Guangzhou offers up to RMB 1m per AI project as subsidy window closes

Most Chinese provinces have already closed, or are closing, their 2026 round of AI-specific subsidies. The deadline is the last day of September. Guangzhou selects ten robotics application projects a year, each eligible for up to RMB 1m; Shanxi recognises only one innovation consortium per track. The dates matter more than the amounts: applications cluster between July and September. Miss this round and the next option is either next year or nothing.

The money is flowing hard in this direction

Allianz Research published a set of figures, reposted by a financial infographics account on 26 September: globally, the cumulative number of subsidy measures in force for AI-related goods rose from fewer than 500 in 2015 to more than 3,600 in 2025. Its share of all subsidy measures also climbed from about 12% to nearly 16%.

The channels are fiscal support, tax credits and state funding, aimed at critical materials, advanced semiconductors and data-centre components.

Abroad is one picture; at home the moves are denser. Since July this year, multiple provinces and municipalities have issued targeted support policies for large models, AI agents and embodied intelligence. A policy review aimed at AI companies lists several categories:

• Guangzhou: embodied intelligence, full humanoid robots, supporting large models and deployment scenarios. Ten robotics application projects are selected each year, each eligible for up to RMB 1m in subsidies; before 2029 the city aims to make breakthroughs in 15 core embodied-intelligence technologies.
• Shanxi: "AI Plus" industry innovation consortia, focused on vertical large models and industry-specific AI agents in energy, manufacturing, healthcare and education. Only one consortium is recognised per track across the province.
• Guangdong: AI benchmark projects for new industrialisation. Companies across the province register online in a central reserve; registered projects later receive special subsidies.
• Anhui: a special fund for data elements, covering enterprise digital-transformation data platforms, datasets dedicated to large-model training, and computing infrastructure.
• Sichuan: a major science and technology programme for AI, covering general foundation models, multimodal industry agents and embodied-intelligence robot control systems. Companies are encouraged to apply in groups with universities.

Read the list and it becomes clear the threshold is not only whether the technology is advanced enough. Shanxi's one-per-track rule is closer to a qualifying round; Guangdong requires registration before disbursement, so position early.

Most companies are not unqualified, they just do not know this exists

The policy review records a candid line from the CTO of an AI startup: "Our AI product is technically ahead, but it never occurred to us that the government had special subsidies we could apply for."

According to the review, companies miss out for three reasons. On awareness, founders with technical backgrounds tend to assume policy is the government's business. On information, special policies are scattered across the industry and information technology, science and technology, and data bureaus, with no single entry point. On execution, application materials require translating technical language into policy language, which is hard work for an R&D team.

The article also offers a three-minute self-check. Hit two or more and it is time to act:

• The product involves large models, AI agents, embodied intelligence or humanoid robots;
• The company holds AI-related invention patents or software copyrights;
• The technology has a clear industry application, such as manufacturing, healthcare or energy;
• The company is a high-tech enterprise or a technology-based small or medium-sized enterprise;
• The company is a "specialised, refined, distinctive and innovative" enterprise, at provincial or national level;
• R&D spending accounts for 5% or more of revenue.

Of the last three, high-tech enterprise and technology-based SME status are the dividing line for many AI companies. Without that label, even the strongest technical credentials struggle to reach the review table.

Also in September, another kind of subsidy was making noise, though of a different nature. Tencent News reported on 24 September that ByteDance's Doubao Work issued 30-day free subscriptions to all staff twice within a month; Qwen Office and Workbuddy gave new registrants 2,000 credits plus 100 credits for daily logins; and Workbuddy's enterprise edition was upgraded into a suite, with Tencent Docs, Tencent Lexiang and Tencent Drive among the first to integrate.

Personal-edition pricing differs little: Doubao Work at RMB 68 a month, Workbuddy at RMB 99, Qwen Office at RMB 98, with discounts at the latter two.

These subsidies buy usage habits. They are a separate matter from whether you can win a government designation. Ding Daoshi, an internet industry analyst, argues that this is the year large models went paid, and that subsidies are effective for opening the market, but retention ultimately depends on whether the product gives users a sense of gain.

Our view: platform subsidies can save a few months of tooling costs and are worth taking; policy subsidies confer qualifications, financing credit enhancement and ecosystem access, and are worth applying for. Schedule the two separately. Do not assume that because you have taken up a free subscription, your "AI credentials" have been officially recognised.

How to proceed: register first, apply second

Registration is not complicated. Contact the local industry and information technology bureau or science and technology bureau and ask whether any AI-specific programme is open. At the same time, gather proof of technical achievements, intellectual property certificates and evidence of R&D spending. If the locality uses a reserve system, get the project into the reserve pool first and wait for the disbursement notice — according to the review, some projects become eligible for follow-on subsidies once registered.

The real bottleneck for most teams is the paperwork: application documents running from dozens to hundreds of pages, in which every requirement must be matched against the original guideline text, with evidence clearly identified. This is where SHEYU ZHISHEN (舍予AI智审) comes in. It organises science and technology project application materials into a verifiable evidence table, reading dozens to hundreds of pages in seconds. Every extracted value carries a page number and original text, and every conclusion can be traced back to the guideline. Data stays on the internal network. What the person writing the application needs is not "make it sound impressive" but knowing, before submission, which point will not survive questioning.

Beyond that, treat this round as practice. The experience of a special programme is itself a track record for the next, higher-level application, and selection lists are often treated by banks as a plus for loan credit enhancement.

One closing point: AI subsidy money rarely goes to the most technically advanced company. It mostly goes to the company whose materials are most complete, and which knew when the window would close.