ByteDance builds 1,000-person AI data unit to fuel no-distillation push

ByteDance's new AI Data & Security department consolidates a data organization of more than 1,000 people, betting that proprietary data — not distillation — will decide the frontier-model race. The unit, led by Adam Wang, a former TikTok Live head, sits parallel to core AI businesses Seed, Flow, and Douyin, marking the first new first-level AI department since those were created in late 2023.

"ByteDance's large language models will firmly pursue proprietary R&D, solidify the fundamentals, accept short-term lagging, and persist in optimizing for the long term," CEO Liang Rubo said at the company-wide meeting Aug 5. Founder Zhang Yiming told the Seed team in late July the company would "resolutely not distill," arguing that building real capability requires "long-termism and delayed gratification."

The department integrates the roughly 100-person Global Data team formed in 2023 by TikTok founding member Fu Yue, the corporate data platform DMC, and Flow's AI data platform AIDP. ByteDance's data budget for world and coding models exceeded $10 million in early 2026, with the ability to add more at any time. The data evaluation team for the Seedance video model alone exceeds 1,000 people, with more than a dozen data colleagues supporting each algorithm engineer.

The move reflects a structural shift across the industry: as public internet data is exhausted, high-quality real-world data has replaced algorithms and computing power as the core variable in model competition. ByteDance's refusal to distill means model improvement now depends almost entirely on proprietary data it must source, synthesize, and clean itself.

Data Becomes the Decisive Variable

The bet extends beyond ByteDance. Anthropic allocated more than $1 billion for reinforcement-learning data in 2025 alone. Silicon Valley data-services firm Mercor saw annualized revenue climb from $500 million in 2025 to $2 billion by mid-2026, with 91 percent coming from OpenAI and Anthropic, and its valuation reached $20 billion in three years. Alibaba and Tencent have both increased data procurement budgets, while Tencent has spent the past six months poaching ByteDance data staff at salaries up to three times higher.

The scarcity stems from a shift from public to private data. Models that must work like doctors, lawyers, or researchers need "process data" — the traces of how experts interpret ambiguous intent, search for context, make mistakes, and correct them. That data sits in corporate code repositories and daily workflow artifacts, not on the open web.

The Challenge of Scale

ByteDance's data team runs an internal competition mechanism, dividing teams by world models, coding, and advanced academic subjects, with each project required to calculate its return on investment. The core challenge is keeping a thousand-person organization responsive while model capability boundaries shift rapidly — the most scarce data today can lose its value within six months.

The company is also pushing AI from consumer to enterprise applications. On July 30, ByteDance merged the Feishu product team into Doubao, with Doubao head Zhao Qi taking unified responsibility, making Doubao the entry point for the enterprise AI market. Whether it can grow into a second "thick main trunk" after Douyin, as Liang Rubo described, will be a key test for the next decade.

ByteDance is also reportedly training a model that could reach 10 trillion parameters, according to the Financial Times, led by Seed under former Google DeepMind scientist Wu Yonghui. That would be more than three times larger than Moonshot AI's Kimi K3 at 2.8 trillion parameters and larger than Anthropic's estimated 8-trillion-parameter Mythos 5. Doubao has about 324 million monthly active users in China.

For investors, the data arms race raises the cost of frontier AI development and concentrates value in data infrastructure and services. ByteDance's private status limits direct exposure, but listed peers Alibaba and Tencent face rising data procurement costs, while data-services firms like Mercor capture the spending. The question is whether the market has priced in a data war that is only beginning.

This article is for informational purposes only and does not constitute investment advice.