Chinese local governments (Wuhan, Qingyang, Guiyang) wage AI compute subsidy bidding war with Token credits and cash incentives.

When computing power becomes a consumer voucher—exploring who benefits and who pays in the Token subsidy boom

In the Nantaizi Lake Innovation Valley of Wuhan Economic Development Zone, an AI startup founder ordered a coffee and received 200,000 Tokens' worth of free computing power.

What can 200,000 Tokens do? Roughly equivalent to generating tens of thousands of words of text on mainstream large language models, or processing dozens of documents—with commercial value between $0.50 and $1.50. This is less a substantive subsidy than a tangible signal: computing power has become a new bargaining chip in local government business recruitment.

In March 2026, Liu Liehong, director of the National Data Bureau, officially designated "词元" (Token) as the Chinese term, defining it as the value anchor of the intelligent age and the settlement unit connecting technological supply with commercial demand.

One concept requires clarification: strictly speaking, computing power is the commodity, and Token is merely the unit of measurement for computing power consumption. Tokens issued by local governments are essentially computing power deduction quotas for designated local computing platforms, not standardized assets directly convertible to cash or freely tradeable. Tokens currently serve only as billing units for AI services and are strictly prohibited from securitization, tokenization, and public trading or monetization.

Previously, local government recruitment relied on providing land and tax incentives. Now a new bargaining chip has emerged. In 2026, a nationwide competition centered on Tokens is quietly unfolding across China.

According to IDC's "China AI Computing Power Market Tracking Report" and public data from Alibaba Cloud and Huawei Cloud (measured by public cloud API call volume, excluding private deployments), China's daily average Token call volume was approximately 100 billion in early 2024; by the end of 2025 it jumped to 1 quadrillion; by March 2026 it had surpassed 1.4 quadrillion—a growth exceeding a thousandfold over two years. Token call volume has become a key metric for measuring AI model activity and industrial value.

Token spans the "energy-chip-computing power-model-application" five-layer value chain, becoming the core benchmark for measuring AI industry vitality. A representative from Wuhan's Economic and Information Technology Bureau AI Industry Division stated plainly: "During the AI industry's rapid development phase, everyone starts from the same starting line. Whoever is early, fast, and proactive is most likely to seize new opportunities."

Yet Token consumption does not equal value creation. A customer service bot consuming a million Tokens and a code generation tool consuming 100,000 Tokens may see the latter generating greater commercial value. Token measurement quantifies only volume, not quality. Without high-quality data training, even the cheapest Tokens can only produce low-quality content or hallucinations. Current Token subsidies primarily cover computing power consumption, but subsidies for acquiring high-quality corpora, data annotation, and compliant dataset construction remain scarce—meaning Token usage costs have fallen, but the value of generated outputs has not necessarily risen proportionally.

Qingyang, Gansu is a typical case. As one of eight national-level hub nodes in the "Eastern Data, Western Computing" project, Qingyang enjoys electricity prices as low as 0.398 yuan per kilowatt-hour, with data center PUE values consistently below 1.2. According to public data from Qingyang's municipal government, it has connected (through signed letters of intent or site inspections) with 8,126 digital economy enterprises, signed agreements with 1,760, and registered 571. Leading models including MiniMax, Kimi, Zhipu, and DeepSeek have all deployed training operations in Qingyang—it has become a core computing power supply center for the Kimi large language model.

In January 2026, the Maize Zhiyun 10,000P Green Intelligent Computing Center project, with a total investment of 20 billion yuan, was signed and landed in Qingyang, planning to deploy 200,000 servers with 8 to 10 ten-thousand-GPU clusters, achieving core computing capacity of 100,000 PFlops.

In April 2026, Wuhan's Jianghan District convened a Token Economy Conference, announcing an all-out effort to build the "first Token economy district." Jianghan District coordinated 50 million yuan in annual special-purpose funds and rolled out a combination of measures: (1) 10 million yuan in annual Token vouchers to subsidize enterprise Token usage costs; (2) 10 million yuan in annual computing power vouchers to support enterprise priority application of green computing; (3) 5 million yuan in annual model vouchers supporting large model and agent development; (4) 5 million yuan in annual data vouchers supporting data product procurement and trading; (5) 10 million yuan in annual scenario funding for selecting benchmark application scenarios; (6) 5 million yuan annually supporting online Token service platform construction; (7) 5 million yuan annually supporting offline OPC (AI One-Person Company) ecosystem communities.

According to Jianghan District government public information, through green electricity and computing power subsidies, enterprises are expected to reduce computing costs by 30-40%. By 2028, Jianghan targets the Token economy-related industry scale to exceed 20 billion yuan, aggregating over 200 enterprises.

