Industry insiders report a surge in the purchase of private workflow data to train large language models. An email from a source named Cora warned that “Public data is exhausted, and the labs are spending over $45B on private data to build better models.” The message suggested that the emerging market could rival the billions already earmarked for AI compute infrastructure.
One startup, Handshake, recently offered a freelance writer a $1,000 payment for a single article, though the writer declined. A Handshake spokesperson explained, “We license professional documents across a lot of fields so the labs can test and train how models handle real knowledge work,” adding that journalism is among the targeted sectors. The company also purchases graduate-level papers and coursework for the same purpose.
Mercor’s chief executive Brendan Foody described the data market as poised for rapid expansion, noting that the firm began as an AI recruiting service before recognizing labs’ frustration with low-quality training material. Foody said, “Big AI has two bottomless needs: compute and data.” He also claimed that Mercor removes any personal identifying information before passing data to clients, asserting that “this de-identification can be verified.”
Micro1’s founder and CEO Ali Ansari estimated current annual spending on data by labs, enterprises and robotics firms at roughly $40 billion to $50 billion, though the figure could not be independently confirmed. He projected, “In the next two to four years, this is going to be over a trillion dollars a year.” Micro1 both uses purchased datasets to improve its own models and resells them to larger AI organizations.
The practice attracted public scrutiny when the bankrupt Spirit Airlines listed its digital assets for sale, including 80,000 email accounts, 500 million internal messages and millions of Sharepoint files. Google initially led with a $10 million bid, but Micro1 outbid it by $2.5 million, prompting a dispute. A flight-attendants union later challenged the sale, alleging that the transaction exposed workers’ private information.
Looking ahead, Ansari predicts that selling workflow data will become a routine revenue stream, even surpassing advertising as a common line item for corporations. He stated, “I think it’s going to become more common than ads as a revenue line item for companies.” The author declined to provide WIRED’s editorial records to a data-seeking startup, noting growing resistance among workers to having their daily activities harvested for AI training.