A Shenzhen founder watches the seed round shrink with every GPU billing cycle, long before the product is real. The city's answer is not a speech but a coupon: a piece of paper that turns public money into cheaper compute, cheaper data, and cheaper models.
On 18 December 2024, the Shenzhen Municipal Industry and Information Technology Bureau published the Measures for Building an AI Pioneer City (深圳市打造人工智能先锋城市的若干措施), stating the document was "issued for implementation (现印发实施)." This is a local industrial policy that is in force, not a consultation draft. It is built around vouchers — a mechanism Shenzhen calls by three names.
The policy, and its legal status
The measures run to four areas and eighteen specific actions. Their legal footing is two higher documents: the Guangdong provincial opinion on building a leading generative-AI hub (粤府〔2023〕90号) and the Shenzhen Special Economic Zone Artificial Intelligence Industry Promotion Regulation (深圳经济特区人工智能产业促进条例), a local regulation. The vouchers are capped annual budgets, not open-ended commitments — "up to (最高)" is the recurring word.
All figures below are official caps from the Shenzhen government text. RMB conversions use roughly 7.2 yuan to the US dollar and 1.09 yuan to the Hong Kong dollar, so treat them as approximate.
The voucher stack
Three coupons target the three things a model startup cannot live without.
- Compute vouchers (训力券) — up to RMB 500 million (about USD 69 million / HKD 545 million) a year. A company, university, or research institute renting intelligent compute for large-model training can get a subsidy of up to 50% of the contract, capped at RMB 10 million (about USD 1.4 million / HKD 10.9 million); for early-stage firms the rate rises to 60%.
- Corpus vouchers (语料券) — up to RMB 50 million (about USD 6.9 million / HKD 54.5 million) a year. Buying non-affiliated training data through a data-trading platform earns up to 30% back, capped at RMB 2 million (about USD 0.28 million / HKD 2.18 million). Organizations that open quality public-data corpora can win up to RMB 1 million (about USD 0.14 million / HKD 1.09 million) each.
- Model vouchers (模型券) — up to RMB 100 million (about USD 14 million / HKD 109 million) a year. Firms building services or agents on a state-filed generative model can reclaim up to 30% of model costs, capped at RMB 2 million.
Add the three pots and Shenzhen is putting up to roughly RMB 650 million (about USD 90 million / HKD 710 million) a year in direct vouchers aimed at the cost of building models.
Grants for real deployment
Vouchers cut build cost; grants push models into the world.
- Industry application — up to RMB 100 million (about USD 14 million / HKD 109 million) a year across advanced manufacturing, services, and research. Demonstration projects can take up to RMB 2 million; benchmark projects, up to RMB 10 million (about USD 1.4 million / HKD 10.9 million).
- Government scenarios — up to RMB 50 million (about USD 6.9 million / HKD 54.5 million) a year, with at least 50 application scenarios opened annually in city governance and public services; developers of adopted products can receive up to RMB 5 million (about USD 0.7 million / HKD 5.4 million).
- Frontier research — up to RMB 300 million (about USD 42 million / HKD 327 million) a year for basic research and breakthroughs in embodied intelligence, autonomous driving, and AI chips.
The voucher and grant ceilings together reach over RMB 1.1 billion (about USD 153 million / HKD 1.2 billion) a year.
A ten-billion fund behind it
Beyond annual coupons, the measures set up an AI industry fund targeting a RMB 10 billion (about USD 1.4 billion / HKD 10.9 billion) capital ecosystem, with a stated bias toward early, small, long-horizon, and hard-tech bets, plus loan interest subsidies and insurance support to lower the cost of capital.
The logic is a chain: cheap compute trains the model, cheap corpus feeds it, cheap model vouchers get it deployed, open government scenarios give it a first customer, and the fund finances the company through the gap.
Why coupons, not just rhetoric
Shenzhen is already a dense AI cluster, and the measures read as an attempt to keep that lead by attacking the three bottlenecks every founder names — compute, data, and the first paying user. Handing out vouchers rather than writing strategy papers means the support is claimed at the point of spend, which is harder to ignore than a slogan. For a startup, a 50% compute rebate changes the burn rate immediately.
Honest limitations
- It is local. The measures apply within Shenzhen. A firm registered elsewhere does not automatically qualify, and the amounts are caps, not entitlements — vouchers are applied for and competed for.
- The money is conditional. Subsidies ride on contract value, on using state-filed models, on buying through designated data platforms, and on delivery milestones. The percentages and ceilings are real but so are the strings.
- It is industrial policy, not a legal framework. This does not set safety duties or labeling rules; those come from national regulations. The Shenzhen text sits on top of, and assumes compliance with, the national layer.
- Budgets can shift. Annual caps depend on continued fiscal allocation; a future year could trim them without rewriting the document.
What readers can do now
- If you run an AI startup in Shenzhen, map your annual compute, corpus, and model spend to the three voucher pots now — the 50% compute rebate (60% for early-stage) is the largest single lever on your burn.
- If you sell models or data, get on the eligible lists (state-filed generative models, recognized data-trading platforms); without that status your customers cannot use their vouchers on you.
- If you watch China's AI industrial race, track whether other first-tier cities match Shenzhen's voucher levels — this is a measurable proxy for how seriously a city is subsidizing its model economy.