Wuhan at the municipal level made an even bigger move—announcing the distribution of 100 million yuan in computing power vouchers, with annual usable computing power reaching 15,000 PFlops, enabling Token-based services through the computing power public service platform accessible to universities, research institutions, enterprises, and individual entrepreneurs in OPC communities.

Guiyang, leveraging its foundation as China's data hub, began distributing 140 million yuan in computing power vouchers nationwide starting April 2026, with eligible enterprises enjoying computing subsidies up to 30%, coordinating efforts across five dimensions: computing power, data, models, business formats, and scenarios.

Baoding, leveraging its advantages as a national data annotation base, is simultaneously developing Token economy policies. In policies released in March 2026, enterprises purchasing over 300,000 yuan in annual computing power receive up to 300,000 yuan in digital intelligent voucher subsidies.

The nationwide computing power recruitment competition is intensifying. Shanghai's annual computing power voucher distribution reaches 600 million yuan, Wuhu offers enterprise computing subsidies up to 1 million yuan, and Guangdong, Anhui, and other regions are following suit. Computing power subsidies have diffused from "Eastern Data, Western Computing" hub nodes to major digital economy strongholds, becoming a standard tool in local government industrial recruitment.

Multiple industry insiders point out that many confuse computing power and Token. The vast majority of AI enterprises' cost pressures stem from Token consumption, yet they receive computing power vouchers they cannot use. One expert notes that in the artificial intelligence era, every Token's movement is traceable—who provided computing power, how much was used, how many resources the model consumed, all crystal clear. Technically there are no obstacles; the problems stem from policy design—the deeper contradiction lies in misalignment between policy supply and market demand. Some policies sound attractive but are difficult to implement because they divorce from enterprises' actual business scenarios; others set high application thresholds and tedious processes, excluding small and medium enterprises needing support.

Synthesizing computing power voucher application rules across regions, most areas establish revenue and team size entry thresholds. Taking Wuhan Jianghan District's publicly released computing voucher application details and feedback from multiple interviewed AI entrepreneurs, applicants must have prior-year revenue of at least 5 million yuan or teams of at least 20 people; the vast majority of sub-10-person teams lack basic eligibility. Even when meeting conditions, subsidy vouchers are bound to local computing platforms, which adapt less effectively to mainstream inference frameworks like PyTorch and vLLM than the cloud services enterprises originally used; additional migration and adaptation labor costs could already offset Token price advantages.

Industry warnings note that current online verification mechanisms for computing consumption lack full coverage, creating regulatory gaps for fabricating consumption data to fraudulently claim subsidies. Multiple regional financial audit reports have already flagged risks of enterprises registering elsewhere and artificially inflating Token consumption through computing platforms to capture policy benefits under locality-bound subsidy models.

Local governments subsidize the demand side but not the supply side. According to Liaowang News Weekly, a state-owned enterprise official disclosed that serving government recruitment was an important mission of their computing center. "Governments want to offer cheaper computing to attract enterprises; private computing enterprises are unwilling, so only state-owned enterprises do it—effectively state enterprises absorb government recruitment costs. Data center construction and operation expenses are very high; government subsidies go only to computing users, not to us. We can barely break even on operations and cannot cover high depreciation costs."

The overlooked side: the silent cost to traditional industries. As fiscal resources tilt toward the new AI track, subsidies for traditional industry digitalization are diverted. Third and fourth-tier cities lacking computing, data, and scenario advantages risk wasteful fiscal resource allocation if they blindly follow suits with computing subsidies, struggling to achieve comprehensive industry transformation and upgrading.

General-purpose inference computing will face continued price declines due to local subsidy competition, but high-precision training computing, private deployment computing, and secure-compliant computing remain scarce and expensive. Most subsidy benefits go to leading large model enterprises; smaller startups with seemingly cheap Tokens face hidden costs—computing resources squeezed by major vendors, peak-hour queues, inflated computing specs, and service degradation.

Deeper risks include quantity-quality mismatch. With current uniform Token pricing by call volume, identical Token consumption sees equal pricing for high-commercial-value scenarios and casual chat, potentially triggering adverse selection—enterprises deliberately compress prompts and reduce model output quality to economize Token costs rather than optimize commercial value. This reflects a fundamental flaw in market pricing mechanisms.

Currently, no national unified Token measurement and pricing standards exist. Different model architectures, precisions, and deployment methods have inconsistent Token consumption rules; private deployment Tokens don't enter public statistics and cannot be included in unified pricing; whether green-compute and compliant-data-trained Tokens receive differential pricing has no clear rules.

