Refounding Handshake: zero to $1bn in AI data in 15 months
Key takeaways from Knuckle Up with Nakul interview with Handshake co-founder and CEO Garrett LordDealroom has a profile for this one. Try Dealroom → (July 2026):
The Christmas-party origin. In December 2024 Lord spent an entire party talking to one researcher at a frontier AI lab, who mapped model-building into pre-training, mid-training and post-training and argued most future gains would come from human data feeding reinforcement learning. Two weeks later Handshake was recruiting STEM PhDs for that same lab.
Zero to $1bn in gross revenue in about 15 months. Handshake AI has scaled from a single lab relationship to working with essentially every frontier model builder, making it one of the fastest-growing AI companies; Lord notes he set an internal target of $25–$50m run-rate that seemed "insane" at the time.
Refounding a $3.5bn company. Rather than protect the decade-old career network (18m students and alumni, 1,600 schools, most of the Fortune 500, valued at $3.5bn after a $200m Series F in 2022), Lord bet the company on a business barely a year old — spinning up a separate team reporting directly to him, then reshaping the whole org around it.
The network as moat. 89% of people who worked on Handshake AI last month came from the existing network, giving a structural customer-acquisition and quality advantage: Handshake can target experienced accountants, lawyers, farm supervisors and other specialists at scale — paying college students ~$55/hour to "monetize their intelligence" producing post-training data.
Olympic Pace and the human cost. Lord took a remote workforce back to five days a week in-office and codified a value called Olympic Pace (the Michael Phelps analogy of "stacking days"); around 100 people left in the reset. He triages problems red/yellow/green and says he spends 0% of his time celebrating success.
Systematic expansion across data types. From an STEM-reasoning insertion point, Handshake AI moved into generalist/multimodal data, then professionals, then a safety team (preventing models from aiding bio-weapons or harming children) and an international robotics team — acquiring eight companies along the way.
The bear case on human data. Lord argues "there will be human data for as long as humans are needed in the economy," contending that white-collar domains like tax and accounting lack the GitHub-style verifiable corpus that accelerated software engineering, so bespoke human data remains essential to model progress.
Read more: Knuckle Up · Lenny's Newsletter
