Why Chinese robotics companies are looking outside China for data services
China's humanoid robotics market has grown faster than its domestic annotation and data collection infrastructure can scale. With 150+ active humanoid development programs, 40+ government-backed training centers, and a national "AI-plus" initiative that treats robot training data as a strategic asset, demand for high-quality robot training data in China in 2026 is running significantly ahead of the supply of qualified annotators and collection operators.
The state-funded training centers - including the Shanghai facility that targets 10 million data entries annually and the network of provincial centers announced across Guangzhou, Shenzhen, and Hangzhou - are real and are producing meaningful volumes of data. But they are not positioned as commercial services for private robotics companies. Priority goes to state-aligned robotics programs and national champion companies. Commercial access to these facilities is limited, competitive, and subject to terms that may not suit proprietary development programs.
Private Chinese robotics companies - including Unitree, Fourier Intelligence, UBTECH, DEEP Robotics, and Horizon Robotics - are therefore evaluating external data service options. The evaluation is not primarily about cost. It is about capacity, specialization, IP protection, and access to managed programs that can scale to the dataset sizes that production deployments require.
Vietnam-based managed data programs have emerged as a credible option for Chinese robotics teams based on three structural advantages: a 1-hour timezone overlap with Beijing that enables same-day feedback cycles, a workforce with meaningful physical AI annotation experience developed through APAC programs, and cost structures that remain competitive with domestic commercial providers (distinct from the government-subsidized state facilities).
1. Cost comparison: Vietnam vs Chinese domestic commercial providers
The cost comparison that matters is not Vietnam vs. state-funded Chinese facilities - those are not commercially accessible for most private robotics programs. The relevant comparison is Vietnam vs. domestic Chinese commercial annotation companies like Datatang, CloudWalk, Ruijin Technology, and Data Hall.
For physical AI training data annotation (video labeling, action segmentation, sensor fusion annotation), domestic Chinese commercial providers charge RMB 150-300 per hour of annotation labor, which at current exchange rates corresponds to approximately $20-42 USD. High-complexity annotation requiring specialist reviewers (surgical simulation, precision manipulation, safety-critical scenarios) commands RMB 400-600 per hour from domestic Chinese providers.
Vietnam-based managed annotation programs for equivalent task complexity run $15-35 per annotation labor hour in USD, depending on QA tier and task complexity. The headline cost advantage is real but should be evaluated alongside overhead factors: management communication overhead (even with Chinese-speaking PMs), potential for format iterations if the delivery pipeline differs from domestic standards, and the cost of cross-border data transfer if the program involves proprietary hardware data.
For simple, high-volume annotation tasks (basic object detection, frame-level classification), domestic Chinese providers often match or undercut Vietnam costs due to their workforce scale. The Vietnam cost advantage is most significant for specialist annotation tasks requiring domain expertise in robotics - action segmentation, contact state annotation, demonstration quality scoring - where the domestic supply is constrained.
2. IP protection: practical safeguards for cross-border programs
IP protection is the primary concern for Chinese robotics companies evaluating overseas data vendors. The concern is not theoretical - training data that reflects proprietary robot kinematics, task designs, and operational scenarios represents genuine competitive advantage that should not be accessible to competitors through a data vendor.
For Vietnam-based programs, IP protection rests on three layers: contractual, technical, and operational. At the contractual level, the data processing agreement must specify that collected and annotated data is owned exclusively by the customer, that the vendor will not use or reference the data for any purpose outside the specific program, and that all data will be deleted from vendor systems within a defined period after delivery. Data deletion should be verifiable through audit logs.
At the technical level, programs with high IP sensitivity should require that annotation work occurs in air-gapped environments - annotation workstations with no internet access, where data cannot be exfiltrated via email, cloud upload, or USB. Vietnam's established data annotation industry has the infrastructure to support air-gapped programs at production scale. The audit trail for air-gapped programs should include session logs showing workstation access events and delivery checksums that verify no files were modified after QA sign-off.
At the operational level, the vendor should sign NDAs that extend to individual annotators working on the program - not just the company-level agreement. Annotators with access to proprietary demonstration data should be subject to individual confidentiality obligations, not just team-level policies.
3. Language support for Chinese-language programs
Chinese robotics programs that train VLA (vision-language-action) models on Chinese-language instruction following require annotators who are native Simplified Chinese speakers with sufficient technical vocabulary to produce consistent natural language task descriptions.
