Data Marketplaces & Intermediaries

AI labs are buying data from startups that shut down — and it reframes what counts as a valuable corpus

Source: MIT Technology Review · Oct 13, 2026

One of the more telling developments of the last year is quietly becoming an industry: when a startup shuts down, its data is being sold to AI labs. Wind-down services that used to help founders liquidate laptops and reassign trademarks now broker something far more valuable — the company's accumulated operational exhaust. Internal chat histories, project tickets, work email, source code, the messy record of a real team solving real problems. One platform in this market reports well over a hundred deals, with a rising share now including code or data rather than just conventional assets.

The reason labs are reaching for this material is the same constraint showing up everywhere: the easy data is exhausted. The open web has largely been scraped, and synthetic data, however useful, tends to smooth out exactly the irregularities that make human data worth learning from. What is left — and what is suddenly worth paying for — is authentic records of people actually working: the dead ends, the corrections, the domain-specific judgment that nobody writes down for a public audience.

That reframes what a valuable corpus even is. It is not the polished, published artifact; it is the process behind it. The reasoning, the back-and-forth, the decisions made under real constraints. This is the same premise that makes interview and assessment data valuable, and from a cleaner starting point: an interview is authentic human problem-solving that was captured deliberately, with consent, in a structured setting — not reconstructed after a company folded, and not entangled in the question of whether the people who generated it ever agreed to its reuse.

The market is, in other words, going to unusual lengths to source genuine human work product because the authentic article has become the scarce thing. The durable advantage belongs to data that is not only authentic but consented and provenance-documented from the start — which is the distinction between scavenging for signal and owning it.

Key Points

• A wind-down-services market has emerged that resells failed startups' real operational data — internal chats, tickets, work email, code — to AI labs
• One platform reports 150+ deals, with a growing share now including code or data rather than just physical and IP assets
• Labs are reaching for this material because public and synthetic data are running out of the signal they need
• The common thread is authenticity: records of real people doing real work, which is exactly what scraped or generated text cannot replicate