As that ceiling approaches, OpenAI and Anthropic have turned their attention to a previously untapped trove: internal corporate collaboration data . Records of live human interactions—emails containing negotiations, meeting transcripts with real-time decision-making, and Slack threads showing authentic workplace dynamics—offer the kind of organic conversational data that has become scarce on the scraped public web.

The Prices: From $10,000 to $300,000 per Deal

A new ecosystem of intermediary firms has sprung up to broker these sensitive transactions. Two names appear most frequently: Mercor and SimpleClosure .

• Failed and acquired startups are selling full Slack archives, email corpora, and document repositories for $10,000 to $100,000 per deal through specialized wind-down services .
• Mercor, an AI training data company, is offering up to $300,000 for Slack threads, GitHub commit histories, and meeting transcripts .
• Warmly, an AI startup recently acquired by HubSpot, received four separate offers of up to $300,000 each for its internal business communication data—meeting minutes, emails, and more—after signing the acquisition agreement, and reportedly turned them all down .

Broader market rates, reported by Reuters in 2024, offer a sense of the value per unit: approximately $0.001 per word for text, $1 to $2 per image, $2 to $4 per short video, and $100 to $300 per hour of longer film or video .

SimpleClosure, a startup wind-down firm, has processed nearly 100 such transactions over the past year, with payouts ranging from $10,000 to $100,000 per company . The firm's "Asset Hub" platform helps founders scrub personally identifiable information (PII) from the data before licensing it—though the effectiveness and consistency of that anonymization remain uncertain and a growing point of contention .

Anthropic's $1.5 Billion Bet on Books: The Project Panama Settlement

The lengths AI companies will go to secure training data were laid bare in the extraordinary case of Project Panama, Anthropic's secret internal program to build a massive book training dataset .

The project had two tracks. First, Anthropic purchased physical print books, cut off their bindings, and destructively scanned them page by page, converting them into searchable PDFs and discarding the paper originals. The goal: convert up to 2 million books in six months . An internal memo advised using a codename "because we don't want it to be known that we are working on this" .

Second, the company downloaded more than 7 million pirated ebook files from shadow libraries including LibGen and Pirate Library Mirror . That track led to a class action lawsuit filed in 2024 by authors Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber, who accused Anthropic of using nearly half a million pirated books to train Claude .

On July 20, 2026, a federal judge in San Francisco approved a $1.5 billion settlement—the largest known copyright payout in U.S. history . More than 500,000 authors and publishers were part of the class action, and they will receive an estimated $3,000 per eligible work .

Crucially, the court drew a legal distinction that has major implications for AI training. Using pirated ebooks was infringing and triggered the settlement. But buying physical print books and destructively scanning them internally for training was ruled to potentially qualify as fair use, because each physical copy was replaced with a single digital copy in a process the court called "exceedingly transformative" . The $1.5 billion covered the pirated book liability; the physical destruction program was not penalized .

The Emerging Challenges: Anonymization and Cost

As companies rush to monetize internal data, significant questions remain. The effectiveness of data anonymization by brokers like SimpleClosure is uncertain and a growing point of contention . Employee Slack messages and emails contain deeply personal and sensitive information, and it is not clear whether current stripping techniques are sufficient to prevent re-identification.

Meanwhile, training costs are spiraling. A single GPT-4-scale training run may already cost $100 million or more. As high-quality public data runs out, the combined costs of licensing proprietary corporate data, paying landmark settlements like Anthropic's $1.5 billion, and building synthetic data pipelines are all escalating dramatically. The broader AI training dataset market is projected to grow rapidly, from an estimated $3.6 billion in 2025 to $23 billion by 2034, as human text licensing becomes a formalized asset class .

What This Means for Your Data

For individual users, the landscape has shifted. As of late 2025, both OpenAI and Anthropic changed their default policies to allow training on consumer-tier chats (ChatGPT Free and Plus, Claude Free, Pro, and Max) unless users actively opt out . Enterprise and API customers remain excluded from training by default under contractual Data Processing Agreements .

The message is clear: in the race to feed AI's insatiable hunger for human-generated data, the boundaries between public and private, personal and commercial, are being redrawn—one Slack archive at a time.

