Data Markets · Essay
The best teams don't collect data. They manufacture it — and the factory gets bigger every year.
I joined Google in 2007. Not long after, I got a cappuccino in a micro kitchen with Hal Varian, Google's chief economist and the author of a key textbook I studied in graduate school at Princeton. I asked him why he had left academia.
“A simple reason,” he said. “The best tech companies create the best data. And data is precisely what you need for the best answers.”
That discussion stuck with me. The word that matters is create. Not have. Not collect. Create.
Varian himself later published the strongest case against the thesis people now build on top of that quote: in “Artificial Intelligence, Economics, and Industrial Organization” he wrote that “data typically exhibits decreasing returns to scale like any other factor of production,” and that what circulates as “data network effects” is really old-fashioned learning by doing — “certainly not a network effect!” He is right, and it relocates the thesis rather than weakening it, because volume has decreasing returns and capability does not. The same paper says so plainly: “the problem is not lack of resources, but is lack of skills.” The honest version of the compounding claim is also the narrower one — Andrei Hagiu and Julian Wright show that learning across many customers' data, not repeat use by one, is the kind that behaves like a network effect, and that takes a team good enough to relate one customer's records to everybody else's.
So the asset is not the pile of data. It is the plant that makes the pile and keeps improving it. I have come to call that a data factory, and I think it is the most important company shape of the next decade.
What reCAPTCHA Actually Was
In 2008, Luis von Ahn and his colleagues at Carnegie Mellon published reCAPTCHA in Science. Words that OCR had failed to read in scanned books were served as security challenges, and humans proved they were human by transcribing them. In roughly the first year, humans solved more than 1.2 billion CAPTCHAs, deciphering over 440 million words — “equivalent to over 17,600 books manually transcribed” — at 99.1% word accuracy against 83.5% for the OCR it was correcting. As von Ahn put it: “It would take more than 1,500 people working 40 hours a week at a rate of 60 words a minute to match our weekly output.”
Google bought it in 2009, pointed it at Google Books, then at Street View house numbers — close to 100 million of them, the addresses that make Maps work. Same plant, new input, new product.
That was not a data-collection project. It was a manufacturing process that produced an asset Google could not have bought at any price.
Collection Is the Input, Not the Business
This is where most data pitches go wrong. Founders describe what they have access to, as though access were the asset. Access is the loading dock.
A data factory has four stages, and the first one is the cheapest. Collect — get to where the data is born: a sensor, a transaction, a signed license from an owner nobody else thought to ask. Structure — decide what the records are, and whether this entity is the same as that one. Refine — score it, for how fresh it is and how likely it is to be true. Ship — cut it for one buyer, in their schema, graded against their outcome.
The middle two are where teams fail, and they are the two that make everything else compound. Add a related record to a well-structured graph and you have not just added a record — you have resolved ambiguity in the records you already held.
Oil in a tank does not make the oil beside it more valuable. A record in a graph makes every record it touches more legible.
Collecting data nobody else can reach, and refining it in a way nobody else can, is the pair that wins. One without the other is a commodity.
Data Volume
The pro-volume case is real. In 2001, Michele Banko and Eric Brill scaled a spelling-disambiguation task from one million training words to a billion and watched all four algorithms they tested climb from roughly 75% accuracy to 98 or 99%, with the ranking inverting along the way — the worst learner with a lot of data beat the best learner with a little. Google researchers called it the unreasonable effectiveness of data. But the counterexamples now come from the same labs: Microsoft's phi-1 reached 50.6% on HumanEval with 1.3 billion parameters on about 7 billion deliberately curated tokens, and Meta's LIMA, fine-tuned on 1,000 chosen examples with no reinforcement learning from human feedback, still had human raters prefer or tie it against GPT-4 in 43% of comparisons. The variable was not volume. It was the judgment of the people who picked the examples.
The cautionary case is Google's own. Google Flu Trends had proprietary search data at a scale no public health agency could match. It missed the 2009 H1N1 pandemic entirely, then overestimated flu prevalence in 100 out of 108 weeks. Lazer, Kennedy, King and Vespignani named the failure in Science: big data hubris. Nobody had structured the data or asked what the signal meant. They had correlated it.
The Data Factory Map
Cross the two variables and you get the frame I now use in diligence. On one axis, how much data. On the other, how good the team is at structuring and refining it.
| Quadrant | Volume × Skill | What it means for capital |
|---|---|---|
| The Hobby Bench | A little · Low skill | Nothing compounds. Resold access, priced as a commodity. |
| The Data Lake | Vast · Low skill | Varian's case exactly: huge amounts of data, nothing done with it. |
| The Boutique | A little · High skill | Beats much larger corpora. Real, but hard to turn into scale. |
| The Data Factory | Vast · High skill | Proprietary collection plus proprietary refining. Where we underwrite. |
The useful thing about the map is which axis moves faster. Volume gets bought, licensed, scraped and eventually matched. Skill gets hired, retained and compounded. Vertical movement is the durable one.
Why We Invested in Nyne AI
Nyne AI describes itself as a data factory. That is a claim about process rather than inventory, which is what got my attention.
The graph is roughly 4,000 data features across 2.2 billion people and 380 million businesses. The size is not the interesting part. The interesting part is that every element carries two scores: how recent it is, and how likely it is to be true. Most data companies I meet cannot tell me the freshness of a given field. Nyne scores it by default.
Then there is the second line. Nyne uses the graph to find long-tail businesses, licenses their operational data with consent, de-identifies it while preserving the relationships inside the records, and delivers it to frontier labs as training and reinforcement-learning material. The graph that is the product in one business is the discovery mechanism for the other — the plant's output became the plant's input. By our estimate, labs will spend on the order of $8.5 billion on third-party data in 2026. For Nyne's data partnerships, clear licensing, permitted use, and careful de-identification are part of the product specification.
The team. I have known Michael Fanous, the CEO, for about a year through the founder community I run, and he is in the top 1% of founders I meet — fast, decisive, and straight with me about what is not working. His father Emad, the CTO, has twenty-five years in people data, including as CTO of a public internet holding company. This is a team that has structured messy entity data at scale, through a full regulatory cycle. I have also watched them win head-to-head technical evaluations against incumbents — the only review that counts in this business, because the buyer tests the records against reality instead of reading the deck.
The risks are legible. Revenue concentration is the live one; this category has a history of suppliers losing a dominant customer with a weekend's notice. Privacy regulation in California and the EU is a standing exposure for anyone holding a consumer graph. And discovery may not be the bottleneck — the hard parts are legal clearance of third-party data inside someone's ERP, and de-identification that survives an audit.
So here is the test I will be watching, and the one that would prove me wrong: does gross margin rise as token prices fall? If the factory is real, cheaper intelligence makes it more profitable — agents do the sourcing and verifying while the price of verified quality holds. If it is a reseller, cheaper intelligence lets the buyer do the job itself.