AI Training Data Is Becoming the Next Big AI Infrastructure Business — And India Could Be a Major Winner

For two years, the AI race has been a story about chips. Whoever had the most GPUs, the biggest data center, the fattest compute budget — that’s who was winning. AI training data just quietly became the next battleground, and the number that proves it is $350 million.

That’s what Snorkel AI raised in a Series E on September 22, 2026, at a $3.5 billion valuation — nearly triple what it was worth 17 months ago. Insight Partners and S32 led the round, joined by more than a dozen backers including Alphabet’s GV. The San Francisco-based company, founded out of a Stanford research project in 2019, says the round confirms something the industry has been circling for a while: once you have enough compute, the bottleneck moves to the data good enough to actually make a frontier model smarter.

From “More Data” to “Better Data”

The first wave of large-scale AI training was a volume game — more text, more code, more human feedback, scraped from wherever it could be found. That worked when models were comparatively weak. It stops working once a model is already good at ordinary tasks. What moves the needle now is much harder to produce: examples of how an experienced engineer actually diagnoses a gnarly production failure, or how a senior lawyer reasons through a genuinely ambiguous clause — not another million lines of routine code.

Snorkel’s own framing for this, per CEO Alex Ratner’s announcement, is a shift from “Data 1.0” to “Data 2.0” — from labeling logistics to expert-grade datasets and reinforcement-learning environments built specifically for what a model still gets wrong.

The Numbers Behind the Round

Reuters, which broke the story, reported that Snorkel’s annualized revenue run-rate has crossed $350 million, up from roughly $20 million a year earlier — driven almost entirely by the data-as-a-service business it launched in September 2025. Ratner’s own blog post claims a slightly higher figure: a $375 million run-rate, up more than 18x in a year. Both numbers point the same direction; they’re just not identical, and it’s worth knowing which one you’re quoting. Neither is independently audited.

India’s Opening: From Annotation to AI Data Engineering

India already runs a large share of the world’s data-annotation and BPO work. The Snorkel round suggests the more valuable tier above that — expert-grade datasets, evaluation benchmarks, domain-specific reasoning data — is where the real margin is shifting. That’s a different business: healthcare datasets that need a doctor’s judgment call, legal reasoning benchmarks built around actual Indian case law, financial-compliance data, or high-quality datasets across India’s languages rather than the English-centric material most frontier models are still trained on.

India’s government is already building policy infrastructure for exactly this. The National Data Governance Framework rests on four pillars — data-sharing policies, standardisation, data-exchange platforms, and governance architecture — and its data-exchange pillar names API Setu, NAPIX, AI Kosh, IUDX, DigiLocker and Entity Locker as the platforms meant to carry it. AI Kosh specifically functions as a repository of non-personal, anonymized Indian datasets intended to support AI model training and reduce reliance on Western-centric data like ImageNet.

The Catch: Trust, Not Just Volume

None of this works if the data is bad. Poorly sourced, duplicated or contaminated datasets make models worse, not better — and once personal, medical or financial information is involved, consent and de-identification aren’t optional. India’s own framework explicitly builds in consent mechanisms and dataset classification for this reason. The opportunity isn’t “produce more data”. It’s produce data that can survive the question: where did this come from, and can you prove it’s clean?

Bottom Line

The next AI infrastructure race may not be won purely on GPU count. Snorkel’s $350 million bet says the data layer is becoming a real business category of its own — and for India, with its software talent, domain expertise and linguistic diversity, that’s a more defensible opportunity than trying to out-compute Silicon Valley.