Why AI Labs Are Buying Failed Startups’ Data Archives

👁 113 reads

A crypto startup founder walked away with $25,000 for his source code and expects another $15,000 for his company’s archived Google Workspace, Notion, Figma, and Slack data - all after deciding to shut down the business. The experience of Keelan Miskell highlights a fast-growing secondary market emerging from tech bankruptcies: AI labs are paying cold cash for the digital remains of failed startups.

Startup liquidation is becoming a data transaction

Data from Carta and AngelList reveals startup shutdown rates climbing between 26% and 56% year over year. A cottage industry of wind-down services has sprung up around this wave, ranging from closure coaches and “death doulas” to legal and tax dissolution specialists. Now, specialized data brokers are connecting defunct companies directly with artificial intelligence labs.

One key player is SimpleClosure, which rolled out its Asset Hub marketplace in April. The platform allows founders to list proprietary codebase repositories alongside workspace archives - including emails, direct messages, project specs, and meeting notes. Buyers include foundation model labs, teams building synthetic training environments for autonomous agents, and established tech corporations training proprietary models. Asset Hub has handled over 150 transactions to date. According to CEO Dori Yona, the platform expects to return over $10 million directly to winding-down startups by the end of 2026.

Yona notes that such asset sales were once rare and taboo. Today, data sales represent 47% of all liquidation cases handled by SimpleClosure - a structural change in how founders exit.

Competitors are moving in quickly. Sunset, a direct rival to SimpleClosure, runs its own expanding data brokerage arm. Brendan Mahony, founder and CEO of Sunset, values assets based on team headcount, active operating years, data volume, and internal communication workflows. He framed the opportunity with a clear analogy:

“These operational data points are like sawdust that everyone assumed was worthless - yet it turns out to be immensely valuable.”

Behind that analogy lies a clear technical bottleneck facing the AI industry.

Why AI labs need corporate Slack channels and repos

The sudden surge in demand for corporate post-mortem data took off last year as frontier AI labs hit the ceiling of public web scrapes and synthetic generation. Most accessible public web content has already been consumed, while synthetic data has clear failure modes - models trained heavily on generated text produce recurring hallucinations, such as referencing non-existent calendar dates or fabricating procedural steps.

Real corporate archives provide something synthetic data cannot: authentic work sequences, messy collaborative decision-making, and documented trade-offs. Nick Heiner of Surge AI - a firm acquiring datasets to build reinforcement-learning gyms where AI agents practice complex skills ranging from debugging cloud infrastructure to drafting legal motions - points out that an active Slack workspace with nuanced technical back-and-forth can be worth far more than another company’s combined email, calendar, and chat exports.

Valuation remains speculative. Heiner readily acknowledges that buyers throw out wild figures because pricing hinges on speculative research theses - estimates of how many discrete training episodes can be extracted from 100,000 lines of proprietary code. Because no provider has published public benchmarks comparing models trained on bankrupt startup logs against purely synthetic baselines, the market is expanding on intuition rather than standardized metrics.

Google vs. Mercor: The $10 million Spirit Airlines auction

The clearest sign that enterprise data values are escalating arrived in August. During the Spirit Airlines bankruptcy auction, Google outbid AI data provider Mercor to secure decades of airline operational archives for $10 million. The package covered 100 million emails, 500 million Microsoft Teams messages, and 176,000 employee records. Bloomberg reported in September on similar aggressive bids by major technology companies, signaling that operational corporate data harvesting is becoming an established strategy.

The Spirit Airlines acquisition illustrates two shifts. First, operational logs from large operations - capturing live troubleshooting, crisis handling, and bureaucratic approvals - offer far richer agent training material than simulated datasets. Second, dedicated AI data firms like Mercor are now bidding directly against hyperscalers for distressed data assets.

Employee privacy and the legal fallout

For shutting-down founders, data offloading redefines the narrative: closure shifts from pure failure into a partial liquidity event. As Keelan Miskell pointed out, having someone pay for the archive means the previous two and a half years of work still retained measurable market value.

Yet the legal mechanics remain murky. Workspace exports from Slack, Figma, and Notion contain far more than founder thoughts: they capture sensitive employee correspondence, customer communications, confidential performance reviews, and internal disputes. Whether liquidating estates have the clear legal right to transfer these personal and business communications to commercial model developers - particularly under frameworks like the EU’s GDPR - remains contentious. None of the major liquidation platforms have offered a comprehensive public breakdown of their compliance frameworks.

As startup failure rates stay elevated, more corporate communication archives will end up on the bidding block, pushing data privacy regulators to examine the trade directly.

// read moreAI Data Startup Micro1 Hits $500M Gross Run Rate in Eight MonthsMagazineFive Times AI Failed Me (And Why I Still Won’t Turn It Off)MagazineWorld Labs Unveils Atlas: Fei-Fei Li’s 3D World ModelNews

Kinetyra