China's AI Computing Vouchers Cut Costs for Small Firms and Target Idle Data Centers

Public subsidies connect smaller AI teams with unused servers

Shanghai has earmarked about 600 million yuan for computing vouchers that can cover up to 80 percent of rental fees, while Shenzhen has announced a program worth 500 million yuan a year. Across China, local governments are using these subsidies to tackle two problems together: smaller businesses struggle to afford artificial intelligence computing, while some data centers have more machines than customers.

Beijing, Shanghai, Chengdu, Shenzhen and other locations are offering credits that reduce the cost of renting powerful servers. Small and medium enterprises, often called SMEs, can use that capacity to build, adapt and run AI models without buying expensive equipment themselves.

Shanghai is also setting aside 100 million yuan for data vouchers connected to large language model training. Shenzhen offers standard computing subsidies of up to 50 percent for enterprises and institutions, rising to 60 percent for startups. Chengdu is expanding a pilot with 100 million yuan for research institutions, while Shandong has made 30 million yuan available and plans a further 1 billion yuan investment in local AI infrastructure.

As of September 6, 2025, Beijing and Henan have also begun taking applications for similar schemes. The programs connect an immediate business need, cheaper access to AI tools, with a larger public investment challenge: making use of China's rapidly expanded data center network.

Industry accounts describe some facilities operating at only 20 to 30 percent of capacity. Much of the construction has taken place in western provinces, where electricity and land are cheaper. Vouchers aim to bring paying workloads into those facilities while helping smaller teams move their ideas beyond an initial demonstration.

How the national policy developed

The National Development and Reform Commission rolled out the voucher mechanism in December 2024 to reduce research and development costs for smaller firms and connect more users to a nationwide computing network. Local governments are translating that approach into programs with different budgets, subsidy rates and eligible services.

China's wider strategy seeks to expand AI capabilities by 2030. Public funding, laboratories backed by the state and talent programs sit alongside subsidized access to computing. The China Academy of Information and Communications Technology has estimated that each yuan invested in computing power can generate three to four yuan in economic output, helping explain the interest among local authorities.

The policy sequence also shows that regional support was developing before the national mechanism:

• November 2024: Guizhou updated its voucher rules to include services for adapting software to domestic systems and training services.
• December 2024: The National Development and Reform Commission rolled out the national voucher mechanism.
• January 2025: Guizhou authorities reported 177 vouchers issued nationwide, alongside computing transactions worth more than 10.5 billion yuan.
• Early 2025: Electricity consumption rose sharply in Gui'an New Area as computing activity expanded.
• May 2025: Renewable capacity had risen to roughly a third of total generation in Guizhou.
• September 6, 2025: Multiple cities and provinces had announced funding, expanded pilots or opened applications.

The vouchers also fit the AI Plus initiative referenced in the 2025 Government Work Report. That policy encourages AI use across traditional industries and the development of intelligent devices and robotics. Subsidized computing provides a practical connection between those goals and projects in factories, warehouses and clinics.

What a computing voucher actually buys

Computing vouchers are credits administered by local governments. Eligible recipients can include smaller businesses, startups, universities and research institutes. They redeem the credits at participating data centers, which may be owned by the state, operated by universities or run commercially.

The support generally pays a share of rental charges for clusters of graphics processing units, or GPUs, together with storage and networking. GPUs perform many calculations at once, making them useful for the large amounts of mathematical work involved in AI. Renting a cluster gives a team access to several of these processors working together.

Supported tasks range from pretraining, when a model learns broad patterns from large amounts of data, to fine tuning, when an existing model is adapted for a narrower purpose. Computer vision projects, which process images, and inference at scale are also covered. Inference means using a trained model to produce an answer or prediction.

Data vouchers address another expense. They can reduce the cost of training data, data labeling or curated collections of text used by large language models. Labeling adds information that helps a system learn, while a curated collection organizes material selected for a particular training purpose.

Coverage varies by location. The common principle is to lower the user's bill, with participating providers supplying the equipment and services. A business therefore needs both an eligible project and a provider whose systems can run it.

Shanghai and Shenzhen support more than computing

Shanghai's combination of about 600 million yuan for computing and 100 million yuan for data addresses two major costs in building AI systems. Computing vouchers support the processing needed for training and fine tuning. Data vouchers help pay for acquiring or preparing training material, including data for large language models.

