In May 2025 the Beijing Municipal Bureau of Economy and Information Technology issued its 2025 action plan for AI-enabled industrial modernisation. The plan places its support in two places: the compute cost of industry models, and general-purpose agents that can be replicated across an industry.
Why the subsidy goes to compute cost
The bottleneck in bringing large models into industrial settings is often cost rather than algorithm: the compute bill for training and inference is hard for smaller companies to carry. For industry models that reach a leading domestic or international level, the plan offers support of up to RMB 30 million against compute cost — in substance, absorbing the cost of trial and error.
This logic matches MIIT’s direction on AI plus manufacturing. See our notes on the State Council opinion on the AI plus initiative and on the AI plus manufacturing action plan.
Why general-purpose agents are singled out
The plan also supports the operation of general-purpose agents with industry-wide reach, again against the compute cost of calls. Supporting agents separately from models reflects a judgement: models supply capability, agents supply deliverable service. What policy wants is the party that can enter a production process and be reused.
This approach is consistent with the wider direction of agent-related policy. See our note on the opinion on the development of agent applications.
Echoing the conference calendar
Around the plan’s release, policy moves on AI and manufacturing kept coming. See our note on the 2025 AI industry and industrialisation conference.
Where companies can start
First, translate the subsidy threshold into an application roadmap, identifying what evaluation or scenario evidence a “leading domestic or international level” requires.
Second, separate compute subsidies from scenario subsidies; the former targets cost, the latter application, and the application materials differ.
Third, assess how compute spending interacts with data compliance. Compute investment often comes with data concentration, and data processing rules need checking at the same time. See our notes on the national integrated compute system and large models and on the action plan for building high-quality industry datasets.
The value of a local subsidy lies not in the amount but in the direction it shows policy endorses. Those who can read that direction into a product roadmap are the ones who get support.
This article is based on the Beijing Municipal Bureau of Economy and Information Technology’s 2025 Action Plan for AI-Enabled Industrial Modernisation and public information and is for reference only; it does not constitute legal advice.