Vietnam-based annotation providers vary significantly on Chinese-language capability. Large providers like DataX Power maintain Chinese-speaking project management teams for coordination with Chinese client engineering teams, and can source Chinese-language annotation for VLA instruction labeling through managed partner networks. Providers without Chinese language capability require English as the coordination language, which introduces translation overhead and potential for terminology inconsistency in instruction annotations.
For programs requiring Traditional Chinese (for Taiwan or Hong Kong markets), the supply of qualified annotators in Vietnam is more limited and should be confirmed explicitly before program commitment. Simplified Chinese annotation at scale is more readily available.
Chinese-language VLA instruction annotation should be reviewed by domain-expert reviewers who understand both the robotics task and the Chinese language - not just the language. Generic Chinese-language annotators who do not understand robot manipulation tasks will produce instruction annotations that are grammatically correct but semantically inconsistent with the task execution.
4. Timezone and iteration speed
Vietnam (GMT+7) is 1 hour behind Beijing (GMT+8). This is the smallest timezone gap available from any offshore data service location, and has practical implications for program iteration speed that matter in active development programs.
A Chinese robotics engineering team that submits a task specification update at 6pm Beijing time reaches the Vietnam data team at 5pm local time - within the same working day. Feedback on a delivered batch can be reviewed and returned before end of the Vietnam working day. This cycle is 16-24 hours faster than programs run from India (GMT+5:30, 2.5 hours behind Beijing) and far faster than programs run from Europe or the US.
The implication for active development programs is that Vietnam-based programs can support weekly iteration cycles - deliver a batch on Monday, provide feedback by Tuesday, revised collection begins Wednesday - that are not practical with larger timezone gaps. For teams in early-stage model development where task specifications change frequently, this iteration speed matters more than marginal cost differences.
5. Scale and capacity comparison
Domestic Chinese commercial annotation providers have the largest absolute annotator pools available for non-specialist tasks. For programs requiring thousands of annotators for high-volume frame annotation, domestic providers can scale faster than any offshore alternative.
Vietnam-based providers have an advantage in specialist capacity for physical AI programs. The concentration of physical AI annotation expertise - annotators and reviewers trained on robot manipulation data, egocentric video quality assessment, and sensor fusion annotation - is higher in established Vietnam providers than in domestic Chinese commercial annotation companies, which have built their capacity around general computer vision and NLP annotation rather than the physical AI specialization that humanoid programs require.
The practical outcome: for programs requiring 1,000+ annotators doing basic labeling, domestic Chinese providers can staff up faster. For programs requiring 50-200 specialist annotators with physical AI domain knowledge, Vietnam-based providers can often scale more quickly because they are not competing against 150+ humanoid programs for the same specialist workforce.
DataX Power works with Chinese robotics teams on physical AI training data programs from Hanoi - including Chinese-speaking project management, IP-protected air-gapped environments, and RLDS-format delivery for VLA model training.
Discuss your program with our team- Does outsourcing data annotation to Vietnam violate China's Data Security Law?
- China's Data Security Law (2021) regulates "important data" and "core data" with specific cross-border transfer restrictions. Robot training data does not automatically qualify as regulated data unless it involves critical infrastructure, national security contexts, or the specific data categories defined in sector-specific regulations. For most private commercial robotics programs, cross-border data transfer to Vietnam for annotation does not require CAC approval. Programs involving military, healthcare, or critical infrastructure robotics should conduct a specific regulatory review before any cross-border data transfer.
- Can DataX Power annotate Chinese-language instruction data for VLA training?
- Yes. DataX Power maintains Chinese-speaking project managers for coordination and can source Simplified Chinese native-speaker annotation for VLA instruction labeling. For high-volume Chinese-language annotation programs (1,000+ hours), we coordinate annotation through managed partner networks with established Chinese-language NLP annotation experience. Annotation quality review is performed by reviewers with both Chinese language fluency and robotics task domain knowledge.
- How does DataX Power handle IP protection for Chinese robotics clients?
- Standard IP protection measures include: data processing agreements with explicit customer ownership and deletion provisions, NDAs extending to individual annotators, air-gapped annotation environments for sensitive programs, no cross-customer data sharing, and audit logs for session access and data transfer events. For programs with particularly sensitive proprietary content (novel robot designs, unreleased product scenarios), we can discuss on-customer-premises annotation infrastructure options.
