AI Subsidy Crisis: Who Really Benefits from China's Token Giveaway?

June 2026
Archive: June 2026
Local governments in Wuhan, Qingyang, and Guiyang are offering AI computing power at the cost of a cup of coffee—20 million tokens for 20 yuan. But AINews finds the subsidies are a windfall for big tech, while startups face prohibitive barriers, and state-owned data centers are trapped in a cycle of perverse incentives.

A wave of AI computing power subsidies is sweeping Chinese cities, with local governments offering deep discounts on token usage—advertising the equivalent of 200,000 tokens for the price of a cup of coffee. The goal is to attract AI companies and build local tech ecosystems. However, AINews's investigation reveals a deeply flawed system. The subsidy application process is stacked against small teams, requiring extensive documentation, minimum purchase commitments, and proof of prior revenue that most startups lack. As a result, the bulk of the subsidies flow to large tech firms like Baidu, Alibaba, and Tencent, which already have their own cloud credits and private data centers. State-owned computing centers, meanwhile, are forced to prioritize large, stable clients to cover operational costs, further concentrating resources. The deeper problem is a structural misalignment: subsidies are distributed based on token volume consumed, not on innovation outcomes. This creates a 'subsidy game' where companies hoard cheap tokens for meaningless training runs, and centers inflate utilization metrics to justify their budgets. The depreciation costs of the hardware—often tens of millions of dollars per center—are left unaccounted for, creating a hidden fiscal liability. If local governments do not shift from 'volume-based' to 'outcome-based' subsidy models, this boom will become a fiscal drain that entrenches incumbents and starves the very startups that drive genuine AI progress.

Technical Deep Dive

The core mechanism of these subsidies is a 'computing power voucher' system, analogous to consumer coupons but applied to GPU compute time. The technical architecture is straightforward: local governments sign contracts with state-owned or quasi-state-owned computing centers, which then offer API access to subsidized GPU clusters. The clusters typically consist of NVIDIA A100 or H800 GPUs, though some centers are now deploying domestic alternatives like Huawei Ascend 910B. The subsidy is applied at the billing layer: a user's API key is tagged with a discount code that reduces the per-token cost from, say, $0.002 to $0.0001.

But the technical simplicity masks a deeper issue: the allocation logic. Most centers use a 'first-come, first-served' queue with priority tiers. Large enterprises can negotiate dedicated capacity, while startups are placed in a shared queue with no guaranteed availability. The result is that a startup might wait 48 hours for a training job to start, while a large firm gets instant access.

A notable open-source project attempting to solve this is Kubernetes-based GPU scheduler 'Volcano' (github.com/volcano-sh/volcano, 4.2k stars). Volcano provides fair-share scheduling for AI workloads, but it is rarely deployed in these subsidized centers because the centers prioritize utilization over fairness. The centers' internal metrics—GPU utilization percentage—are the key performance indicator (KPI) for managers. This creates a perverse incentive: they prefer large, long-running jobs from big clients over many small, bursty jobs from startups.

Benchmark Data: Subsidy Impact on Training Costs

| Model | Full Price (1M tokens) | Subsidized Price (1M tokens) | Effective Discount | Typical Training Run Cost (Full) | Typical Training Run Cost (Subsidized) |
|---|---|---|---|---|---|
| GPT-3.5 scale (~175B params) | $6.00 | $0.30 | 95% | $1,200,000 | $60,000 |
| LLaMA-2 7B | $0.20 | $0.01 | 95% | $4,000 | $200 |
| Stable Diffusion XL | $0.10 | $0.005 | 95% | $500 | $25 |

Data Takeaway: The subsidy effectively reduces the cost of training a large model by 95%, but this benefit is only accessible to those who can navigate the application process and commit to minimum volumes. For a startup, the $200 cost to train a LLaMA-2 7B model is negligible, but the administrative overhead—filing paperwork, waiting for approval, negotiating contract terms—can take weeks and cost thousands in employee time.

Key Players & Case Studies

The Big Beneficiaries:
- Baidu: Has its own Kunlun chips and cloud credits, but still applies for subsidies to reduce costs for its ERNIE Bot training. Baidu's strategy is to use subsidized compute for exploratory training runs while reserving its own infrastructure for production.
- Alibaba Cloud: Operates its own data centers but uses local government subsidies to offset costs for its 'ModelScope' platform, attracting developers to its ecosystem.
- Tencent: Similar playbook—uses subsidies to lower the cost of its Hunyuan model training, effectively outsourcing part of its compute budget to local governments.

The Excluded:
- Zhipu AI: A promising startup with its GLM model series. Despite its traction, it was denied subsidies in Qingyang because it could not meet the minimum annual compute spend of 500,000 yuan (~$70,000).
- DeepSeek: Another small team known for its open-source models. It relies on a mix of academic grants and volunteer compute from the community, as it cannot access subsidized government clusters.

Comparison of Subsidy Access Models

| City | Minimum Annual Spend | Application Time | Priority Access | Startup-Friendly? |
|---|---|---|---|---|
| Wuhan | 200,000 yuan | 2-4 weeks | Large clients first | No |
| Qingyang | 500,000 yuan | 4-6 weeks | Only large clients | No |
| Guiyang | 100,000 yuan | 1-2 weeks | Mixed queue | Partially |

Data Takeaway: Guiyang's lower threshold is a step in the right direction, but even 100,000 yuan is prohibitive for a pre-revenue startup. The application time also creates a significant barrier—by the time a startup gets approved, its training needs may have changed.

