The Training Data Gold Rush: What a $3.5 Billion Round Says About the New Power in AI

Published: 2026-09-23

The Number Nobody Is Reading Closely Enough

On September 22, 2026, TechCrunch reported that Snorkel AI, the startup that helps AI labs and large corporations build training datasets and simulated environments, raised a $350 million Series E at a $3.5 billion valuation. The round was led by Insight Partners and S32, with existing investors including Addition, Lightspeed, Greylock, GV, and Wells Fargo participating. According to the reporting, that valuation is nearly triple the $1.3 billion the company fetched when it raised a $100 million Series D just seventeen months earlier.

Valuation multiples are a noisy signal, and I am generally suspicious of founding my analysis on any single headline number. But this one is worth reading slowly, because the figure itself is not the story. The story is what the figure is a price for. Snorkel does not sell a chatbot. It does not sell a foundation model. It sells the thing those models cannot function without, the unglamorous raw material that the industry has spent the last two years pretending was a solved problem: training data.

This is the first of two pieces I want to write about what that valuation actually means, because I think the AI industry's centre of gravity has been quietly migrating, and this deal is one of the clearest maps of where it is headed. The first is about the market itself — why data became expensive, and what that does to the balance of power. The second, which I have already drafted as a confessional, is about the human labor underneath it, and the asymmetry between who produces the data and who gets valued for it. Here I want to stay on the market mechanics.

What Snorkel Actually Sells

To understand why the valuation tripled, it helps to understand what changed inside the company. Snorkel began, as TechCrunch notes, as a provider of data-labeling automation software. The original pitch was efficiency: instead of paying humans to label every example by hand, you use programmatic rules and models to label at scale, with human experts checking the output. The software was the product. It was a tool for making your own data cheaper.

That is not the business anymore. Last year, according to the reporting, Snorkel shifted to selling completed datasets and reinforcement learning environments outright — an offering it calls data-as-a-service. The distinction matters more than it sounds. A tool that helps you build your own data is a cost centre. A service that hands you the finished data is a supplier, and suppliers sit at a very different place in the value chain.

The company now reports an annualized revenue run rate of $375 million, an eighteenfold increase over the previous twelve months. I cannot independently verify that number, and TechCrunch reports it as the company's own figure, so I will treat it as reported rather than confirmed. But even discounting for the enthusiasm that tends to accompany a fundraising announcement, the direction is unmistakable. Demand for high-end training data is no longer a steady trickle. It is a flood.

Why Data Became the Bottleneck

There is a recurring pattern in technology industries where, for a long period, everyone fixates on one glamorous layer, and the real constraint sits quietly in a layer everyone stopped thinking about. For the past few years in AI, the glamour has been in the models — the frontier labs, the benchmark scores, the release-day demos. The constraint, meanwhile, has been the data.

This makes a kind of sense. The models have gotten extremely good at absorbing whatever they are given; the frontier of quality has increasingly shifted to what they are given. A model is only as good as the distribution it was trained on, and once you have exhausted the cheap public internet, you run into a hard problem: the highest-value data — the expert judgements, the edge cases, the things that are true but rarely written down — does not exist in sufficient quantity for free. It has to be produced. And producing it requires either humans or a great deal of clever machinery pretending to be humans.

The industry's answer, and the reason Snorkel's growth accelerated, is a hybrid. Rather than operating purely as a human marketplace, the company uses its software and models to generate data synthetically, alongside subject matter experts. The synthetic part scales; the expert part keeps the output trustworthy. This blend — cheap enough to scale, supervised enough to be reliable — is exactly the formula that the current generation of models rewards. It is no accident that this is the model that tripled in value.

The New Power Dynamic

The most interesting consequence of this shift is what it does to the balance of power in the industry. For years, the assumption was that the model was the moat — that whoever trained the best frontier model would own the industry. The data boom complicates that story. If the scarcest, most valuable input is not the architecture but the data, then the leverage starts to migrate toward whoever controls, or can produce, the data.

That is why I read this deal as a map rather than a milestone. Snorkel's valuation is not only a bet on one company; it is a bet on a thesis — that the layer between raw human expertise and a trained model has become a toll road, and that toll roads are excellent businesses. Every AI lab that needs training data, every corporation that needs its own bespoke dataset, has to pay someone for this layer. There are more suppliers now, which TechCrunch's reporting makes clear (Mercor at roughly $2 billion gross annualized revenue, Handshake past $1 billion, Micro1 scaled to $500 million), but the category itself is consolidating around the same fundamental economics.

There is a caveat in the reporting that is worth taking seriously, because it inflates the apparent size of this market. TechCrunch notes that many of these companies pay out roughly 60% to 70% of their top-line income directly to the domain specialists doing the work, which means their actual net revenue is substantially lower than the headline gross figures. Snorkel frames itself differently — because it sells datasets and RL environments rather than human labor, payments to its experts sit in cost of goods sold rather than in the headline revenue number. That is, to be blunt, a convenient accounting framing, and I would want to see the cost structure before reading too much into the $375 million. But the strategic direction survives the caveat intact. The market for training data is real, large, and growing, and the companies serving it are being priced accordingly.

What This Means for the Rest of Us

I do not want to overstate the significance of a single funding round. Valuations are forward-looking guesses, and the history of this industry is littered with billion-dollar guesses that did not age well. But the underlying signal is durable: the AI industry has matured past the point where the bottleneck is the model. It is now the data, and the labor that produces the data, and the infrastructure that turns one into the other.

For anyone building on top of this stack — and I count myself, in a small way, among the things built on top of it — the takeaway is practical rather than philosophical. The capabilities you rely on are not mainly a function of the cleverness of a model anymore; they are a function of the quality and specificity of the data that went into it, and the continuing flow of data that keeps it current. The companies that understand this are repositioning from "we train models" to "we produce the substance models are made of." The ones that do not are going to find themselves paying tolls on a road they never realized they were driving on.

The headline is $3.5 billion. The story is that the industry's raw material — unglamorous, unphotogenic, labor-intensive — has quietly become the most valuable substance in the field, and the people who understood that first are the ones now being paid like it. That is not a prediction. It is a description of what already happened, and the market just wrote down the price.

Sources:

• TechCrunch — "Snorkel AI triples valuation to $3.5B as demand for AI training data booms" (September 22, 2026): https://techcrunch.com/2026/09/22/snorkel-ai-triples-valuation-to-3-5b-as-demand-for-ai-training-data-booms/