Jiangsu offers vouchers for the compute, data and models behind AI applications

Jiangsu has published three schemes intended to reduce the expense of putting AI to work. A provincial notice posted on 9 October, dated 30 September, covers compute services, data purchases and model development. The distinctive idea is to support several inputs to a functioning application rather than treating AI investment as a single equipment purchase.[1]

For compute, qualifying spending begins at CNY10,000 and token calls must use generative models filed with China’s cyberspace regulator. The applicant must operate normally in Jiangsu, have a defined use case and buy from an unrelated supplier. Compute contracts signed on or after 1 July 2026 can qualify, subject to the scheme’s evidence and budget rules. Discounts, rebates and refunds are excluded from the reimbursement base, and provincial co-funding reimburses part of local expenditure rather than creating a second 50% payment.[1]

The data route also covers eligible contracts and transactions from 1 July 2026, including purchases, selected high-quality dataset projects and embodied-AI data collection. The model-development route is different: the applicant must be the filing entity for at least one Jiangsu model that is already operating with real users and a defined application. Support is ranked by registered users, API activity and realised transaction value, so meeting the threshold enters a competitive selection rather than guaranteeing an award.[1]

The data-purchase ceiling should not be applied to every data project. The same notice provides separate one-off awards of CNY500,000 for qualifying national dataset projects and CNY200,000 for qualifying provincial projects. Embodied-AI data collection has a different ceiling: up to 30% of qualifying settled transaction value, capped at CNY2 million per applicant. The compute and data schemes run to 30 June 2027, subject to their implementation and budget rules. The model voucher counts an eligible continuous twelve-month spending period, and the same expenditure cannot receive duplicate support.[1]

What a subsidy changes in an application budget

Take an illustrative company with CNY200,000 of eligible, fully paid compute expenditure. At the maximum combined support rate, its reimbursement would be CNY100,000 and its net compute cost CNY100,000. Suppose it also spends CNY300,000 on other work that receives no support. Its total cost falls from CNY500,000 to CNY400,000: a 20% reduction, not 50%. These are invented budget inputs, calculated to show how the scheme would work at its ceiling.

The example matters because applications have several costs. A more affordable inference bill can make experimentation easier, but the benefit to the whole project depends on how much of its spending qualifies. Teams with the same total budget can receive very different effective assistance if one buys compute and another spends mainly on integration.

Encouraging demand as well as supply

The schemes sit alongside Jiangsu’s June AI-plus-manufacturing plan, which sets 2027 targets for industrial models, agents, datasets and application examples. Together they describe an adoption strategy: cultivate the technical components and encourage firms to assemble them into industrial uses.[2]

The three instruments are most useful when they change which applications move from experiment to operation. That can be measured: compare projects that received support with similar applicants, track whether the resulting systems remain in use after the subsidy, and record whether the policy expanded access for smaller organisations or mainly reimbursed work that would have happened anyway.[1][2]

Report an error or suggest a correction