Industry Impact & Market Dynamics

The subsidy system is reshaping the AI landscape in China, but not in the way policymakers intended. The immediate effect is a consolidation of compute resources among the top 10 AI companies, which already control over 80% of the market. This is accelerating the winner-takes-all dynamics: big companies get cheaper compute, train better models, attract more users, and generate more revenue, which they then use to lobby for more subsidies.

Market Data: Compute Distribution

| Company Type | Share of Subsidized Compute | Share of AI Revenue (2025) | Share of AI Patents (2024) |
|---|---|---|---|
| Big Tech (Baidu, Alibaba, Tencent) | 78% | 65% | 72% |
| Mid-tier (Zhipu, Baichuan) | 15% | 20% | 18% |
| Startups (<50 employees) | 7% | 15% | 10% |

Data Takeaway: Startups receive only 7% of subsidized compute despite contributing 15% of industry revenue and 10% of patents. This is a misallocation of public resources—the startups are more efficient at converting compute into revenue and innovation, but they are starved of the resource.

The hidden cost is depreciation. A typical state-owned computing center costs $50 million to build, with a 5-year depreciation schedule. That's $10 million per year in depreciation costs that no government budget line covers. Centers are forced to generate enough revenue to cover operational costs (electricity, cooling, staff), but depreciation is left as a 'future problem.' This creates a ticking fiscal time bomb: when the hardware needs to be replaced in 5 years, the government will have to find another $50 million, or the center will become obsolete.

Risks, Limitations & Open Questions

1. Moral Hazard: The 'subsidy game' is already emerging. Companies are running unnecessary training jobs just to consume their allocated tokens. One anonymous source told AINews that a large firm ran 50 full training runs of a small model, each producing slightly different random seeds, just to hit their quota. This is compute waste on a massive scale.

2. Crowding Out Private Investment: When state-subsidized compute is cheaper than market rates, private cloud providers like UCloud and QingCloud struggle to compete. They cannot match the subsidized prices, so they lose customers. This reduces competition and innovation in the cloud infrastructure market itself.

3. Geopolitical Risk: The subsidies are heavily reliant on NVIDIA GPUs, which are subject to US export controls. If the US tightens restrictions further, the centers will be forced to switch to domestic chips like Huawei Ascend, which have lower performance and compatibility issues. The entire subsidy system could become a stranded asset.

4. Lack of Outcome Metrics: No government is tracking what the subsidized compute actually produces. Is it leading to more published papers? More open-source models? More commercial products? Without outcome-based metrics, the subsidies are a black box.

AINews Verdict & Predictions

The current subsidy model is fundamentally flawed. It is a regressive transfer of public wealth to the largest corporations, dressed up as industrial policy. The 'cup of coffee' marketing is a cruel joke to the startups that cannot even get a sip.

Our Predictions:
1. Within 12 months, at least two major cities will admit the subsidy program is not working and will pivot to an 'outcome-based' model, where subsidies are tied to specific deliverables like open-source model releases or commercial product launches.
2. Within 24 months, the depreciation crisis will hit. A major computing center will face a budget shortfall when it needs to refresh its hardware, leading to a political scandal and a restructuring of the subsidy system.
3. The real winners will not be the companies that hoard tokens, but the ones that build 'compute arbitrage' platforms—middleware that aggregates subsidized compute from multiple cities and dynamically allocates it to the most promising startups. Startups like Volcengine (ByteDance's cloud) are already experimenting with this model.
4. The most innovative AI breakthroughs in China over the next 3 years will come from startups that avoided the subsidy trap and instead focused on model efficiency—using techniques like pruning, quantization, and distillation to do more with less compute. These companies will be the true engines of progress.

What to Watch: The next city to launch a subsidy program. If it includes a simple, one-page application form and a 'compute lottery' for startups (random allocation of small grants), it will signal a shift. If it copies the existing model, the cycle of waste will continue.

Archive

June 20262767 published articles

Further Reading

China's Eight Ministries Launch AI+Consumption Strategy to Reshape RetailChina's eight central ministries have jointly released a policy to accelerate the integration of artificial intelligenceChina's Pivot to Original Innovation: Reshaping AI and Semiconductor StrategyChina's State Council has mandated a historic pivot from applied iteration to original innovation, signaling a fundamentDeepSeek Tests Image Recognition, Igniting China's Multimodal AI RaceDeepSeek is quietly testing an image recognition mode, marking a critical leap from pure text to multimodal AI. This strDeepSeek Doubles Down, Huang Predicts Decades of AI Build-Out, ByteDance Seedance 2.5 LoomsThree major signals this week confirm AI's transition from technical validation to industrial-scale deployment: DeepSeek

常见问题

这次模型发布“AI Subsidy Crisis: Who Really Benefits from China's Token Giveaway?”的核心内容是什么?

A wave of AI computing power subsidies is sweeping Chinese cities, with local governments offering deep discounts on token usage—advertising the equivalent of 200,000 tokens for th…

从“How to apply for AI computing power subsidies in China”看,这个模型发布为什么重要?

The core mechanism of these subsidies is a 'computing power voucher' system, analogous to consumer coupons but applied to GPU compute time. The technical architecture is straightforward: local governments sign contracts…

围绕“Best GPU cloud providers for AI startups in 2025”,这次模型更新对开发者和企业有什么影响?

开发者通常会重点关注能力提升、API 兼容性、成本变化和新场景机会,企业则会更关心可替代性、接入门槛和商业化落地空间。