NVIDIA Gave India a Recipe, Not Data
The corpus is a one-time gift. The recipe is a standing order for compute. The smartest way to sell compute is to give away the recipe that eats it.
NVIDIA didn’t publish a dataset this week. It published a demand curve.
The announcement is Nemotron-pretraining-Indic-v1: about 2.8 trillion tokens across the 22 languages listed in India’s Constitution, built with IIT Madras and its AI4Bharat lab, “releasing soon” along with the open, GPU-accelerated pipeline that produced it. The token count is the headline. The pipeline is the part that matters.
Start with the strongest case for taking this at face value, because it’s a good one. NVIDIA says those 22 languages are the first languages of 97% of Indians, nearly 1.2 billion people at the 2011 Census. Most of those people have been served by models that learned mainly from English. AI4Bharat has spent years building open-source Indic tooling. A corpus this size, with native-script data for all 22 languages and romanized data for twelve, fills a real gap, and anyone can download it once it ships. That’s a public good. I believe it is one.
Now look at how it gets made. The blog lays out the path “from English source documents to Indic training data”: translation, quality assessment, and repair, run as LLM workflows with prompts and quality checks, all GPU-accelerated. Then comes the sentence that matters. The pipeline “captures the full generation recipe… so developers can adapt it to their own source material and generate more data for their needs.”
That sentence is the business model. The corpus is a one-time gift. The recipe is a standing order for compute. Every new token anyone makes this way comes from running models on GPUs. NVIDIA has taken the thing that stops a ministry, a Bengaluru startup or a university lab from training an Indic model, which is too little data, and turned it into a hardware purchase.
This is commoditize-your-complement, the old Joel Spolsky rule, played very well. NVIDIA sells compute. The main thing that goes with compute is a reason to use it. A shortage of data caps how much compute a language market will ever buy. Remove the shortage by making data something you produce with compute, and the cap goes away. NVIDIA has already given away the software layer and open Nemotron models. Giving away the data recipe is the next layer down. Each giveaway pays off in the same place: the GPU bill.
This isn’t a scandal. It’s the best kind of business: you give away the thing that makes your product necessary, and the gift is real. Taking the money doesn’t make the public good fake, and the public good doesn’t make the money incidental. Both are true, and the second one decides who holds the leverage.
The sovereignty angle is the interesting part. Every government now wants its own model, and “sovereign AI” is the pitch that sells the hardware. Look at what this corpus is built from, though. It’s synthetic. NVIDIA built it by running documents from 15 English Nemotron-CC-v2 subsets through a translation and repair loop. NVIDIA’s own limitations section says it plainly: translation “does not replace text originally written in Indian languages or guarantee coverage of local knowledge and culture.” Credit to them for writing that down. It’s sensible engineering, since the open web doesn’t hold trillions of tokens of native Indic text waiting to be scraped. It also means you get sovereignty at the language layer while depending on two others: the English source canon and the silicon that converts it.
Power structures matter more than stated policies. The stated policy here is openness: open data, open pipeline, adapt it as you like. The power structure is that the cheapest route to the next trillion Indic tokens runs through one vendor’s recipe on one vendor’s kind of hardware. In principle anyone can fork it and run it on any accelerator. My bet, and I’m labeling it a bet, is that a pipeline NVIDIA built with AI4Bharat and pitched as GPU-accelerated will run best where NVIDIA wants it to.
None of this makes me pessimistic. Open data compounds. An open recipe beats a closed one by a mile, and India’s labs are better off with it than without it. AI4Bharat gets reach it couldn’t buy. Indian developers get a starting point that didn’t exist last month. Technology compounds, and this compounds in India’s favor too.
Trust gets earned one action at a time, though, and right now this action is an announcement. The corpus is “releasing soon.” Judge it when the files land, the license is readable, and someone has rerun the pipeline on non-NVIDIA hardware and published the numbers.
For anyone running a national AI program, the lesson isn’t “don’t take the gift.” Take it. But ask the right question. Having your own model is the easy part now. The hard question is who you call when you need the next trillion tokens. If the answer is the same company every time, you didn’t build sovereignty. You picked a supplier.
The smartest way to sell compute is to give away the recipe that eats it.