The Liquidators Turning Startup Corpses into Training Data

From Wind-Down to Windfall

Keelan Miskell had resigned himself to writing off two and a half years of effort when his crypto venture folded earlier this year. The user base was pseudonymous, the runway exhausted, the pivot options nil. Then a broker approached him with an unexpected proposition: sell the company's internal data. Within weeks, Miskell closed a deal worth roughly $25,000 for the code repository, with another transaction pending that should net around $15,000 for the accumulated detritus in Google Workspace, Notion, Figma, and Slack.

At Opentechwire, we've tracked the rising tempo of venture-backed shutdowns across Southeast Asia and North America for the past eighteen months, but this development marks a structural shift. The buyers lining up for these digital remains are not scrap dealers or compliance auditors. They are foundation model labs, vendors constructing simulated environments for agentic AI, and established technology companies building proprietary training corpora. The transactions are being brokered by a cluster of firms that specialise in startup euthanasia, and their order books are growing fast.

SimpleClosure, the intermediary in Miskell's case, launched its Asset Hub platform in April 2026. The service connects founders winding down their companies with organisations willing to pay for archives of code, email threads, chat logs, and collaborative documents. SimpleClosure has processed more than 150 such deals to date and expects cumulative payouts to exceed $10 million by the end of the year. Dori Yona, the company's chief executive, notes that nearly half of all SimpleClosure engagements now include a code or data component, a sharp increase from twelve months ago when such sales were considered niche and faintly distasteful.

Sunset, a rival outfit in the wind-down space, reports similar momentum. Brendan Mahony, Sunset's founder, likens the phenomenon to particleboard manufacture: what was once regarded as waste, sawdust and wood shavings with no intrinsic value, can be compressed into something structurally useful. His firm appraises defunct ventures by headcount, operational tenure, data volume, and the granularity of tool usage. A three-person team that ran lean and documented sparingly will fetch less than a fifteen-person squad that left dense trails across project management and communication platforms.

Why Labs Want the Exhaust

The appetite for startup exhaust emerged abruptly in late 2025, driven by the depletion of high-quality public datasets and mounting scepticism about synthetic alternatives. Reinforcement learning from human feedback, the technique underpinning conversational agents and code assistants, depends on examples that reflect genuine decision-making under uncertainty. Public repositories and crowdsourced annotations offer breadth but often lack the contextual messiness that characterises real work: disagreements over architectural trade-offs, ambiguous requirements, evolving priorities, and the informal knowledge embedded in chat threads.

Surge AI, a vendor building what it calls reinforcement-learning gyms, acquires archives to construct training environments where agents practise complex tasks such as debugging distributed systems or drafting legal briefs. Nick Heiner, who leads environment design at Surge, explains that a single company's Slack history can prove more valuable than a competitor's combined Slack, email, and calendar exports if the former captures a richer variety of conversational dynamics and edge cases. Authentic archives also sidestep artefacts common in synthetic data, such as references to non-existent dates or implausibly polished prose, though Heiner acknowledges that no lab has yet published controlled comparisons quantifying the performance delta.

Pricing remains volatile because valuation models rest on untested assumptions about task density. Buyers quote widely divergent figures, Heiner says, because each is operating from a research thesis that estimates how many useful training episodes can be extracted from, say, 100,000 lines of commented code or 50,000 threaded messages. The market is too young for benchmark transactions to have established floor prices.

The Big Deals and the Big Questions

Scale matters. In August 2026, as Spirit Airlines navigated bankruptcy proceedings, Google outbid Mercor, an AI data specialist, to acquire the carrier's operational archives. The package included approximately 100 million emails, 500 million Microsoft Teams messages, and 176,000 employee records spanning decades. The purchase price was $10 million. Separately, Bloomberg reported in September that SpaceX's AI division has explored acquiring data from distressed startups, though no transactions have been confirmed.

These high-profile deals have drawn scrutiny. Spirit Airlines' flight attendants' union filed an objection on privacy grounds, and legal observers are beginning to flag the risks. Greg Ewing, a solicitor advising AI ventures, has characterised such datasets as legally radioactive. Both sellers and buyers deploy data scientists to strip personally identifiable information and exclude human resources records, but Ewing argues that residual exposure remains substantial. Emails and calendar entries often contain incidental personal details: medical appointments, childcare logistics, social plans. The current market, he suggests, resembles a Wild West phase that is only beginning to transition towards formal licensing frameworks and due diligence protocols.

There is also the distributional question. Proceeds from asset sales typically flow to senior creditors and equity investors under standard liquidation waterfalls, leaving employees, the individuals who generated the data, with nothing. Ifeoma Ajunwa, a law professor at Emory University and author of The Quantified Worker, describes this arrangement as captured capital. Workplace monitoring is legally permissible when justified by operational necessity, but that rationale evaporates once a firm ceases operations. Ajunwa argues that employers should secure express consent before monetising worker-generated data, though she notes that no federal statute in the United States currently prohibits such sales outright.

Opacity and Asymmetry

Founders participating in these transactions face significant information asymmetry. They have limited visibility into the identity of buyers, the intended use cases, or the downstream licensing arrangements. The brokers provide indicative valuations, but sellers have few benchmarks against which to judge whether an offer is fair. Yona and Mahony both acknowledge this opacity but argue that founders, facing insolvency, have little negotiating leverage and that any recovery is preferable to writing off the assets entirely.

The market may mature quickly. If demand continues to grow and more transactions close, pricing conventions will likely stabilise over the next twelve to eighteen months. Alternatively, if investor appetite for generative AI cools or if regulatory intervention constrains permissible data use, the nascent trade could stall. Either outcome will keep the liquidators busy. Shutdown rates tracked by Carta and AngelList have climbed between 26 per cent and 56 per cent year-on-year in recent periods, and the pipeline of distressed ventures shows no sign of thinning.

For now, the tech sector's newest intermediaries occupy an unusual niche: part insolvency practitioner, part data broker, appraising the informational remnants of companies that failed to find product-market fit but succeeded, inadvertently, in generating training signal. Whether that signal proves durable enough to justify the legal and ethical complications remains an open question, but the market is moving faster than the answers.