Dead Startups' Data Is Becoming AI Training's New Goldmine

AI model training requires massive amounts of data, and high-quality text on the public internet is gradually being "eaten clean." As a result, an unexpected data source has surfaced: startups that have already gone bankrupt or are shutting down. According to The Download, a newsletter under MIT Technology Review, AI labs are buying data from dead startups to train their models.

An Undervalued "Legacy"

One of the key figures in this story is Keelan Miskell. When he began liquidating his cryptocurrency startup, he didn't think the data the company had accumulated would be worth much. But the reality is that this seemingly unremarkable business data—user behavior records, transaction logs, internal documents, customer service conversations—may hold considerable value for AI trainers.

Data generated by startups during operations often has several characteristics: domain focus, real-world scenarios, and clear labeling. Compared with the messy text scraped from the internet, this data is "cleaner" and better reflects the real language patterns of specific industries. For companies looking to improve model performance in vertical domains, such data sources have a natural appeal.

Why AI Labs Are Eyeing "Dead Data"

Large models' hunger for data has already outpaced the supply of public corpora. High-quality human-written text is limited, and while synthetic data can make up the volume, it falls short in diversity and authenticity. Against this backdrop, the data assets of failed startups have become an overlooked stock of resources.

From the seller's perspective, data is often seen as a burden when a company shuts down—storage costs money, compliance requires attention, and deleting it feels like a waste. Selling it to AI labs can at least recover part of the cost. From the buyer's perspective, this is a shortcut to bypass competition for public data and obtain differentiated corpora.

The variable that needs attention is compliance. Startup data may contain users' personal information, and the transaction process needs to address privacy protection and data authorization issues. The report did not elaborate on specific transaction terms, but this step is undoubtedly key to whether such data circulation can scale.

Synthetic Data refers to data generated by AI models or algorithmic programs rather than directly produced by humans. Common practices include using large models to rewrite or generate conversations, instruction-response pairs, code snippets, and more. Its advantages are low cost, unlimited mass production, and the ability to targeted supplement specific capabilities; its weakness is that "models teaching models" easily leads to narrowed distributions and accumulated errors, and long-term use may cause model outputs to become homogenized and lose coverage of long-tail real-world scenarios. For this reason, the industry generally treats synthetic data as a supplement to, not a replacement for, real corpora—real human text remains the scarce foundation for training high-quality models. This also explains why AI labs are willing to pay for the real business records of failed startups—they provide real language distributions and domain details that synthetic data struggles to reproduce.

Data compliance and authorization involve at least three layers of issues in such transactions: first, personal information protection—user behavior records and customer service conversations may contain identifiable information, and anonymization or de-identification is usually required before sale; second, data ownership—whether the service agreement between the startup and its users allows transferring data to third parties often determines whether a transaction can proceed; third, legality of data sources—if the original data collection process itself had flaws, transfer cannot cure the problem. In addition, during bankruptcy liquidation, data assets may also be constrained by bankruptcy proceedings, creditor interests, and regulatory filings. Therefore, such transactions are often accompanied by due diligence, data anonymization, and contractual use-restriction clauses, and compliance costs are also a practical reason why they are difficult to scale quickly.

Interspecies Communication: Another Frontier Signal

The same issue of The Download also mentioned the topic of interspecies communication. Looking at these two items together points to the same underlying trend: AI is penetrating domains that were previously not datafied.

Interspecies communication research attempts to use AI to decode animal signals, which is essentially also a search for new data sources—except this time the data comes from non-human life forms. Whether it is the business records of failed companies or the sounds and behavioral patterns emitted by animals, AI labs are transforming "uncollected information" into trainable signals.

Interspecies communication research typically uses machine learning to perform pattern recognition on animal vocalizations, behavioral footage, and physiological signals, attempting to establish correspondences between signals and specific meanings. Representative work includes analysis of cetacean calls, primate alarm calls, and bird songs. The difficulty of such research lies in the lack of "parallel corpora"—humans cannot directly confirm the exact meaning of a sound and can only verify it indirectly through behavioral responses, contextual comparisons, and similar methods. The value of AI lies in identifying statistical regularities and structural patterns in massive recordings that humans find hard to detect. It is logically isomorphic to buying data from failed companies: both transform signals that were not previously collected in a structured way into data assets that can be used for model training.

Implications for Entrepreneurs and the Industry

For entrepreneurs, this trend offers a pragmatic reminder: a company's data assets should not be assumed to be worthless when shutting down and liquidating. Sorting out data ownership, compliance status, and potential buyers in advance may add another path to monetization at exit.

For the AI industry, data procurement reaching into failed companies shows that the boom period for public corpora is narrowing. Future data competition will be more dispersed, more segmented, and more demanding in terms of grasping compliance boundaries. Whether this model can become a stable data supply channel depends on whether transaction standards and privacy protection mechanisms can keep up.

Data assetization refers to treating data accumulated by a company during operations as an asset that can be valued, traded, and recorded on the balance sheet, rather than merely an operational byproduct. For startups, this means that during normal operations they should consciously record data sources, obtain user authorization, and maintain data dictionaries and quality documentation, because these efforts directly determine whether the data can be smoothly transferred at exit. In practice, founders can sort out a data inventory alongside financing and compliance audits, clarify which data can be traded externally and which are restricted by contracts or regulations, and retain complete evidence of the authorization chain. The value of data assets depends heavily on their scarcity and compliance cleanliness; advance planning often realizes more value than waiting until liquidation to look for buyers.

One-line Summary

AI labs are turning to buying data from failed startups to break the dilemma of depleted high-quality corpora.

Paragraph Summary

High-quality text on the public internet is gradually being exhausted, so AI labs have begun buying business data from bankrupt or shutting-down startups for model training. Such data includes user behavior records, transaction logs, internal documents, and customer service conversations, and is characterized by domain focus, real-world scenarios, and relative cleanliness, giving it differentiated value compared with scraped corpora. For sellers, data is otherwise a burden during liquidation, and selling it can recover part of the cost; for buyers, it is a shortcut around competition for public data. But transactions involve key variables such as user privacy, data ownership, and authorization compliance, which determine whether the model can scale. The article also mentions interspecies communication research as a similar signal of AI penetrating non-datafied domains, and reminds entrepreneurs to include data assets in value considerations during exit liquidation.