NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

From US$0.42M to US$1.4M: Shenzhen more than triples its AI training-compute voucher

Shenzhen's 2025 "training-compute voucher" (训力券) raises the annual cap to 10 million yuan for AI model training and inference, up from 3 million in 2024.

2026-11-12 · 968 words · NeuroAI
From US$0.42M to US$1.4M: Shenzhen more than triples its AI training-compute voucher

A small team in Shenzhen trains a Chinese-language legal model, but the GPU bill hits before the model is any good. Across the city, a robotics lab needs inference capacity it can never justify buying outright. Both face the same wall: smart compute (智能算力) is cheap to need and expensive to own. On March 25, 2025, Shenzhen opened a doorway through that wall — and made it more than three times wider than a year earlier.

Key takeaways

  • Shenzhen launched its 2025 "training-compute voucher" (训力券) application round on March 25, 2025, with filings due by April 25, 2025.
  • The annual cap per applicant rose to 10 million yuan (≈ US$1.4M / HK$11M), up from 3 million yuan (≈ US$0.42M / HK$3.3M) in 2024.
  • Vouchers offset up to 50% of a qualifying smart-compute service contract, or 60% for companies less than a year old; the floor to qualify is 500,000 yuan (≈ US$70,000 / HK$550,000) of annual spend.
  • The vouchers cover contracts signed between January 1, 2025 and December 31, 2026 with approved providers.
  • It is a targeted subsidy with gates: service providers must clear a 300 PFLOPS (FP16) capacity bar and hold the right telecom permits.

What the voucher actually buys

The "training-compute voucher" (训力券) is a recharge-style subsidy issued by the Shenzhen Science and Technology Innovation Bureau (深圳市科技创新局). Recipients do not get cash; they get a coupon they can spend with an approved local smart-compute provider to rent GPUs for large-model (大模型) training and inference. The work it covers is specific: pretraining, supervised fine-tuning, reinforcement learning, and validation inference of AI models.

The money is real but capped and conditional. For most applicants, a single voucher redemption can cover no more than 50% of one service contract's value. A company founded less than a year before the 2025 guide was published can claim up to 60%. Either way, the applicant must first spend at least 500,000 yuan (≈ US$70,000 / HK$550,000) of its own money on qualifying compute in a year before any voucher applies. The annual ceiling per applicant is 10 million yuan (≈ US$1.4M / HK$11M), and unused quota from the prior year rolls into the current year's cap.

Who qualifies, and who doesn't

On the demand side, eligibility is narrow: a legally registered, independent legal-person entity operating in Shenzhen (including the Shenzhen–Shanwei cooperation zone) — a company, university, or research institute. Applicants must be clean on research-integrity and credit blacklists, and the same project cannot already be drawing more than 50% of its contract value from other government funds.

On the supply side, the bar is technical. To join the provider registry, a smart-compute operator needs the Ministry of Industry and Information Technology's IDC/ISP telecom license, a Shenzhen presence, and the ability to offer at least 300 PFLOPS of intelligent compute measured at FP16, with metering logs that track actual usage. In 2024, only 11 providers made the cut — among them Huawei Cloud's Shenzhen unit, Cloudwalk (云天励飞), the Pengcheng Laboratory, and the National Supercomputing Center in Shenzhen (国家超级计算深圳中心).

There is one hard exclusion worth flagging: an entity cannot be both a voucher-claiming user and a registered provider in the same round.

Why Shenzhen is doing this

Shenzhen has been rolling out a stack of "vouchers" — for compute, for corpora (语料券), and for models (模型券) — as a way to lower the cash barrier to building AI products without the government picking winners or owning data centers. The compute voucher is the most concrete of the three so far.

The raise from 3 million to 10 million yuan signals that demand, not just intent, is the bottleneck the city wants to remove. High-performance compute is consumed by everything from autonomous driving and film rendering to biomedicine and smart-city systems, and the city would rather see that demand met by local providers than by teams quietly renting overseas capacity.

How it fits the bigger subsidy stack

The 2025 voucher is the second year of a program that began in 2024. It sits alongside the city's broader AI push — including an action plan for AI terminals (2025–2026) and an embodied-intelligence robotics plan (2025–2027) — where the compute voucher is explicitly named as the mechanism to cut model-training costs.

The application itself is unglamorous but strict: file the contract through the city's science-technology business system, pass an electronic review, a commissioned check, a public notice, and final approval. The bureau warns that using a paid agent to write the application gets it rejected at the formality stage, and false claims land the applicant on a research-integrity watchlist.

Honest limitations

This article covers only the 2025 Shenzhen training-compute voucher round as published by the Shenzhen Science and Technology Innovation Bureau; it does not cover the separate corpora or model vouchers, the city's other AI subsidies, or programs in other cities. Figures come from the official 2025 application guide and municipal news coverage. Exact approval rates, how many vouchers were actually redeemed in 2024, and the full size of the 2025 applicant pool are not published in the sources reviewed, so they are not stated here. Details such as the per-project funding mix can change in later guidance, and we did not model the effective discount after taxes or the time cost of applying.

What readers can do now

  • If you run an AI team in Shenzhen, track the next voucher round early — the 2025 filing window closed on April 25, 2025, so 2026 planning should start well before spring.
  • Compare the voucher's 50% (or 60%) offset against a direct cloud contract; the 500,000 yuan (≈ US$70,000 / HK$550,000) floor means it only helps teams with real, recurring compute bills.
  • If you operate compute infrastructure in or near Shenzhen, review the 300 PFLOPS (FP16) and IDC/ISP license requirements now, because provider-registry openings are short and capped.

Related coverage

Cite this article

NeuroAI (脑机网), "From US$0.42M to US$1.4M: Shenzhen more than triples its AI training-compute voucher," 12 November 2026, https://neuroai.site/a/na-policy-shenzhen-training-voucher

Machine-readable: BibTeX · RIS · JSON · Markdown

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

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories