AI Labs Are Buying Defunct Startups' Data to Train Models

When a startup shuts down, it often leaves behind not just a story of failure, but also a pile of seemingly useless user data, code repositories, and digital archives. But today, these "digital relics" are being treated by AI labs as valuable fuel for training models.

According to an RSS source, when Keelan Miskell began shutting down his cryptocurrency startup this year, he didn't think his company's data was worth much—users were anonymous, and the company had only been around for a few years. However, he subsequently turned the company's digital archives into real money: the code repository sold for about $25,000, and another payment is still under negotiation, potentially for a higher amount.

This is not an isolated case. A quiet trend of "acquiring defunct startup data" is emerging in the AI industry.

Why Are AI Labs Eyeing "Dead" Startups?

Training large models requires massive, diverse, real-world data. While public internet data is vast, its quality is uneven and it faces copyright and privacy disputes. In contrast, a startup that has operated for a few years often holds internal data with unique value:

• User behavior data: Even when anonymized, users' clicks, dwell time, and conversion paths can reflect real demand patterns.
• Code repositories and engineering docs: For code generation models, code from real production environments is excellent fine-tuning material.
• Industry vertical data: Startups in crypto, healthcare, finance, and other sectors carry strong domain-specific characteristics in their data, which can help models perform better on specific tasks.

This data is typically not on the public web—it belongs to "private corpora," a scarce resource for AI labs seeking differentiated capabilities.

"Private corpora" is a concept relative to public internet data. Public corpora include web pages, Wikipedia, open-source code repositories, and other content that anyone can scrape, while private corpora refer to non-public data such as internal user logs, customer service records, proprietary code, and internal documents. The core value of private corpora lies in scarcity and distributional difference: public internet text has been repeatedly trained on by mainstream large models, with diminishing marginal returns; private corpora reflect real language and behavior patterns in specific business scenarios, helping models gain differentiated capabilities on vertical tasks. For this reason, private corpora are often obtained through licensing, partnerships, or acquisitions rather than scraping. Defunct startup data falls exactly into this category—it once operated in real production environments and possesses distributional characteristics that public data cannot replace.

How Do These Transactions Work?

Based on Miskell's case, the process is not complicated: after a startup founder decides to close the company, they sell data assets through intermediaries or by directly contacting AI labs and data brokers. The price depends on data volume, uniqueness, cleaning difficulty, and potential legal risk.

Miskell's code repository sold for about $25,000, while the user data is still awaiting a buyer. This suggests a pricing logic: code repositories can be directly used for training code models, so their value is relatively clear; the value of anonymous user data depends more on buyer assessment and carries higher uncertainty.

You may not have noticed that such transactions currently sit in a legal gray area. Startup privacy policies typically promise not to sell user data to third parties, but after a company goes bankrupt, it is legally unclear whether asset liquidation covers data.

"Data brokers" refer to intermediary institutions that specialize in collecting, organizing, and reselling data. In traditional industries, data brokers mainly serve advertising and risk control, aggregating information from public records, consumption behavior, and other channels and selling it to businesses. In the AI training data market, the role of data brokers is expanding: they help AI labs find private corpora in specific domains, while also assisting data sellers in assessing compliance risks, cleaning data, and performing anonymization. Because AI labs usually lack direct channels to a large number of defunct companies, data brokers and intermediary networks become key nodes connecting supply and demand. However, this link also increases transaction opacity, making it harder to trace data provenance and ensure informed user consent.

What Does This Mean for the AI Industry?

A new vein of ore amid data exhaustion. High-quality public text data is being depleted, forcing AI labs to seek new sources. The data pools of defunct startups may become the third major data source after web scraping and open-source datasets.

A new exit path for startups. In the past, failed companies usually ended with fire sales of assets or outright deregistration. Now, data itself has become a sellable asset, offering founders some return and potentially spawning funds or intermediaries dedicated to acquiring "digital legacies."

Privacy and ethical concerns. Even anonymized user data can be re-identified through cross-referencing. If AI models "memorize" this data during training, it could trigger new privacy leakage risks. Regulators have not yet issued clear rules for such transactions.

"Data exhaustion" has been a core issue in the AI field in recent years. Multiple studies estimate that high-quality public text data available for training large models will be consumed within a few years—the growth rate of model scale far exceeds the rate at which new public data is generated. This is forcing AI labs to turn to multimodal data, synthetic data, and private data. Synthetic data is generated by models themselves, but carries risks of model collapse and amplified bias; private data, while high-quality and strongly domain-specific, faces bottlenecks such as high acquisition costs and complex compliance reviews. The data pools of defunct startups are therefore seen as a compromise path: they have real-world distributional characteristics, yet unlike public data, they have not been fully utilized. But whether they can scale depends on the maturity of legal frameworks and transaction infrastructure.

Risks and Controversies

From Miskell's case, the transaction amounts are not astonishing—$25,000 may be a drop in the bucket for a company that operated for a few years. But the signal matters more than the amount itself: the liquidity of data as an asset is increasing.

On the other hand, when AI labs buy this data, they must face due diligence challenges. Is the data source legal? Do users know? Will model outputs leak sensitive information? There are no standard answers to these questions.

Currently, such transactions are still conducted privately and non-publicly, lacking transparency. But as AI data demand continues to swell, the market for defunct startup data may gradually surface, possibly even forming specialized trading platforms.

For entrepreneurs, this may be a reminder: even if the company fails, the data may have a second life. But before selling, be sure to clarify legal boundaries to avoid hidden risks.

The "anonymization paradox" refers to the fact that even after data is anonymized—removing direct identifiers such as names and email addresses—individual identities may still be re-identified through cross-referencing and auxiliary information. Research shows that with just a small amount of additional external data (such as public social profiles or location information), anonymous records can be linked to specific individuals. For large models, the risk goes further: a model may "memorize" certain training samples during training and reproduce them verbatim during generation, causing originally anonymous information to be accidentally leaked. This means that even if a startup anonymizes data before selling, downstream models can still become a channel for privacy leakage. Currently, techniques such as differential privacy and federated learning have been proposed to mitigate this problem, but they have not yet become standard practice in actual transactions.

One-line Summary

AI labs are acquiring defunct startups' code repositories and anonymous user data as a new source of private corpora for large model training.

Paragraph Summary

As high-quality public text data gradually dries up, AI labs are turning their attention to the "digital relics" left by defunct startups—code repositories, user behavior data, and industry vertical data. One cryptocurrency startup founder earned about $25,000 by selling a code repository, while user data is still for sale. Such transactions sit in a legal gray area: whether privacy policies cover data asset liquidation, whether anonymous data can be re-identified, and whether models will leak sensitive information all lack clear regulation. For the AI industry, this may be the third major data source after web scraping and open-source datasets; for entrepreneurs, failed companies' data gains a second chance at monetization, but legal boundaries must be carefully clarified.