Why High-Quality Data Is the New Bottleneck in Large Model Competition

In a nearly four‑hour investor session, DeepSeek founder Liang Wenfeng repeatedly emphasized compute, open‑source, agents and continual learning, but highlighted that high‑quality data is the key competitive factor; the gap in post‑training performance between domestic and overseas models correlates with the accumulation of such data, and this cannot be closed quickly after financing because the bottleneck is time.

During pre‑training, massive corpora give models broad knowledge, but post‑training aims at better reasoning, alignment and fixing concrete defects, so data value is judged by whether it hits the model’s ability boundary rather than sheer scale.

DeepSeek‑R1 technical report illustrates the workflow: start with a few thousand cold‑start samples for fine‑tuning, then reinforcement learning for reasoning; after convergence, generate and filter reasoning trajectories to collect ~600 k inference‑related samples plus ~200 k other samples, forming ~800 k supervised fine‑tuning data, followed by another round of RL. The crucial point is the iterative alternation of data production and training, where each training round exposes new problems that are turned into tasks, annotated, and fed back.

Liang argues that high‑quality data is not a purchasable commodity but an organizational capability. Building a stable data‑production mechanism requires time; capital can buy compute instantly, but a mature pipeline needs repeated refinement on real tasks. Five capabilities are needed: converting poor model performance into concrete tasks, enabling domain experts to express tacit knowledge as learnable samples, establishing stable labeling, review and evaluation standards, identifying valuable hard cases from failures, and feeding training results and online feedback back into data creation.

He also notes that high‑end data annotation does not enjoy a Chinese low‑cost advantage for knowledge‑intensive tasks such as medical, finance, research, code, or complex agents; only simple classification can be tool‑automated.

The bottleneck is not the dataset itself but the production system behind it; competitors may copy sample size or some methods, but cannot quickly replicate how a team selects problems, organizes experts, judges data value, and links data to model iteration.

The 2026 national policy on high‑quality data sets defines them as collections that directly improve AI model performance across pre‑training, instruction fine‑tuning, RL, evaluation and agents, and calls for a data flywheel: scenario‑driven data, data‑driven models, model‑empowered applications, and application‑generated value. This shifts data governance from static accuracy/completeness to continuous validation of model learning and production impact.

For enterprises, “Data First” must evolve from merely storing data to operating a data flywheel. Besides factual and metric data, enterprises need to retain task trajectories, tool calls, failure reasons, human corrections, approvals and business impact. A practical loop starts with high‑value scenarios, defines tasks and success criteria, records model actions and outcomes, converts errors and human edits into a problem bank, produces samples with experts, validates via evaluation and deployment, then scales.

The overall insight: as models become stronger, data work moves closer to research, relying on expert knowledge and continuous verification. The long‑term competitive moat is not the amount of data owned today, but the ability to continuously turn every real‑world success and failure into the next incremental capability.