Andrew Chen: Acquisitions in 2012 Were for Users, in 2026 for Training Data
a16z General Partner Andrew Chen summarizes the evolution of tech acquisition logic: acquisitions in 2012 were for users, in 2021 for engineers, and in 2026 for training data.
This observation reflects a shift in the focus of mergers and acquisitions in the AI era from traffic and talent to proprietary datasets and usage logs.
In recent large transactions, user interaction data, code trails, and behavioral records of target companies have become core valuation drivers.
Acquirers are accelerating model training and iteration through data integration rather than merely acquiring products or teams.
The priority in tech acquisitions has shifted from user scale and talent density to data assets, with capital flowing towards targets that possess high-quality training corpora. Beneficiaries are companies that accumulate unique data flywheels, while those under pressure are pure algorithm teams with scarce data. The event-driven AI competition has shifted from computing power and model parameters to data moats.
Source: Public Information
ABAB AI Insight
Andrew Chen has long observed the consumer and growth sectors, previously focusing on network effects and user acquisition at a16z. The capital path has extended from early internet user growth to the importance of AI data.
Resources are shifting from traffic acquisitions and talent grabs to data asset integration, motivated by the fact that high-quality training data has become a key bottleneck for model differentiation, and acquisitions can directly shorten the data accumulation cycle.
Analogous to the acquisition waves of social platforms for users in the 2010s and AI labs for researchers in the early 2020s, we are currently in a phase transitioning from talent competition to data flywheel competition, with the industry as a whole shifting from "buying people and users" to "buying data and trajectories."
Essentially, this represents a transfer of pricing power, where training data has shifted from being a byproduct to a core asset, as model performance heavily relies on unique, large-scale, and continuously updated real usage data.
ABAB News · Cognitive Laws
Acquisition logic changes targets every decade
Data is scarcer than engineers
Training trajectories are the new moat