ByteDance Makes Data Procurement a First-Level Department

01The department, in facts

Integration began in early June 2026, merging the Global Data team founded by Fu Yue in 2023 with roughly 100 people, the group data middle platform DMC, and Flow's AI data platform AIDP. Global Data first served Dola (Doubao's overseas version) and TikTok, then Seed's training data procurement and quality control. It spans standard-setting, sourcing, synthesis and cleaning, and quality evaluation across ByteDance's models.

Two numbers give the scale: Seedance's data evaluation team exceeds 1,000, and world model and coding model data budgets were already at tens of millions of US dollars in early 2026. Standards, quality inspection, and security now have permanent owners that outlive projects; vendors who once reached the algorithm team directly now pass three standing functions first.

02Working with four reviewers

Procurement asks pricing structure and delivery schedules; QC operations asks the sampling plan, agreement rates, and rework terms; security and compliance ask where data came from and through whose hands; algorithms ask the evaluation-set increment.

One delivery needs four sets of materials: quotation details and schedule for procurement, sampling records for QC, source and personnel lists for compliance, small-batch trial results for algorithms. QC and compliance release last, and the schedule follows the slowest party — quote delivery timelines against their pacing, not algorithm-team hopes.

The standing objection — in-house teams will absorb data production — holds for evaluation: Seedance's 1,000-person team is in-house. But internal teams take standards, QC, and security first, not all capacity. Expert-scarce, volatile directions like hard sciences and world models still outsource; what changed is that acceptance authority returned to the customer.

03Decision rights and the two choke points

The budget sits with the data department; projects start with algorithm teams. Two-step approval: algorithm teams decide what data and how much; the data department decides from whom and at what price. Persuading only the algorithm team does not close a contract; procurement and QC hold the final veto.

QC acceptance: the sides disagree on what pass means; a sampled agreement rate below the agreed line means the whole batch is reworked and payment deferred. Compliance review: data origin, expert identity, cross-border transfer — anything unclear stalls the process; algorithm-team pushing does not help. Remedy: hand over identity verification, collection records, and transfer paths at quote stage.

One passed review shortens the next cycle: its materials are reused. Condition: if by the first half of 2027 data procurement at top Chinese labs still runs on direct orders from individual algorithm teams, this judgment is wrong.

04Evaluation increments and the contract

ByteDance's data teams split by direction — world models, code, hard sciences — and every project gets return-on-investment accounting. The buyer's question: which evaluation set moved by how many points, at what unit cost. Vendors quoting only item counts and unit prices cannot answer; the vendor who can quote an increment wins — given a customer-accepted evaluation set and small-batch trial capability — deliver 1,000 items, show the delta, then discuss 100,000.

Four contract recommendations. Acceptance clauses name the evaluation set, the sampling ratio, and a numeric pass line — an agreement-rate number, not words like "high quality." Payment runs in batches, each released after acceptance, the trial batch preconditioning later payments; cash-flow pressure rises, so the advance funding belongs in the quote. Exclusivity clauses state scope — which data, how long, who owns derivatives — and since exclusivity consumes capacity, its price should exceed non-exclusive. Key-person clauses stand alone: the hard-to-copy assets are the expert-managing project manager and the QC lead; any staged buyout goes into the contract, not a verbal promise.

The scarcest data loses value within six months; contract terms shorten, and renewal depends on the next round of increments, not the last delivery. Condition: if 2027 tenders still score item counts and unit prices as the main criteria, evaluation increments have not become the mainstream acceptance standard.

05Alibaba's role and the shared procurement list

Alibaba's holding company posted a Data Procurement Expert - Large Models role at 40-60K per month, requiring 5 to 10 years of experience (招聘平台). Three duties: build an ecosystem map of global LLM data suppliers — the business models and technical moats of annotation, synthetic data, real-machine collection, and pretraining data providers; run technical due diligence on automation, cost control, and iteration roadmap; push procurement categories from buying data to buying capability. Plus points: data compliance, copyright, cross-border transfer.

"Buying capability" matches a NetEase account article's taxonomy of Chinese post-training data players — expert data, benchmarks, RL environments, synthetic data — commercialized through volume subscriptions, project-based deals, or cloud bundling. The last three sell reusable capability, not a batch of data.

Both point to one list: delivery with QC evidence, sources that survive compliance review, effects as evaluation increments, exclusivity spelled out, vendor risk under evaluation. Vendors still pricing by item count get squeezed first in renewal talks — buyers now hold comparable increments from peers. If 2027 renewal prices still track per-item pricing, this judgment does not hold.

The customer's shape has changed: from one project contact to a four-party review. Vendors arriving with evidence will go further than vendors arriving with data.