NeuroAI NEUROAINEUROAI.SITE
ESC

Jiangsu hands out three AI vouchers — token, corpus and model — up to ¥5M per firm

Jiangsu's five departments launched token, corpus and model vouchers to cut AI costs; a single firm can claim up to ¥5M (≈US$700k) in compute aid through mid-2027.

2026-10-27 · 801 words · NeuroAI
Jiangsu hands out three AI vouchers — token, corpus and model — up to ¥5M per firm

A startup in Suzhou wants to train a vertical model but balks at the compute bill. A factory in Nanjing needs quality datasets but cannot price the licensing. A research lab hesitates to launch an agent (智能体) because inference costs never end. These are the small frictions that decide whether "AI+" stays a slogan or becomes a habit.

Jiangsu, one of China's largest provincial economies, just turned those frictions into paper coupons.

Three vouchers, one pipeline

On October 9, 2026, five Jiangsu departments — the Development and Reform Commission, the Cyberspace Administration, the Department of Industry and Information Technology, the Department of Finance and the Data Bureau — issued an implementation plan introducing three AI subsidy vouchers:

  • Token voucher (词元券). Reimburses purchases of intelligent compute services and compute leasing for model training, inference and agent use. Provincial and city support together cover up to 50% of actual spend; a single entity can receive up to ¥5 million (≈US$700k). Contracts must be signed on or after July 1, 2026, with spend of at least ¥10,000.
  • Corpus voucher (语料券). Covers buying and building high-quality datasets and industry corpus. Up to 50% of cost, capped at ¥1 million (≈US$140k) per entity. National pilot dataset projects get a one-time ¥500,000; provincial excellent projects ¥200,000. Embodied-intelligence robot data products get up to ¥2 million (≈US$280k) at 30% of settlement value.
  • Model voucher (模型券). For in-province entities with a generative-AI service filed with the CAC. Covers 30% of R&D spend, up to ¥1 million (≈US$140k).

Why vouchers, and why now

Jiangsu is executing its provincial "AI+" action plan. Rather than building one big compute center and hoping firms show up, it is subsidizing the metered inputs firms actually buy — tokens, data, model R&D. The design matters: a "token voucher" pays by usage, which matches how models are billed today far better than a fixed hardware grant.

According to Li An, chief scientist at BrainNet (脑机网), metering by token instead of by hardware is a smarter subsidy design because it follows real consumption instead of speculative capacity. The Yangzhou city-level pilot that preceded it showed metered token subsidies were claimed quickly, which is why the province is scaling the model rather than reinventing it.

How it is built to avoid waste

The vouchers are "pay-on-proof" fiscal grants, not discounts at the register. Applicants submit contracts, invoices and token-metering evidence through Jiangsu's unified compute-monitoring platform; funds settle after verification. Rules forbid stacking the same expense across similar provincial AI subsidies ("higher-of, no double claim"), and provincial money is capped and allocated by competitive score for the model voucher.

The plan also selects provincial AI industry clusters by competition, funneling fixed funds into public technology-service platforms — a second lever to pull smaller firms into the ecosystem. Contracts must clear the ¥10,000 floor, and rebates or refunds are excluded from the subsidy base.

Who can apply, and how the money splits

The vouchers are open to enterprises, universities, research institutes and OPCs (one-person companies). Provincial fiscal covers up to 50% of what each city actually pays out, with the rest from local matching — so the real cap a firm sees depends on its city's own ratio within the 50% ceiling. The model voucher is the tightest: it is allocated by competitive score across registration users, API daily calls and transaction value (weighted 30/30/40), and it runs out when the budget does. That means early, well-documented applicants fare better than late ones.

The stage — it is live

This is an implemented support policy, not a proposal. The notice was issued October 9, 2026, and the scheme runs through June 30, 2027. It builds on a Yangzhou city-level token voucher pilot now being scaled province-wide — a detail other provinces are watching closely as a test case for "AI vouchers" as industrial policy.

Honest limitations

We verified the departments, issue date, voucher caps, percentage rates and validity window against Jiangsu provincial-party and People's Daily-affiliated reporting plus a Shanghai securities outlet. We have not seen the full eligibility fine print — for example the exact credit-score weights used to rank model-voucher applicants — and all figures describe the published plan, subject to "total control" budget caps that could lower payout ratios if funds run short.

What readers can do now

  • If you are an AI firm in Jiangsu: register on the provincial compute-monitoring platform now and keep token-metering logs — the ¥10,000 minimum spend is low enough that even small pilots qualify.
  • If you sell compute or datasets: get your service listed so customers can apply vouchers against it; demand will tilt toward approved providers.
  • If you are in another province: watch Jiangsu's redemption rate; it is the clearest live test of "AI vouchers" as industrial policy and a likely template elsewhere.

Related coverage

Cite this article

NeuroAI (脑机网), "Jiangsu hands out three AI vouchers — token, corpus and model — up to ¥5M per firm," 27 October 2026, https://neuroai.site/a/na-policy-jiangsu-ai-vouchers

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version