For a small team, paying for servers alone may leave too little money to prepare useful data. Shanghai's paired approach is intended to help businesses move from a proof of concept, an early demonstration that an idea can work, toward something ready for testing in practice.

Shenzhen's annual program offers up to 50 percent support to companies, universities and research institutions renting intelligent computing capacity. Startups can receive up to 60 percent, giving young businesses a higher subsidy when early training bills are difficult to manage and reserving equipment far in advance may be impractical.

Shenzhen's measures also refer to model vouchers, corpus vouchers and caps on training support for individual projects. A corpus is a collection of material used to train a model. The city's approach reaches beyond infrastructure to model service providers and product developers that register their work with the National Internet Information Office.

Enterprise chat assistants, coding copilots and AI features in consumer devices are among the applications this approach seeks to bring into the market. The policy combines help with the inputs needed to build models and an interest in getting usable products into customers' hands.

Chengdu, Shandong, Beijing and Henan take different routes

Chengdu is expanding a pilot that includes 100 million yuan in voucher support for research institutions. Supporting those institutions can help generate local projects and develop people with the skills to work on AI, linking computing access with research and training.

Shandong is pairing an initial voucher pool of 30 million yuan with a proposed investment of 1 billion yuan in local AI infrastructure. The two measures address different sides of the market: vouchers help customers pay for services, while infrastructure investment expands the equipment available to serve them.

Providers considering investment may find a pool of subsidized users attractive because it offers a more predictable source of demand. Whether that demand continues after public support ends will matter to facilities built around those expectations.

Beijing and Henan are at earlier stages, with application windows open and data center partners enrolling. Their participation broadens the reach of the approach, although the size and terms of local support remain uneven.

Guizhou links domestic equipment, storage and AI

Guizhou illustrates the inland computing strategy. Under the national East Data West Computing approach, facilities in inland provinces serve demand from the coast. The aim is to connect places with substantial computing needs to locations that can supply land, electricity and large facilities at lower cost.

Guizhou's intelligent computing capacity has been reported at 85 EFLOPS. EFLOPS is a measure of computing speed, describing how many mathematical operations equipment can perform each second. More than 98 percent of the reported capacity is classed as intelligent computing rather than traditional storage, and about 90 percent of the hardware is described as domestically produced.

The province has 48 key data centers operating or under construction, including 28 large sites. Storage capacity totals 25 exabytes. Together, those figures describe a shift from facilities centered on storing information toward a mix that places greater weight on AI training and inference.

Guizhou's November 2024 voucher changes extended support to domestic adaptation services and training. Reported coverage reaches 30 percent of eligible costs, with a ceiling of 1.5 million yuan for adaptation and 5 million yuan for training. Adaptation services help software work with domestic equipment and systems, an important task when developers move projects between different computing platforms.

Power supply is central to the expansion. Electricity consumption in Gui'an New Area rose sharply in early 2025, and local authorities added substations and dedicated lines to supply racks containing densely packed equipment. Hydropower is a large contributor, with renewable capacity rising to roughly a third of total generation by May 2025.

Guizhou's package therefore connects infrastructure, energy supply and user subsidies. Its goal is to turn inland capacity into computing services that businesses across China can actually use.

Can vouchers fill the empty racks?

China's data center construction expanded rapidly during the past few years. East Data West Computing encourages companies and public bodies to send data westward over long distance networks, taking advantage of cheaper land and power while continuing to serve users elsewhere.

Some facilities opened before enough demand arrived. Industry analyses and press accounts describe pockets of utilization at 20 to 30 percent. For their operators, vouchers are a way to attract customers, use more of the installed equipment and recover construction and equipment costs.

AI training often arrives in bursts. A team may need thousands of GPUs for a run lasting several days, followed by a quieter period while it studies the results. Coordinating voucher calendars and job queues could help providers fit these bursts together, use quieter periods and offer startups more predictable booking slots.

For users, a lower price per GPU hour can extend the time available for experimentation. A budget that would otherwise support two months of work might stretch to six months, allowing more attempts and revisions. Teams adapting existing models can be especially well placed to benefit because fine tuning can produce useful results on more modest clusters than those needed to train a very large model from the beginning.

The practical value depends on matching jobs to suitable equipment. An unused server becomes useful to a business when the hardware, software, network connection and booking schedule meet that business's needs.