As that ceiling approaches, OpenAI and Anthropic have turned their attention to a previously untapped trove: internal corporate collaboration data . Records of live human interactions—emails containing negotiations, meeting transcripts with real-time decision-making, and Slack threads showing authentic workplace dynamics—offer the kind of organic conversational data that has become scarce on the scraped public web.

The Prices: From $10,000 to $300,000 per Deal

A new ecosystem of intermediary firms has sprung up to broker these sensitive transactions. Two names appear most frequently: Mercor and SimpleClosure .

• Failed and acquired startups are selling full Slack archives, email corpora, and document repositories for $10,000 to $100,000 per deal through specialized wind-down services .
• Mercor, an AI training data company, is offering up to $300,000 for Slack threads, GitHub commit histories, and meeting transcripts .
• Warmly, an AI startup recently acquired by HubSpot, received four separate offers of up to $300,000 each for its internal business communication data—meeting minutes, emails, and more—after signing the acquisition agreement, and reportedly turned them all down .

Broader market rates, reported by Reuters in 2024, offer a sense of the value per unit: approximately $0.001 per word for text, $1 to $2 per image, $2 to $4 per short video, and $100 to $300 per hour of longer film or video .

SimpleClosure, a startup wind-down firm, has processed nearly 100 such transactions over the past year, with payouts ranging from $10,000 to $100,000 per company . The firm's "Asset Hub" platform helps founders scrub personally identifiable information (PII) from the data before licensing it—though the effectiveness and consistency of that anonymization remain uncertain and a growing point of contention .

Anthropic's $1.5 Billion Bet on Books: The Project Panama Settlement

The lengths AI companies will go to secure training data were laid bare in the extraordinary case of Project Panama, Anthropic's secret internal program to build a massive book training dataset .

The project had two tracks. First, Anthropic purchased physical print books, cut off their bindings, and destructively scanned them page by page, converting them into searchable PDFs and discarding the paper originals. The goal: convert up to 2 million books in six months . An internal memo advised using a codename "because we don't want it to be known that we are working on this" .

Second, the company downloaded more than 7 million pirated ebook files from shadow libraries including LibGen and Pirate Library Mirror . That track led to a class action lawsuit filed in 2024 by authors Andrea Bartz, Kirk Wallace Johnson, and Charles Graeber, who accused Anthropic of using nearly half a million pirated books to train Claude .

On July 20, 2026, a federal judge in San Francisco approved a $1.5 billion settlement—the largest known copyright payout in U.S. history . More than 500,000 authors and publishers were part of the class action, and they will receive an estimated $3,000 per eligible work .

Crucially, the court drew a legal distinction that has major implications for AI training. Using pirated ebooks was infringing and triggered the settlement. But buying physical print books and destructively scanning them internally for training was ruled to potentially qualify as fair use, because each physical copy was replaced with a single digital copy in a process the court called "exceedingly transformative" . The $1.5 billion covered the pirated book liability; the physical destruction program was not penalized .

The Emerging Challenges: Anonymization and Cost

As companies rush to monetize internal data, significant questions remain. The effectiveness of data anonymization by brokers like SimpleClosure is uncertain and a growing point of contention . Employee Slack messages and emails contain deeply personal and sensitive information, and it is not clear whether current stripping techniques are sufficient to prevent re-identification.

Meanwhile, training costs are spiraling. A single GPT-4-scale training run may already cost $100 million or more. As high-quality public data runs out, the combined costs of licensing proprietary corporate data, paying landmark settlements like Anthropic's $1.5 billion, and building synthetic data pipelines are all escalating dramatically. The broader AI training dataset market is projected to grow rapidly, from an estimated $3.6 billion in 2025 to $23 billion by 2034, as human text licensing becomes a formalized asset class .

What This Means for Your Data

For individual users, the landscape has shifted. As of late 2025, both OpenAI and Anthropic changed their default policies to allow training on consumer-tier chats (ChatGPT Free and Plus, Claude Free, Pro, and Max) unless users actively opt out . Enterprise and API customers remain excluded from training by default under contractual Data Processing Agreements .

The message is clear: in the race to feed AI's insatiable hunger for human-generated data, the boundaries between public and private, personal and commercial, are being redrawn—one Slack archive at a time.