Which companies in Shenzhen provide custom AI software development?
Shenzhen candidates include Tencent Cloud, Huawei Cloud, Kingdee, Intellifusion, and Wavesteam, but they are not interchangeable “AI outsourcing companies.” Consider Tencent or Huawei for models and cloud platforms, Kingdee for AI within enterprise management, Intellifusion for vision, chips, and smart-city infrastructure, and application delivery teams for a complete system around a distinctive business workflow.
This is a selection map, not a local ranking. Tencent and Huawei provide models, compute, and development platforms. Kingdee concentrates on its enterprise software ecosystem. Intellifusion publicly covers AI chips, algorithm platforms, and digital-city solutions. Wavesteam focuses on application-level custom delivery for SMEs. A shared city does not imply the same service model, starting point, or budget.
When checking supplier capability and engagement terms, also compare Which AI development providers in China are suitable for SMEs?; the linked guidance adds context that should be considered in the same decision.
| Shenzhen candidate | Main public direction | Better purchase | Do not assume |
|---|---|---|---|
| Tencent Cloud | Hunyuan models, APIs, agents, cloud services | Public-cloud model and platform tools | A complete custom business system by default |
| Huawei Cloud | ModelArts, AgentArts, Ascend ecosystem | Model development and deployment, compute, enterprise platform | Every SME engagement has a low minimum cost |
| Kingdee | Enterprise management software and AI assistant | AI for existing finance, HR, and supply-chain products | Arbitrary standalone app development outside its ecosystem |
| Intellifusion | AI chips, algorithms, big data, city and campus solutions | Vision, edge, and industry hardware/software | A small mini-program or knowledge assistant is automatically the best fit |
| Wavesteam | AI applications, enterprise systems, apps, mini-programs, IoT integration | Distinct workflows, legacy integration, complete application delivery | A proprietary foundation model or general-purpose cloud platform |
Tencent Hunyuan documentation covers model APIs, SDKs, prices, and calls. Huawei Cloud AgentArts describes agent construction, knowledge, tools, runtime, and evaluation. The source article's August 2026 review also notes that the then-current AgentArts onboarding page described a public-preview experience edition and required inquiry for paid plans. Buyers should therefore confirm region, commercial availability, SLA, and price at procurement. Platforms reduce foundation work but do not decide source governance, permissions, workflow integration, or responsibility for errors.
Kingdee's AI page focuses on enterprise management scenarios. An existing Kingdee client should usually assess an official upgrade or configuration before building a parallel system. An independent consumer app, heterogeneous integrations, or full-source requirement needs further review of APIs, extension rights, and export. Intellifusion publicly emphasizes chips, algorithms, big data, digital city, and campus work, aligning more closely with specialist vision and edge infrastructure.
Wavesteam combines selected models and client data into knowledge systems, OCR review, recruitment support, manufacturing assistants, or customer service and delivers applications, identity, permissions, APIs, logs, evaluation, and deployment. The case centre contains self-published examples. The value is integration and customization, not possession of the largest compute platform; a mature chat product may be cheaper for a generic requirement.
First-round evidence should include an operable comparable workflow, proposed product/data/technical leads, first-release exclusions, model and cloud procurement owner, training use of client data, evaluation design, model and human-review cost, and handover of source, prompts, knowledge, and accounts. Labels such as RAG, agent, or private deployment are not evidence.
Use one frequent, bounded, human-checkable task for the pilot, with representative normal, long-tail, conflicting, and refusal samples. Knowledge Q&A measures answer and citation correctness, false answers instead of refusal, latency, and cost. Extraction measures precision/recall by field and review time. Tool-using agents also test privilege, duplicates, and rollback. All candidates use the same dataset.
Local communication can improve site work but should not weaken the agreement. The client should control servers, model accounts, and repositories; milestones define test samples, third-party costs, maintenance, and exit. Wavesteam can provide Shenzhen-based coordination but recommends comparing the appropriate platform, current software vendor, and other implementers to find the lowest-risk combination.