Hong Kong's 70% Compute Subsidy and Chengdu's Token Vouchers: Which AI Grant to Apply for First

In the second half of 2026, the density and size of AI subsidies in China rose another notch. The more important change is not that the money grew, but that the target of the subsidies has shifted. In the past, subsidies went to "installing systems and building platforms." Now they go to the processing and consumption of data itself: registration and listing, corpus processing, token calls.

For operators, whether you can get the money depends on how much measurable usage you have and how verifiable your records are. The window is roughly the next 12 months.

Q: Does a small team of about ten people have anything to do with these subsidies? We don't train large models

More than you might think.

On 29 August 2026, the General Office of the Ministry of Industry and Information Technology (MIIT) issued the "AI SME Entrepreneurship Support Plan (2026–2028)," publicly released on 4 September. This is MIIT's first special plan that takes "AI SME entrepreneurship" as its integrated target. Its four main lines cover factor supply, entity cultivation, ecosystem building and service guarantees. The three-year goal is to cultivate 10,000 technology-based SMEs and 2,000 "little giant" specialized and sophisticated enterprises.

It departs from the old playbook in two places.

First, its definition of "AI entrepreneurship" does not stop at the model layer. The compute vouchers, the inclusive compute special action, and the SME zone on the China Computing Power Platform serve those who use compute, not those who build it. Teams doing vertical applications, industry tools and data processing are all on this line.

Second, service guarantees have shifted from after-the-fact reimbursement to full-cycle accompaniment. Patient capital, investment-incubation linkages, compliance services and talent housing are embedded across the entire entrepreneurial cycle. In plain terms: don't wait until you're done to claim reimbursement. There is a point of contact from day one.

Q: Hong Kong's HK$3bn compute subsidy — can overseas teams reach it?

According to public information, the Hong Kong SAR government has allocated HK$3bn to subsidise enterprises, universities and R&D institutions using compute at the Cyberport AI Supercomputing Centre, with the subsidy ratio generally no less than 70%. There are also the "New Industrialisation Funding Scheme" and the "New Industrialisation Acceleration Scheme," the latter offering up to HK$200m per enterprise, aimed at AI applications in manufacturing, engineering and testing.

The subsidy targets are written as enterprises, universities and R&D institutions — not local giants only. But whether you can reach it depends on three things: whether you have a viable entity in Hong Kong, whether the project falls within the funded industry scope, and whether compute usage can be measured and attributed. The eligibility clauses in the guidelines must be checked one by one.

That 70% ratio is substantial — it cuts the cost of trial and error by more than half.

Q: Data vouchers, compute vouchers, corpus vouchers — which to apply for first

First, see clearly that the three types of vouchers correspond to three links in the data value chain. Matching yourself to the right one matters more than casting a wide net.

• Data vouchers govern rights confirmation and circulation: Wuhan issued its digital-intelligence economy policy on 24 September, effective from 26 October for two years. Data vouchers support data trading and asset registration; the compute vouchers in the same document support the purchase of compute services and token services.
• Compute vouchers / token vouchers govern the consumption side: Chengdu's token vouchers offer up to RMB 2m per entity, with total annual citywide issuance capped at RMB 100m.
• Corpus vouchers govern supply-side processing: Shandong's corpus vouchers offer up to RMB 1.5m per item, with an additional RMB 750,000 for outstanding ones, backed by a RMB 2bn AI industry fund.

There is a layer above that. Jiangsu's "Action Plan for Promoting High-Quality Development of the Token Economy (2026)" sets out 26 measures, aiming to build 2,000 high-quality datasets and 3 data annotation bases. Beijing has written "tokens generated per unit of electricity" into the main indicators of its "15th Five-Year" digital economy plan.

The threshold is not about money. It is about preparatory materials. Data vouchers require data resource ledgers, ownership documents and compliance review records. Compute and token vouchers require measurable usage, with bills and contracts that can be broken down by scenario. Corpus vouchers require data integration or an existing clear processing plan plus a quality evaluation system. If the materials are not ready, an open window is still a closed door.

Q: So who prepares the materials? We can't even break down our token usage

This is the most common pitfall.

The document arrives at the company and lands with marketing or finance, while the hardest part of the application materials — data resource ledgers, registration documents, data quality certificates, corpus processing records — sits entirely with the technical or data team. By the time the two sides align, two or three months may have passed.

Two moves can be made first.

Establish a collection standard for AI usage across three dimensions: application, scenario and department. Calculate training, inference and testing separately. This is not just application material; it is also leverage in later "pay-for-performance" negotiations. Many companies currently have a single mixed account for compute costs and cannot break out how much each business line used.

The other is to let application documents and accounting audit documents share one foundation. Classification code ledgers, ownership chain documents and quality control records are both application materials and audit materials. One set of documents serves two purposes, at very low marginal cost.

How to implement

The most time-consuming part of an application is not writing. It is verification: can every number and every conclusion in the materials be traced back to the original guideline text and source records?

That is what SHEYU ZHISHEN (舍予AI智审) does — turning technology project application materials into a verifiable evidence table, reading tens to hundreds of pages in seconds, with every extracted value carrying a page number and original text, dual-engine cross-checking, and four-value results that never default to pass, with data staying on the internal network. For teams preparing to apply for vouchers or grants, what it saves is the time lost when "one rejection means waiting another round."

On subsidies, the dividing line in the second half of 2026 is not who has the fastest information. It is whose books are clean. Policy runs to 2028, but the first batch of cases usually takes shape in the first 12 months — sort out your usage and ledgers first, then decide which voucher to apply for.