions that lets a programmer tell a graphics chip to do something other than draw a game.

Think of it as the translator between ordinary code and a very specialised kind of chip. 

CUDA sat on Nvidia's balance sheet as dead weight for six years, with investors begging Jensen to shut it down.

Then came 2012, and researchers figured out that the same chips used for video game graphics were weirdly perfect for training a kind of AI called deep learning. 

Every AI researcher in the world wanted those chips, and the only way to speak to them was CUDA. 

This cuts against almost everything governments are doing about AI right now.

Iran has been holding the Strait of Hormuz for fifty years. 

You would think that makes Iran rich.

The average Iranian makes about $4,500 a year, while the average Japanese person, whose country owns no oil at all, makes closer to $33,000.

The countries that got rich from oil were the ones who had somewhere to put it once they bought it: electronics factories in Japan, chemistry labs in Germany. 

Most countries are walking into the AI economy looking to be Iran, and India, despite being in the one position that could let it avoid that fate, is currently one of them.

There are five layers to what we now call AI: electricity at the bottom, then chips, factories that manufacture the chips, AI models built by frontier labs, and on top of all of that, the actual uses people put AI to in their work.

Almost every government conversation I'm in right now is about the middle of that stack, because the concentration there is frightening..

The chips themselves are mostly made by one company, a Taiwanese one called TSMC, which happens to sit on an island China claims as its own. 

The rare metals inside those chips are almost entirely refined in China, which started weaponising them against the US last December. 

A new chip factory takes five years to build.

It is whether your hospital can deploy an AI that reads scans before three years of pilot studies eat the budget. 

Whether the law firms in your country have rebuilt around the tool, or are still using it as a slightly unreliable intern. 

And whether your government can buy a piece of software without writing a two-hundred page tender.

The answer in most countries is no.

But this is where all the wealth will come from, because people won't lose jobs to AI so much as lose tasks.

And the boring repetitive parts of every job get absorbed by the tool while whatever the person can do that the tool can't become the thing markets will pay for. 

That is the country-scale question too: what is your economy still good for, once the chip stops being scarce?

We are third in the world in rare earth reserves and produce less than 1% of them, and the chip factories we are building today, in Dholera and Sanand, will produce chips of a kind the rest of the world was making fifteen years ago.

We have no ASML, TSMC or NVIDIA equivalent, and we won't have them this decade.

The government is running two AI bets at the same time.

The other is the IndiaAI Mission, with a five-year budget of about ₹10,000 crore, which is subsidising shared computing power for Indian startups and universities.

It is fully funding four homegrown startups trying to build India's own AI models (Sarvam notably received ~₹100 crore in subsidised Nvidia hardware for this), and running a national skilling programme. 

Both of these are serious programmes.

But the ratio between them is roughly fifteen to one in favour of manufacturing.

And yet.

We have the highest daily ChatGPT usage globally, a $283 billion services industry where Indian engineers are already doing the AI work for a lot of the Fortune 500, and TCS alone has trained 350,000 of its employees on these tools.

The workforce is already ahead of where it needs to be.

Flip that ratio over the next five years.

Match every rupee of chip manufacturing with fifteen spent on compute, application-layer startups like Sarvam, rebuilding the procurement systems that stop our hospitals and courts and schools from deploying anything new.

That, to my mind, is a trillion-dollar shift in where AI output gets produced, and we are currently positioning ourselves to miss it.

“Nandini, you might be saying, this is convenient.

India happens to be bad at the things you claim don't matter and good at the things you claim to do, and also the government is already doing half of what you're asking for through the IndiaAI Mission.”

The second point is fair.

The Mission is working, and Sarvam getting a government-backed chance to train a sovereign Indian language model is quite significant.

But ten thousand crore for compute and applications against 1.6 lakh crore for fabs is a statement of priorities that shows how we still think the value is in the middle of the stack.

The chips are going to arrive eventually, the same way oil became abundant in the 1980s.

The countries that spent the oil panic buying the resource were not the ones that built the rich industrial economies of the decade that followed.

India already has the engineers, companies, and usage that the rest of the world is still trying to build.

And we are about to spend the next decade solving the wrong problem because we have confused someone else's panic for our own.

If you work anywhere near this, in a company or in policy, reply and tell me what you're seeing. I want to know whether I'm early, or just wrong.

Nandini