--- title: "The Compute Voucher: How Chinese Provinces Subsidise AI Demand" date: 2026-10-20 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-compute-voucher language: en --- # The Compute Voucher: How Chinese Provinces Subsidise AI Demand > Provincial rebates cover 20–30% of a company's compute bill, but only on approved suppliers, so the subsidy steers demand as well as cutting cost. WUHAN — A founder with a laptop and a monthly model bill is now a category in Chinese fiscal policy. Hubei's new subsidy gives one-person companies their own tier, a higher reimbursement rate than a university gets, and a cap that fits a single desk. The instrument is a compute voucher (算力券). It is not a grant, not a loan and not an equity stake. It is a rebate on a bill already paid, which is why the paperwork matters more than the headline percentage. Hubei issued its scheme on 30 September. Liaoning issued its own rules in mid-October. Neither is a national programme, and that explains how the money behaves. ## Key takeaways - Hubei pays 20% of a company's compute bill, or 30% for domestically supplied compute, capped at RMB 3 million (≈ US$420,000) per entity a year. - Liaoning reimburses up to 30%, capped at RMB 2 million (≈ US$282,000), and requires annual compute spending of at least RMB 250,000 (≈ US$35,200). - Both schemes pay only for compute bought from suppliers registered on a provincial platform, so the rebate steers demand as well as cutting cost. - Hubei doubled its rate from 10% to 20% and tripled its cap from RMB 1 million to RMB 3 million (≈ US$141,000 to US$420,000) in one year. - The voucher is a reimbursement, not an advance, so it rewards firms that can carry a large bill for most of a year. ## What the rebate actually covers Read the eligibility clause before the headline rate. Hubei requires an applicant to spend at least RMB 30,000 (≈ US$4,200) a year on compute, low enough to include very small firms. Liaoning sets its floor at RMB 250,000 (≈ US$35,200), which excludes them. The two provinces run different policies under the same name. Hubei is buying breadth: it created a separate tier for one-person companies (一人公司), paying 30% up to RMB 100,000 (≈ US$14,100) a year. Liaoning is buying volume: fewer, larger claimants, and a shortlist of approved providers. Both are reimbursements. Hubei's window closes on 31 October and covers spending from the previous 1 November; Liaoning pays out once a year and lets unclaimed vouchers expire. A firm that cannot carry its own bill for most of a year will not benefit. ## The supplier list is the second subsidy Every voucher scheme is also a procurement rule. In Hubei, the provincial data bureau selects eligible compute providers annually, and a buyer cannot claim the rebate outside that list. In Liaoning, providers must register on the provincial platform with at least 50 PFlops listed. The instrument stops being neutral here. A rebate that only works on approved capacity is also a way of filling that capacity, and China has built a great deal of it: national data-centre capacity is forecast to pass 24 GW by the end of 2026, against roughly 56 GW in the United States (SemiAnalysis, 25 September 2026). The second steering wheel is the domestic premium. Hubei pays 30% instead of 20% when the compute is classified as domestically controllable (自主可控). Ten percentage points is enough to change a procurement decision without banning any supplier. Both provinces also bar claims on compute bought from an affiliate, so a data centre cannot subsidise its own subsidiary. ## Why the money moved to the buyer The visible subsidies of the past two years went to the builder: land, power, server halls and municipal equity funds. That phase produced capacity faster than it produced paying customers. Alibaba, Tencent and Baidu spent roughly US$20 billion on capital expenditure between them in the second quarter of 2026, more than double a year earlier, all three with negative free cash flow (SemiAnalysis, 25 September 2026). A voucher is what a government writes when the constraint has moved from supply to demand. It does not fund a new data centre. It funds the invoice that makes an existing one worth running. The instrument was blessed nationally and funded locally. A four-ministry plan for small-business digitalisation, issued on 12 December 2024, told local governments to support local exploration of cloud vouchers and compute vouchers (MIIT, Ministry of Finance, People's Bank of China and NFRA, 2024). Provinces then wrote their own rules, at their own scale, for their own suppliers. Guizhou's third iteration lifted the rate for intelligent-compute and model services from 3% to 30% and set aside RMB 140 million (≈ US$19.7m / HK$154m) (Guizhou Provincial Big Data Development Administration, 10 October 2026). ## What a voucher cannot do Three limits are visible in the documents themselves. - **It measures spending, not use.** A company that buys compute and leaves it idle still gets paid. Neither scheme ties the rebate to utilisation or to a shipped product. - **It fragments the market.** Rules differ by province, entry thresholds by a factor of eight, and a firm with offices in three provinces faces three processes. - **It is small relative to the bill.** Hubei's entire 2026 allocation is RMB 50 million (≈ US$7.0m / HK$55m), which at the 20% base rate implies about RMB 250 million of supported purchases across the province. The scale is a nudge with a filing cabinet attached, not a demand stimulus. The provinces are testing whether 20% off the bill moves a cautious manufacturer into production. ## Honest limitations The scheme documents are public; the disbursement data is not. Neither province's published notice states what has been paid out, so take-up is unknown. The domestic-compute premium is stated as a rate, not as a list, and the classification rules behind "domestically controllable" are not set out in the two documents reviewed here. The 24 GW and US$20 billion figures are estimates from a research firm, not audited statistics. No official has said that vouchers exist to fill underused capacity; that reading is an inference from the design, not a disclosed motive. ## What readers can do now - **Buyers:** if a supplier cites a voucher as part of its cost advantage, ask which provincial list it is on and when the cash lands. A rebate paid the following year is not a discount. - **Investors:** read the annual provider-selection notice from a provincial data bureau before the rate table. The list decides which data centres can pass the subsidy through. - **Analysts:** treat provincial allocations as ceilings, not spending. Until a province publishes redemption totals, the allocation measures intent. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-compute-voucher Free to quote with attribution and a link to the original.