NeuroAI NEUROAINEUROAI.SITE
ESC

Shanghai alone commits ¥10B in vouchers to cut AI compute bills

From Shanghai to Chengdu, cities issue 'computing vouchers' (算力券) covering up to 30% of AI compute rent, reshaping cost math for startups and global firms.

2026-10-05 · 850 words · NeuroAI
Shanghai alone commits ¥10B in vouchers to cut AI compute bills

A startup founder in Chengdu once shelved a model idea because the GPU bill looked like a year of payroll. A government coupon worth a slice of that bill is what brought it back.

Across China, that coupon now has a name: the "computing voucher" (算力券). It is a digital subsidy that pays part of the cost when a company rents intelligent or super computing. Quietly, a growing list of cities has turned it into a core tool of AI industrial policy.

The coupon that cuts the cloud bill

A computing voucher is not cash handed over the counter. It is a policy-backed digital certificate a qualifying firm applies for, then redeems against a real computing-service contract. The subsidy typically covers a fixed percentage of the bill — often 10% to 30% — up to an annual cap.

The point is to attack the single biggest barrier for small AI teams: compute is expensive and must be paid up front, before any product earns a cent. By absorbing part of the rental, the voucher lowers the cost of the first experiment.

Shanghai's ¥10 billion opening

The largest package so far comes from Shanghai. On July 28, 2025, the city's economy and information technology commission issued "Several Measures to Further Expand AI Application in Shanghai" (上海市进一步扩大人工智能应用的若干措施). It commits:

  • ¥6 billion (≈ US$845 million / HK$6.6 billion) in computing vouchers (算力券);
  • ¥3 billion (≈ US$423 million / HK$3.3 billion) in model vouchers (模型券) for using third-party large models;
  • ¥1 billion (≈ US$141 million / HK$1.1 billion) in corpus vouchers (语料券) for buying training data.

Together that is ¥10 billion (≈ US$1.41 billion / HK$11 billion) aimed squarely at the three inputs — compute, models, data — that drive AI cost. Renters of intelligent computing can receive up to a 30% city-level subsidy, and through city-district coordination some small and early-stage firms can reach far higher effective coverage.

A spreading map

Shanghai is the headline, but the pattern is national:

  • Chengdu — on August 21, 2025, published policies to build an AI industry highland, issuing up to ¥100 million (≈ US$14 million / HK$110 million) in computing vouchers each year, plus a subsidy covering up to 30% of annual model-usage fees, capped at ¥1 million.
  • Henan — on August 9, 2025, rolled out a computing-voucher settlement mechanism with an annual cap of ¥50 million (≈ US$7 million / HK$55 million).
  • Tianjin — on April 27, 2025, launched a trial computing-voucher plan covering 10% of a computing-service contract, capped at ¥2 million (≈ US$282,000 / HK$2.2 million) per firm per year.
  • Beijing — under its 2023–2025 AI computing-voucher plan (京经信发〔2024〕25号), firms can claim up to 20% of contract value, capped at ¥2 million (≈ US$282,000 / HK$2.2 million) annually.
  • Zhejiang / Hangzhou — runs a national AI voucher (国家人工智能券/算力券) funded by ultra-long special treasury bonds, with tiered 10% or 30% subsidies on genuine computing leases.

Chengdu actually pioneered the instrument back in 2023, and a 2024 NDRC, National Data Administration, Ministry of Finance, and Ministry of Human Resources opinion on the data-labeling industry explicitly encouraged localities to use "data vouchers, algorithm vouchers, and computing vouchers" to lower costs.

Who actually qualifies

The rules are deliberately narrow. Vouchers usually go to firms registered locally, with clean credit records, that rent compute from a non-affiliated supplier for real AI work — model training and inference, data annotation, analytics, engineering design. Pure storage, networking, or security services are excluded, and the same contract cannot be double-subsidized.

That targeting matters: the money is meant to seed usage, not to reward large incumbents that would buy compute anyway.

Why global readers should care

Two takeaways travel beyond China. First, the cost structure of AI experimentation is being actively shaped by government subsidy, not just by silicon supply. A startup in a voucher city competes with a different cost base than one that pays full rack rate.

Second, the vouchers tie directly into the western computing hubs described elsewhere on this site. Many voucher programs steer users toward domestic intelligent compute, including the national integrated computing network — so subsidy policy and compute-location policy are moving together.

What readers can do now

  • If your company operates a China entity, check the computing-voucher scheme in your city or free-trade zone; Shanghai, Beijing, Chengdu, Tianjin, and Zhejiang all run active programs with published caps.
  • Keep clean contracts: vouchers require an arms-length computing-service agreement, invoices, and proof of real usage — retrofitting paperwork later is painful.
  • Treat the voucher as a marginal cost cut, not a strategy; pair it with a realistic workload plan rather than renting capacity you will not use.

Honest limitations

Figures here come from municipal and provincial government releases and their professional summaries, not from an independent audit of disbursed amounts; announced caps are policy ceilings, and actual uptake may differ. Subsidy percentages and caps change by city and year, so the numbers should be confirmed against the latest local notice before any business decision. This article describes the policy design and does not assess individual providers, eligibility for foreign-owned entities, or tax-treatment questions, which require local professional advice.

Related coverage

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