AI labs are buying data from dead startups to train their models

A growing industry of startup 'undertakers' is selling code, emails and Slack histories to foundation model labs

By LineZotpaper

Published13 October 2026

Read Time3 min

Faced with shrinking supplies of public and synthetic training data, artificial intelligence companies are increasingly buying the digital remains of failed startups, according to MIT Technology Review. Brokers including SimpleClosure and Sunset are selling dead companies' code bases, emails, Slack histories and other workspace archives to foundation model labs, a trade one executive likened to turning wood shavings into value.

When Keelan Miskell began winding down his crypto startup this year, he did not think its data was worth much. The users were anonymous and the company was only a few years old. He has since turned the firm's digital archives into real money, netting roughly $25,000 for its code base and a still-pending amount of around $15,000 for workspace data held in Google, Notion, Figma and Slack. 'At least all that stuff we were doing for two and a half years is worth something to someone somewhere,' he said.

The sale was brokered by SimpleClosure, one of a growing number of companies catering to dying startups. Estimates from Carta and AngelList show startup shutdowns rising between 26% and 56% annually in recent years, and an end-of-life ecosystem of coaches, 'death doulas' and firms that handle the legal, tax and operational work of closing down has emerged alongside it. SimpleClosure added data sales to its services in April. Its Asset Hub platform helps founders sell code and workspace data, with buyers including foundation AI labs, vendors that build simulated workspaces for AI training, and mature tech companies developing their own models.

Asset Hub has handled more than 150 deals and expects to return over $10 million to startups by the end of 2026. Chief executive Dori Yona said transactions that were once rare and 'somewhat taboo' have become increasingly routine, with 47% of SimpleClosure deals now including code or data sales. Rival Sunset has also built a rapidly growing data resale business. Founder and chief executive Brendan Mahony said the value of a dying client's assets is measured by head count, years in operation, data volume and how the team used its tools. 'This operational exhaust data is like the wood shavings that everyone thought had no value, but it actually has a ton,' he said.

Demand emerged suddenly last year as training labs hit the limits of public and synthetic datasets. Surge AI buys archives to build 'reinforcement-learning gyms' where clients' AI agents practise tasks such as debugging complex IT architecture or drafting appellate briefs. Nick Heiner, who leads the company's work on these environments, said he seeks archives that capture a range of situations, from disagreements to tricky judgment calls. One company's Slack history, he said, can be worth more than another's Slack, email and calendars combined if it captures conversations with variety and depth.

Real archives avoid common errors in synthetic worlds, such as references to dates that do not exist, but the extent of the performance gains remains unclear. No company has published tests comparing real-world data training with the alternatives, and pricing reflects the uncertainty. Buyers quote wildly different numbers, Mr Heiner said, because valuations rest on 'research theses' estimating how many useful tasks come from, say, 100,000 lines of code. Bigger deals are already happening. In August, Google outbid the AI data company Mercor to buy Spirit Airlines' business data for $10 million during the carrier's bankruptcy.

Analysis

Why This Matters

AI training data is becoming a scarce commercial asset, and the archives of failed companies are a new source of varied, real-world material that synthetic datasets cannot easily replicate.
Founders winding down startups can now recover meaningful cash from records they once considered worthless, changing the endgame economics of a failed venture.
The trade raises unresolved questions about how the operational debris of one company gets valued, priced and eventually absorbed into someone else's AI.

Background

AI models have historically been trained on enormous volumes of text and code drawn from the public web. As that pool has become harder to use freely, labs have turned to licensed content and to synthetic data generated by other models, which has known weaknesses such as reproducing narrow patterns and generating impossible details. The sharp rise in startup shutdowns in recent years has produced a new pool of digital archives, and a niche industry has formed to route those archives toward model training. What began as a rare and slightly taboo transaction is becoming a standard part of the closing-down process.

Key Perspectives

Founders of failed startups: The chance to sell code and workspace archives offers an unexpected final return. Keelan Miskell's experience suggests even small, anonymous startups can recover thousands of dollars, though outcomes vary widely. AI labs and data vendors: Buyers such as Surge AI treat real archives as valuable source material for training agents to handle realistic tasks, arguing that real records capture disagreements, judgment calls and operational context that synthetic data misses. Brokers: SimpleClosure and Sunset present themselves as helping dying companies extract value from overlooked assets, with the trade becoming routine: 47% of SimpleClosure deals now include code or data sales, and one rival describes the material as wood shavings with hidden worth. Skeptics: There is no published evidence yet that real-world archives outperform synthetic alternatives. Because valuations are driven by 'research theses' rather than measured results, pricing remains inconsistent, and buyers quote wildly different figures for similar data.

What to Watch

Whether any AI lab publishes benchmark results comparing models trained on real archives with those trained on synthetic data.
The growth of SimpleClosure's Asset Hub and its promise to return over $10 million to startups by the end of 2026.
Larger bankruptcy asset sales, such as Google's $10 million purchase of Spirit Airlines' data, which could set pricing benchmarks for corporate archives.
Possible scrutiny of how data belonging to defunct companies' users is handled once it enters AI training pipelines.

Sources

AI labs are buying data from dead startups to train their models — MIT Technology Review