In China, public procure
The Chinese State became in 2024 the first buyer of generative artificial intelligence technologies on its territory. According to Sinolytics data, 931 tenders were issued by state companies supervised by Beijing for the acquisition or deployment of large language models (LLMS). These public contracts alone represent more than 60 % of domestic demand in the Genai.
A substitution assumed to venture capital
This industrial policy is based on the observation that in a context of restrictions on access to foreign technologies and increasing prudence of private investors, public funds become the main AI dissemination engine. Where venture capital focuses on fast-monetization use cases (commercial chatbots, code assistants), the state puts on the long term, technological sovereignty and the priming of an internal mass market.
National champions like Deepseek, Moshot AI, Baichuan, are the first beneficiaries of this policy. According to internal sources relayed by Sinolytics, 45 % of the sies had already integrated a Deepseek model in their business systems in early 2025, mainly in back office or decision support.
Standardized use cases, pressure profitability
Projects funded via public procurement focus on strategic priorities set by Beijing:
- knowledge assistants (72 % of projects),
- automated code generation,
- data analysis,
- automation of administrative processes.
These uses, often similar from one entity to another, aim for rapid industrialization, but leave little room for radical exploration or innovation. The productivist logic dominates, the models must work, integrate quickly, and adapt to the constraints of large bureaucratic structures.
But this approach has a cost, to win the markets, many startups are forced to offer prices far below the commercial standards. Sector leaders now mention “victories at a loss”, necessary to stay in the state funding loop.
Sponsored incubators and shared infrastructure
Beyond the purchase of licenses, the State supports the structuring of a verticalized ecosystem. Structures like Mosu Space (Shanghai) or the Model Power Camp (Shenzhen) serve as catalysts: subsidized compute (up to 1000p of power provided by Sensetime), access to semi-public data games, mentoring of private actors like Huawei Cloud, and support by Chinese funds like Chinese funds IDG Capital Or Zhenfund.
The objective is to produce a constant flow of specialized applications, tested in a real environment, and directly reusable in public or parapublic entities.
To a localized technological block
This strategy also responds to the geopolitical imperative to reduce dependence on American suppliers, notably NVIDIA (for GPU) or OPENAI (for reference models). By massively subsidizing the sector, China intends to build an endogenous, interoperable technological block supported by a captive demand.
The result is a singular ecosystem, closed on itself, but coherent, capable of cushioning foreign sanctions by a strategic withdrawal on a internal market structured by the State.