Export controls make compatibility more important

US led export controls restrict sales of the most advanced AI accelerators to Chinese customers, complicating procurement for businesses seeking the latest Nvidia hardware. China is responding with greater reliance on domestic chips, universities and local cloud providers, while some companies explore hosting equipment overseas.

Plans to equip dozens of new data centers with large orders of Nvidia Hopper GPUs have also been reported. Details have not been verified, and the proposed purchases face legal and logistical questions under the restrictions.

Domestic alternatives, including Huawei's Ascend line, are improving. Developers nevertheless face a mix of software environments. Many use global open source frameworks such as PyTorch and TensorFlow, while domestic systems including MindSpore and PaddlePaddle are developing alongside them.

A framework supplies tools for building and running AI models. Moving a project to another type of processor may require changes to its code or supporting software. Code portability, the maturity of the software and the experience of the people operating a cluster all affect which facilities a team can use.

Vouchers lower the cost of trying compatible equipment. Domestic adaptation support, such as Guizhou's, addresses the work involved in making that move. Performance differences between the newest US hardware and current domestic alternatives still shape the scale and speed of the jobs a cluster can handle.

China's broader approach combines computing subsidies with investment in domestic hardware and software. Analysis of its wider policy tools appears in RAND's full stack AI strategy discussion and MERICS' AI self reliance analysis.

What smaller businesses gain, and what may slow them down

A business previously able to rent only a few GPUs for short periods may be able to book longer runs, compare more versions of a model and move from a prototype to a field test. Manufacturing, logistics, finance and healthcare are among the sectors where cheaper experimentation could shorten the journey from an idea to a pilot.

Universities and laboratories gain more flexibility in their budgets, potentially bringing more students and researchers at the start of their careers into large model projects. Access beyond the biggest technology platforms can also give smaller firms more room to compete in specialized applications.

Local governments gain a route to earn revenue from idle facilities and make use of earlier capital spending. Better use of inland centers can support power and network planning and reduce the risk that expensive equipment remains unused. Paired data vouchers can also create business for data providers and labeling companies, particularly in Shanghai and Shenzhen.

Applicants may still face paperwork, verification and scheduling delays, especially when demand rises near deadlines. Hardware compatibility can limit choices, and wealthier cities can afford larger subsidies and stronger supporting networks than places with smaller budgets.

Providers face their own risks. Demand may fall when vouchers expire, and rules that channel users toward selected facilities could divert business from private cloud companies. A program's lasting value depends partly on whether supported projects become services that customers continue to buy.

Which results will show whether the spending works?

Guizhou authorities reported that, by January 2025, 177 vouchers had been issued nationwide, supporting computing transactions worth more than 10.5 billion yuan. The transaction total combines large enterprise deals and business backed by vouchers, so the amount attributable to subsidies is still unknown.

Early outcome figures remain sparse. Three useful measures are utilization at western data centers, including the difference between busy and quieter periods; participation by genuine smaller businesses compared with large firms; and the time between applying, booking equipment and completing a job.

Local governments could make programs easier to assess by publishing recipients, funded workloads and project outcomes. Waiting times and booking success rates would also help teams plan work. Providers could publish compatibility tables showing which hardware and software combinations they support, helping applicants choose facilities before committing their time and money.

China's policy arrives amid intense worldwide investment in AI and an effort to align public spending with private innovation. The closest match may be with businesses adapting existing open source or commercial models to their own data, rather than trying to train enormous frontier models. Those projects often fit shorter training runs and subsidized clusters.

The central test is concrete: whether lower prices give more teams reliable access to suitable machines, produce useful applications and bring lasting work to facilities that currently have spare capacity.

Key Points

• Shanghai has earmarked about 600 million yuan for computing vouchers and 100 million yuan for data vouchers, with computing support covering up to 80 percent of rental fees.
• Shenzhen's annual program is worth 500 million yuan, offering up to 50 percent support for eligible users and 60 percent for startups.
• Chengdu, Shandong, Beijing and Henan are expanding funding, planning infrastructure or accepting applications under locally designed schemes.
• The national voucher mechanism introduced in December 2024 aims to lower research costs and connect users with computing capacity.
• Some data centers have reported utilization of only 20 to 30 percent, making subsidized demand attractive to operators.
• Guizhou combines large computing facilities, domestic equipment, power investment and subsidies for adaptation and training.
• US export controls, software compatibility, administrative delays and the durability of demand will influence how much businesses gain.
• Participation by smaller firms, faster access to equipment and sustained use of idle servers are key measures of success.