AI labs are increasingly purchasing data from failed startups to train their models, with brokers like SimpleClosure and Sunset facilitating these deals. SimpleClosure's Asset Hub has handled over 150 deals and expects to return over $10 million to startups by the end of 2026. This reflects a growing market for 'operational exhaust data' from defunct companies.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
As AI training labs hit the limits of public and synthetic datasets, a new industry is taking shape that resells failed startups' code and workspace data to AI labs. But these deals remain full of uncertainty around legal compliance, pricing and revenue distribution.
Tech "undertakers" are selling the data wreckage of failed startups to AI labs, forming an emerging data resale industry. The supply side of this industry comes from a continuously expanding pool: Carta and AngelList estimates show that in recent years the number of startup shutdowns has risen 26% to 56% annually.
SimpleClosure is a representative of this industry. It added data sales to its services in April of this year, and its Asset Hub platform helps founders sell code and workspace data (emails, chat logs, documents). According to the company's disclosures, Asset Hub has processed more than 150 deals and expects to return more than $10 million to startups by the end of 2026. SimpleClosure CEO Dori Yona says that today 47% of SimpleClosure deals include code or data sales.
Industry competitor Sunset also has a rapidly growing data resale business. Its founder and CEO Brendan Mahony says Sunset values data by headcount, years in operation, data volume and the team's tool usage.
Demand for failed companies' data suddenly appeared last year because AI training labs hit the limits of public and synthetic datasets. This judgment comes from the reporting's description of the industry's timing.
The buyers are not only foundation model labs. Surge AI buys data to build "reinforcement learning gyms," where clients' AI agents practice tasks such as debugging complex IT architectures or drafting appellate briefs. SimpleClosure's buyers also include vendors building simulated workspaces for AI training, as well as established tech companies building their own models.
Real archival data, compared with synthetic data, can avoid errors such as citing nonexistent dates, but the size of the performance gain remains unclear. Currently no company has published a test comparing the results of training on real-world data with alternatives.
Pricing reflects this uncertainty. Buyer offers vary enormously, and valuations are based on a "research thesis" estimating how many useful tasks 100,000 lines of code can generate. This means the same piece of code or workspace data may be worth vastly different amounts in different buyers' eyes.
One concrete deal sample is founder Keelan Miskell. He received about $25,000 for his codebase, with a pending payment (possibly about $15,000) for his Google, Notion, Figma and Slack workspace data. The deal was brokered by SimpleClosure.
Larger deals have already appeared. In August of this year, as part of Spirit Airlines' bankruptcy proceedings, Google outbid AI data company Mercor to buy the airline's decades of data for $10 million, including 100 million emails, 500 million Teams messages and 176,000 employee records. In addition, Bloomberg reported in September that SpaceX's AI division is considering buying data from struggling startups.
Lawyer Greg Ewing, who advises AI startups, calls such data "legally radioactive," and points out that emails and calendars hide private details such as medical appointments, school pickup arrangements and date plans.
Spirit's flight attendants' union has opposed the data sale on privacy grounds, which leaves the Spirit deal facing legal challenges with an unclear final outcome. Vendors hire data scientists to scrub personally identifiable information and exclude HR records, and buyers also say they will scrub it themselves, but Ewing's description points to a "Wild West" phase of AI training that is beginning to give way to formal licensing agreements and vetted use.
Legal precedent allows monitoring to help employees do their work, but once a company fails, that rationale no longer exists. Emory University law professor Ifeoma Ajunwa sees this as a form of "captured capital," and argues that employers should obtain explicit consent before selling employee data, while noting that no federal law currently prohibits such sales.
The proceeds from these sales usually flow to senior creditors and investors rather than employees. This is a structural fact about revenue distribution, creating a misalignment with the subjects the data comes from (the employees).
Information on the founders' side is equally opaque. Founders can barely learn who is buying the data or what it will be used for, and it is hard for them to ensure they received a fair offer.
Data prices are expected to stabilize over the next year, but only on the premise that the AI bubble does not burst first. This expectation itself is conditional.
This article is based on a MIT Technology Review report and its extracted evidence pack. Among these, the 26% to 56% annual rise in startup shutdowns comes from Carta and AngelList estimates; the Asset Hub deal volume, SimpleClosure's expected return amount, the 47% deal share, Miskell's sale price, Sunset's valuation method, Surge AI's use, SpaceX's consideration of buying, and the views of the lawyer and law professor are all source statements or single-media reports that have not been independently verified. The Spirit deal amount and data scale, and the flight attendants' union's opposition, are facts confirmed in the reporting, but the outcome of the deal's legal challenges remains unclear. As for how much performance gain real archival data training offers compared with synthetic data, no company has published a comparative test, making it an open question.
Failed startups' data is being repriced as an AI training asset, but the performance evidence, legal framework and revenue distribution rules supporting that pricing have yet to be settled. For founders, it is a visible but hard-to-price residual value; for employees, it is "captured capital" they neither consented to nor can easily share in.
Primary report
Primary source