The Hidden Market for Corporate Data: AI Labs Pay Millions for Your Work Files

AI companies have exhausted the public internet. To push model performance further, labs like Anthropic and OpenAI are now paying premium prices — ranging from $350,000 to $2 million per deal — for internal corporate archives: emails, Slack threads, support tickets, call transcripts, and project documents. This isn't speculative; it's a documented spending spree estimated at over $45 billion across the industry.

A new tier of data brokers has emerged to meet this demand. Firms such as Mercor, Handshake, and Micro1 — some founded barely three years ago — specialize in acquiring, scrubbing, and licensing proprietary workflow data. Mercor, launched in 2023, saw its founders reach billionaire status before age 23. Their pitch: real-world knowledge work is the highest-value training signal available.

The supply chain is surprisingly broad. Brokers purchase graduate student coursework, freelance journalism drafts, and even the complete digital exhaust of startups winding down operations. One firm, Sunset, explicitly markets "helping startups wind down" as a data-harvesting service — anonymizing records before brokering them to labs.

Why it matters for GPU and AI infrastructure

Training on proprietary, high-signal data reduces the compute required to reach target model quality. Clean, domain-specific datasets mean fewer epochs, smaller clusters, and lower energy per token — directly impacting the economics of GPU cloud utilization. For infrastructure providers, the shift toward private-data training signals sustained demand for high-density, low-latency compute optimized for iterative fine-tuning rather than monolithic pre-training runs.