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

TEXT START: Most startups fail.

The Dissection

The article is an autopsy of a new salvage industry, although it calls it an “afterlife.” Dead startups are stripped of code and internal records, then sold to AI labs, agent-training vendors, and incumbent tech firms. What is being bought is compressed evidence of how humans coordinate, argue, decide, debug, schedule, and work around exceptions.

It also exposes the distributional mechanism. Founders and closure firms convert organizational remains into saleable assets. Senior creditors and investors collect. Employees do not. Their private traces become captured capital after the institutional justification for collecting them has disappeared. “Tech undertakers” are building a carcass-processing layer for AI capital.

The Core Fallacy

The framing treats recovered asset value as a form of recovery. It is not. A code sale or $10 million archive purchase can monetize the corpse; it cannot restore wages, jobs, bargaining power, or productive participation.

Legal uncertainty, pricing disagreement, weak benchmarks, and a possible AI bubble are friction and lag—not refutations. The decisive question is whether real operational archives improve machine performance enough to replace cognitive labor. The article says that proof is not yet published. That limits present certainty; it does not eliminate the mechanism.

Under DT logic, the data is not a moat protecting human work. It is feedstock for agents intended to perform that work.

Hidden Assumptions

• Anonymization, scrubbing, and licensing can solve the ownership problem. They may reduce legal exposure; they do not give workers meaningful control over capital extracted from their labor.
• Value reaching creditors or investors counts as value preserved for the broader economy. This confuses liquidation proceeds with continuation of the employment-consumption circuit.
• Data scarcity and uncertainty are durable bottlenecks. They are more plausibly temporary constraints while buyers learn which archives produce useful tasks.
• Organizational knowledge can be detached from its creators without changing the strategic significance of the transaction. That detachment is precisely the automation objective.
• This is merely a niche resale market. The reported 47% deal inclusion rate and expanding buyer pool indicate an emerging supply chain.
• Better regulation can domesticate the process. Regulation can slow extraction or redistribute proceeds. It cannot make human-only cognitive production structurally competitive if P1 and P2 hold.

Social Function

Partial truth, transition management, and ideological anesthetic.

The article accurately records legal ambiguity, uncertain valuations, and concentration of proceeds. Its weakness is the “afterlife” frame: it turns the conversion of human organizational life into machine-training substrate into a clever recovery story. That language makes predation sound like resilience. It directs attention toward fair pricing and privacy while under-describing the larger power transfer: human coordination is being stripped from failed institutions and handed to owners of models, compute, and distribution.

The Verdict

This is concrete evidence of the early carcass economy. It does not prove the full DT sequence: the article provides no performance evidence establishing durable AI superiority, nor does it demonstrate that human institutions have already failed to preserve human-only domains. But it shows the direction with unusual clarity.

Dead companies are becoming training mines. Their failures are harvested to improve systems that may make the surviving workforce less necessary. The market can grow while the mass-employment system decays. The corpse yields value precisely because its accumulated human activity can be detached, reproduced, and potentially used against the living.