Share with your CDO
A quiet secondary market for defunct startup data has matured fast enough that Google paid $10 million in August for Spirit Airlines’ bankruptcy estate, netting 100 million emails and 500 million Teams messages. Platforms like SimpleClosure and Sunset now broker these deals at scale, with 47% of SimpleClosure’s wind-down transactions including code or data sales. Buyers range from foundation model labs to companies building proprietary models, all chasing the operational texture that public datasets and synthetic generation can’t replicate.
What this means for your business
Whether your organization is a buyer, a seller, or simply a former employer of people whose Slack history is now on the market, this story is already about you. The CDO whose team is evaluating training data vendors needs to ask a harder question than “Is the PII scrubbed?” The answer from every seller is yes. The more honest question is whether your legal team has reviewed what “scrubbed” actually means when employee calendars contain doctor appointments and school pickups, and whether your indemnification clause covers the answer.
The pricing chaos here is a signal, not just a negotiating inconvenience. Buyers are quoting wildly different figures because nobody has published a controlled comparison of real-world operational data against synthetic alternatives. That means every acquisition is essentially a research bet dressed up as a procurement decision. Organizations spending on this data are buying a hypothesis, and the hypothesis hasn’t been stress-tested publicly. For a CDO defending a training data budget to a CFO, that’s an uncomfortable position, because the ROI argument rests entirely on vendor-supplied logic about how many useful AI tasks emerge from a given data archive.
The deeper structural problem is the consent gap. Legal precedent in most jurisdictions permitted employers to collect and monitor worker communications to support business operations. Once the business is gone, that justification disappears, as Emory law professor Ifeoma Ajunwa points out, but no federal law currently bars the sale. That regulatory vacuum won’t last. The Spirit Airlines flight attendants’ union objection is the early indicator: labor groups now have a concrete grievance and a named buyer. CDOs building data acquisition strategies around this market should assume the legal environment in 2027 looks materially different from today’s, and architect their data provenance records accordingly. The organizations that will get hurt aren’t the ones who skipped the market entirely, they’re the ones who bought aggressively and documented nothing.
Concept deep-dive: Reinforcement-learning gyms
A reinforcement-learning gym is a simulated environment where an AI agent practices tasks repeatedly, receiving feedback on its performance, much like a flight simulator lets a pilot train without a real aircraft. The value of operational data from defunct companies is that it provides realistic, messy scenarios, genuine IT architecture disputes, actual legal edge cases, that pure simulation tends to flatten into predictable patterns. Richer gyms produce agents that handle real-world complexity better, which is why a single company’s deep Slack archive can outvalue a broader but shallower data set.
Based on reporting from AI labs are buying data from dead startups to train their models, originally published 2026-10-13 05:00:00.