A deeper technical barrier: different vendors' models have inconsistent Token segmentation rules. The same Chinese text consumes vastly different Token quantities across GPT-4, domestic models, and even different versions from the same vendor—potentially differing by factors of several times. If Token cannot even ensure measurement consistency across models, the foundation of a unified settlement unit contains fundamental defects.

An even more fundamental issue is chip ecosystem fragmentation. Computing centers across regions extensively adopt domestic AI chips like Ascend, yet mainstream model training and inference frameworks (PyTorch, vLLM, Triton) show uneven adaptation to domestic chips. The same text may differ by several times in energy consumption and latency generating Tokens on NVIDIA GPUs versus domestic chips. Without unified effective computing conversion rates, Tokens cannot become a cross-platform, cross-architecture universal measurement unit.

Without regional coordination, market fragmentation and high transaction costs will obstruct cross-regional free flow of Token elements. Practitioners in the data elements industry suggest the national level should explore establishing unified Token measurement and pricing standards, perfecting Token economy governance systems.

With annual tens of millions in subsidies per region, if AI industry taxes cannot long cover subsidy expenditures, policy reductions are inevitable. Once subsidies end, startups dependent on cheap Tokens face batch elimination risks. Some analysts warn that the current AI industry runs the risk of high profile with few closed loops; many AI enterprises' revenue growth still relies on capital expenditure expansion, yet to form stable, repeatably purchasable business models.

Historical precedent from photovoltaic and new energy vehicle industries shows both adopted phased subsidy reduction plans—high early subsidies, gradual annual reductions mid-term, complete exit later, coupled with technical thresholds (like photovoltaic conversion efficiency, EV driving range) forcing enterprise upgrades. To avoid dying upon subsidy withdrawal, Token subsidies require similar institutional design—time limits, phased reductions, and technological progress assessments.

Token economics' emergence is reshaping China's computing power map. Yet assessments of western energy converting to Token exports warrant caution.

Western opportunities genuinely exist. Regions like Qingyang and Baotou convert wind and solar resources into computing services priced in Tokens. China Telecom Ningxia Branch launched Token factory generation capacity service procurement in April 2026, with framework agreement scale reaching 17.438 billion yuan. This procurement doesn't directly purchase standardized Token assets but rather conducts large-scale centralized bidding for computing services and model-calling capabilities unified under Token billing.

Yet western limitations are equally evident. Wind and solar face consumption fluctuations and seasonal supply instability; green computing cannot provide all-day stable supply for large-scale model training. The west can only undertake offline large model training; high-frequency real-time inference and agent applications still require eastern low-latency computing. The "Eastern Data, Western Computing" project already includes cross-regional computing dispatch rules; if localities establish regional computing barriers through Token vouchers, this may conflict with top-level whole-nation integrated computing dispatch design.

Top-level design is keeping pace. From the "Eastern Data, Western Computing" project to national integrated computing network construction, from high-quality dataset development to data elements market reform, policy chains are gradually taking shape. Yet balancing local recruitment competition with national element free flow remains an unresolved challenge.

Many enterprises train models on remote public computing platforms; enterprise commercial datasets and user prompts travel cross-regionally with Tokens. Data export, privacy leakage, and training dataset copyright infringement compliance risks have yet to enter local computing subsidy pre-approval screening. Large-scale data breaches or copyright disputes could impact not only affected enterprises but also challenge the entire Token economy's legitimacy. This institutional dimension remains blank.

From "Eastern Data, Western Computing" to Token vouchers, computing power is transforming from national strategic infrastructure into local government recruitment leverage. This is not a race all cities can join. Qingyang relies on western-to-east computing plus ultra-low electricity prices; Wuhan relies on industrial foundation plus clean energy; Guiyang relies on years of data industry accumulation. More third and fourth-tier cities lacking green power, data, and AI scenarios face high-probability fiscal waste following blindly.

Whether Tokens become universally shared resources like hydropower or fall to monopoly by few computing hub cities depends on three variables: when unified measurement rules launch, how subsidy exit mechanisms are designed, and how cross-regional circulation is achieved.

Healthy Token economy development requires three layers of institutional constraints: first, the nation issue cross-model, cross-chip-architecture unified Token measurement national standards; second, each region bind computing subsidies to enterprise local tax, employment, and intellectual property landing performance, establishing 3-5 year phased subsidy reduction mechanisms; third, construct a nationwide integrated computing voucher universal redemption platform, breaking local computing property barriers, enabling Token elements to flow freely.

The real challenge is not choosing market or subsidy, but designing systems letting both coexist—letting markets discover prices, subsidies correct market failures, while avoiding subsidy distorting price signals. This balance point's discovery is only just beginning.