UniPat at $2.5B: China's AI data vendor market opens

01Four kinds of players

The first group is founders out of model companies. UniPat founder Li Kuan: Beihang undergrad, master's from the Chinese Academy of Sciences' Institute of Computing, HKUST PhD candidate; last year an intern on Alibaba Tongyi's DeepResearch project and a lead of the open-source agent WebSailor. Monolith and ByteDance's Jinqiu Fund are early shareholders. Copula Lab founder Liang Li previously ran MiniMax's agent business; technical partner Cai Jiaren worked on MiniMax's open-source releases and Coding post-training for the M2 series. Zhineng Zhishi's core team came from ByteDance Seed and took a June angel round from the Jinqiu Fund and Yaotu Capital.

The second group is in-house expert platforms: ByteDance runs Xpert, and on September 21 Tencent launched Hunyuan Zhinangtuan, covering ten-plus fields — finance, law, medicine, code, translation — where experts take assignments online, are rated dynamically on quality, and receive Tencent-certified credentials. Model companies no longer just buy data; they run expert supply directly.

The third is listed vendors pivoting: in private placement materials, Speechocean (Haitian Ruisheng) named its fundraising project "AI large model training dataset construction" and disclosed it is evaluating an evaluation-services capability — evaluation plan design, evaluation data acquisition, automated evaluation.

The fourth is the synthetic-data camp: Weina Intelligence has models write their own tasks, reasoning and answers, packaged as cQrA quadruples; Lenovo Capital put in a seed round of roughly ten million dollars.

02The US playbook does not transfer

In the US, Scale sells data to every lab and Mercor sells experts to every lab — no cross-shareholding, no sides. When Meta bought into Scale, Google and OpenAI immediately cut back cooperation — the price of independence.

In China, all three seats are taken by the model companies themselves. Supply side: ByteDance and Tencent run their own expert platforms. Capital side: Alibaba led UniPat's round, Tencent followed, ByteDance's Jinqiu Fund is an early shareholder, and in January Sequoia China's xbench co-released the BabyVision benchmark with UniPat — half the industry's majors are stakeholders in a company whose round has not closed. Demand side: industry reporting puts ByteDance's 2026 budget for world-model training data in the tens of millions of RMB, Coding data second. This core post-training work stays in-house; outside teams get the layer in between — expert organization, task production, and some environment construction.

One reverse fact: after visiting China, Dimension Capital's internal letter to investors noted that US data companies — Mercor, AfterQuery, Turing — now supply Chinese model companies. Chinese labs order from independent overseas vendors while using capital and in-house platforms to turn domestic rivals into their own people. Grading authority must stay controllable; that is the underlying rule of the domestic table.

03Expert counts are inflated; the moat is failure modes

Zhineng Zhishi's Yimian Qianshi claims more than 50,000 domain experts; Yuanrang Intelligence's TalentsAI also claims 50,000 across 300-plus sub-disciplines. Two 50,000s side by side and the number is no moat — how many registered experts can reliably deliver hard tasks is unverifiable from outside.

Two things separate vendors. One is knowledge of failure modes: UniPat builds tasks hard enough to expose capability gaps, produces data and rubrics around those gaps, then trains models and checks for gains on held-out sets and external benchmarks. Its ExpertEval covers medical, finance and law: 3,213 expert cases, 21.83 rubric items per task on average, with focus on "critical negative items". Copula Lab starts from why a model fails on real tasks and works backward to tasks, trajectories and rubrics. This loop — find the gap, produce the signal, train and verify — answers the hardest question a data vendor faces: did this data make the model stronger?

The other is verifier engineering. Zhineng Zhishi's Terminal RL Data ships Docker environments, test scripts, Oracle reference solutions and run trajectories, not just questions and answers. Membership for its Xpert Studio data-production platform costs 3,000 RMB a year — the tool itself no longer commands a price; the value is the QC protocol and expert network inside it.

04Who grades the graders

UniPat publishes benchmarks, produces training data, and trains models itself. After Bloomberg's report, multiple analyses flagged the conflict: a company that trains a model and also evaluates it will have its conclusions doubted by external customers. The US already argued this through in the Meta-Scale episode; in China it is UniPat's turn to answer.

In October, Tencent Hunyuan, Fudan and Tsinghua released ExplorationBench, which builds two "alien worlds" a model cannot have prepared for, to test exploration. The paper's honesty paragraph: rerun exploration with the same system and budget, and Kimi K3's endpoint ranges from 5.7% to 79.0%; rank by best-of-three versus rank by mean, and the order flips. The people publishing the leaderboard admit a single score can mislead — the most honest evaluation text of the week.

05What to watch: acceptance authority

What matters next is not whose expert count or valuation is higher, but where acceptance authority lands: in model procurement, enterprise deployment and training pipelines, whose verdict of "this model is good enough" gets trusted. UniPat's $2.5 billion valuation buys exactly that position.