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

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

Most startups fail. Now, a growing industry is giving their digital remains an afterlife—as training data for AI models.

When Keelan Miskell began winding down his crypto startup this year, he assumed its anonymous users and short history meant its data was worthless. Instead, he netted roughly $25,000 for the codebase, with another $15,000 still pending for workspace data from Google, Notion, Figma and Slack. "At least all that stuff we were doing for two and a half years is worth something to someone somewhere," he told MIT Technology Review.

That sale was brokered by SimpleClosure, one of several companies now specialising in the end-of-life of startups. With shutdowns rising between 26% and 56% annually according to Carta and AngelList, an ecosystem of coaches, "death doulas" and closure firms has emerged. Many of them are now adding a new line of business: selling defunct companies' emails, chats, documents and source code to AI labs.

When shutdown data becomes training fuel

SimpleClosure launched its Asset Hub data sales platform in April. It has handled over 150 deals so far and expects to return more than $10 million to startups by the end of 2026. CEO Dori Yona says deals that were once "somewhat taboo" are now routine, with 47% of SimpleClosure deals including code or data sales. Rival firm Sunset has seen similar growth.

The demand surged last year as foundation model labs hit the limits of public web data and ran into the well-known weaknesses of purely synthetic datasets. As Nick Heiner of Surge AI explains, the company buys archives to build "reinforcement-learning gyms" where AI agents can practice realistic tasks like debugging complex IT architectures or drafting legal briefs. Real workplace conversations capture disagreements, judgment calls and messy context that synthetic worlds often miss—including basic errors like references to dates that don't exist.

Yet the performance gains remain unproven. No company has published head-to-head tests comparing models trained on real shutdown archives against those trained on synthetic alternatives. Instead, pricing is driven by research theses. "Buyers quote wildly different numbers," Heiner says, because valuations rest on estimates of how many useful training tasks can be extracted from, say, 100,000 lines of code.

Big money for big archives

That uncertainty hasn't stopped seven-figure deals. In August, as part of Spirit Airlines' bankruptcy, Google outbid AI data firm Mercor to purchase the carrier's decades of data—including 100 million emails, 500 million Teams messages and 176,000 employee records—for $10 million. Bloomberg also reported in September that SpaceX's AI unit was weighing buying data from struggling startups.

For founders, these sales can turn what looked like digital waste into a meaningful final payout. For AI labs, they're a way to acquire domain-specific, real-world interaction data at a time when licensed training corpora are increasingly competitive.

The legal grey area

The trade is raising serious privacy and consent questions. Spirit's flight attendants' union objected to the sale on privacy grounds. Attorney Greg Ewing has called such data "legally radioactive," pointing to deeply personal details buried in emails and calendars—doctor appointments, school pickups, date-night plans.

Brokers and buyers say they scrub personally identifiable information (PII) and exclude HR records before transfer. But scrubbing large, unstructured archives is far from foolproof. As Ewing notes, the industry is still in a "Wild West of AI training" phase, only slowly moving towards proper licensing agreements and vetted use cases.

There's also the question of who profits. Proceeds typically flow to senior creditors and investors, not to the employees whose conversations and work are being sold. Emory University law professor Ifeoma Ajunwa argues this is a form of "captured capital," noting that while workplace monitoring may be justified while a company is operating, that justification disappears once it goes out of business. "Employers should have express consent before selling workers' data," she says—though no federal law in the US currently bans such sales.

A minimal PII filter

For technical teams evaluating data licensing, a basic redaction pass is often the first line of defence. A simple heuristic filter can catch obvious PII like emails and phone numbers before more expensive entity recognition runs:

import re

PII_PATTERNS = [
    r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}",
    r"\b(?:\+?1[-.\s]?)?\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b",
]

def scrub_pii(text: str) -> str:
    for pattern in PII_PATTERNS:
        text = re.sub(pattern, "[REDACTED]", text)
    return text

This is nowhere near comprehensive enough for production scrubbing, but it illustrates the kind of preprocessing that often happens before any training data changes hands.

What comes next

Founders selling their archives often have little visibility into who buys the data or how it will be used, and little assurance they're getting a fair offer. Prices may stabilise over the next year—unless an AI bubble bursts first. Either way, as shutdowns continue, the "tech undertakers" appraising what's left of failed startups aren't going anywhere. The question is whether worker consent and meaningful privacy safeguards will catch up to the market that's already forming around their data.

Sources

• MIT Technology Review: AI labs are buying data from dead startups to train their models
• AI Breaking Wire: AI Labs Are Buying Data From Dead Startups

---