By This Hour AI Development Desk
Artificial-intelligence labs are reportedly buying data from startups that have failed or are being wound down, opening a possible new afterlife for the digital records left behind when young companies close. The reported practice matters because it could turn assets once regarded as operational residue—code, archives and other material accumulated by an unsuccessful venture—into inputs for training models.
The account, published by MIT Technology Review, portrays the transactions as a response to the continuing demand for training material. It also raises questions that extend beyond the price of a defunct company’s remaining assets: what, precisely, is being sold; whether sellers have the right to transfer it; how buyers distinguish code from associated data; and what obligations survive when a company’s commercial life ends.
The available reporting provides only a narrow view of the activity. It says AI labs are purchasing data from failed or closing startups, but does not, in the material available here, identify the laboratories involved, detail the kinds of data they acquired, or set out the contractual terms governing the sales. Those gaps are central to assessing both the scale of the practice and its consequences.
A startup wind-down becomes a saleable archive
One example in the report concerns a crypto startup founded by Keelan Miskell that was being wound down in 2026. The company’s closure appears to have prompted a reassessment of what its remaining digital assets might be worth. Instead of treating them solely as remnants of an abandoned business, the report describes a sale of the startup’s code base.
Miskell reportedly received about $25,000 for that code base. That figure is specific, but its meaning is limited without fuller information about the company, the buyer, the scope of the assets transferred or whether the amount represented a single transaction or part of a wider process. The supplied account does not establish whether the code was itself used to train a model, whether it came with other materials, or whether it was acquired for analysis, development or another purpose.
Still, the example captures the commercial logic implied by the report. A startup can cease operating while retaining technical work produced during its attempt to build a product. For an AI developer seeking material relevant to programming or other model capabilities, that work may retain value even after the original business has failed. The closure of the company and the usefulness of its remaining digital assets are therefore not necessarily the same question.
That distinction is particularly important in reporting that uses the broad term “data.” A code base is a defined category of technical material, while a startup’s wider archive might contain many different kinds of records. The available account does not say which categories of material were included in the Miskell sale, nor does it show that all startup assets in such transactions are handled alike. Treating every digital record as interchangeable would go beyond what has been reported.
Why the reported market could matter to model development
If the reported purchases are occurring as described, they suggest that startup failures may be becoming one source of supply for AI development. Training a model depends on the material a developer is able to obtain and use. A closing startup may hold a compact but potentially specialized collection of software and internal outputs created in the course of its work. That can be commercially attractive even when the startup did not succeed as an independent business.
The report does not establish how important these acquisitions are compared with other sources of training material. It does not quantify how many startups have sold assets, how many AI labs have bought them, how frequently deals occur or how much money is involved beyond the approximately $25,000 reported in the Miskell example. There is no basis in the supplied information for calling this a broad industry standard or a dominant route to obtaining training data.
Yet even a small market could alter incentives around shutdowns. Founders, investors and administrators responsible for winding down a company may view repositories, technical documentation and related archives not just as costs to preserve or discard, but as assets that may be sold. For potential buyers, the attraction may lie in obtaining material assembled around a specific technical effort rather than building a comparable collection from the start.
Whether that produces a durable business line depends on factors the report does not resolve. A buyer’s interest will vary with the nature and quality of the asset. So will the seller’s ability to offer clear rights to it. A code base may be valuable as code without necessarily carrying a right to transfer every record, contribution or attachment that sat alongside it. The public description available here does not explain how those distinctions were negotiated in the reported transactions.
Ownership questions follow the assets
The central uncertainty is not simply whether a shuttered startup possesses material that an AI lab wants. It is whether the company can lawfully and responsibly transfer each component for the buyer’s intended use. A startup’s digital holdings can be described as a single archive, but the rights and restrictions attached to its contents may differ. The supplied reporting does not provide agreements, permissions, due-diligence records or buyer statements that would clarify those issues.
That absence makes it impossible to determine from the available material how the reported buyers addressed use restrictions, confidentiality, or any other constraints that could attach to material held by a company being wound down. Nor does the account identify a particular dataset or explain whether it contained information tied to users, customers, employees, contractors or other parties. The report’s broad framing should not be read as proof that any particular category of sensitive or personal information was transferred.
For the same reason, the Miskell transaction should not be treated as a complete illustration of the reported market. The known detail is that a crypto startup he founded was closing and that its code base was reportedly sold for roughly $25,000. It does not establish the treatment of the company’s other digital materials, the identity of the purchaser, or the purchaser’s eventual use of the code.
These distinctions will shape how consequential the practice proves to be. A sale involving technical code under clear ownership conditions presents one set of questions. A sale involving a broader archive, uncertain provenance or mixed permissions would present another. The available account does not allow a comparison between those scenarios, and it would be misleading to infer one from the other.
Evidence is thin beyond the reported transactions
MIT Technology Review’s report is the sole source supplied for this account. Its central assertion is that AI labs are buying data from failed or winding-down startups for model training. The publication also supplies the example of Miskell’s company and the approximate value attributed to its code-base sale. Taken together, those details point to a possible emerging resale market for digital startup assets.
They do not, however, establish the market’s size, regularity or future direction. No AI lab is named in the supplied claims. No buyer has publicly confirmed a purchase. No startup beyond the company founded by Miskell is identified. There is also no independent account provided here that confirms the reported pattern, no description of a standard acquisition process, and no evidence that the reported arrangements will be repeated.
The lack of detail has a practical consequence for readers evaluating the story. The idea that dormant startup material could be repurposed for model development is plausible within the reported account, but the operational specifics remain unknown. It is unclear whether buyers are primarily seeking code, other forms of data, or a combination of assets; whether they want material for direct training or for adjacent development work; and whether sellers see these deals as exceptional recoveries or a routine part of winding down.
It is also unclear what will happen to the assets once acquired. The available reporting does not describe model releases, product changes or technical results tied to the purchases. It therefore does not support claims that the transactions have improved particular AI systems or changed how any named lab develops models.
For now, the report’s strongest supported conclusion is a narrower one: some AI labs are said to be purchasing material from failed or closing startups, and one founder is reported to have sold a code base for about $25,000 while winding down a crypto venture. That may signal a new outlet for assets from unsuccessful companies, but the breadth and rules of that outlet have not been established. This report has not been independently corroborated.
For further context on this subject, see Pudgy Penguins-Backed Abstract Set to Shut Down After Reported Heavy Losses.
Reporting notes
What is confirmed: The supplied report describes purchases by AI labs and one reported code-base sale. It does not identify the labs or deal terms.
Why this matters: The reported practice could give failed startups’ digital assets a resale value while raising unanswered questions about what is transferred and on what terms.
What remains unclear: The scale of the activity, the buyers, the assets involved, their eventual use and the applicable rights are not established by the supplied material. This report is based on one source and has not been independently corroborated.
