Beijing AI Factory Project Launched: Targeting 100,000 PetaFLOPS and 10 Trillion Tokens Daily Output

In April 2026, Jiuzhang Yunji officially unveiled its "AI Factory" strategy at the 2026 Global Intelligent Computing Technology Summit and Strategy Launch, aiming to bridge the engineering gap from large model R&D to industrial deployment through standardization and industrialization. The project includes two core facilities: a Training Factory targeting 100,000 PetaFLOPS of computing power, which uses reinforcement learning to "refine" general-purpose large models into specialized models for finance, manufacturing, government, and other sectors; and a Token Factory targeting a daily output of 10 trillion tokens, packaging specialized models into callable, metered intelligent services. Jiuzhang Yunji also pioneered the "Degree of Computing Unit" (DCU) metric (312 TFlops×hour), enabling computing power procurement to be billed per degree like electricity. The AI Factory aims to achieve a 1000-fold comprehensive cost reduction and incubate 1,000 high-value models and applications.

Background: From Computing Power Competition to Industrialized Production

As of March 2026, China's daily token call volume exceeded 140 trillion, growing over a thousandfold in two years, but enterprises generally face the dilemma that "models can answer but cannot execute." General-purpose large models perform well in chat scenarios but frequently err in complex business processes (e.g., approvals, quality inspections). Jiuzhang Yunji founder Fang Lei pointed out that the core of intelligent competition has shifted from technological excellence to productivity industrialization, and the AI Factory is precisely designed to address this contradiction.

Key Details: How the AI Factory Operates

The AI Factory forms a closed loop through four key stages:

• Input Side: Pioneering the "Degree of Computing Unit" (DCU), which standardizes heterogeneous GPUs, NPUs, networks, storage, and other resources into 312 TFlops×hour. Customers pay per degree, achieving standardized computing power procurement.
• Production Side: The Training Factory leverages five engineering capabilities (elastic computing, hybrid scheduling, network optimization, storage optimization, multi-tenant queuing) and a reinforcement learning training stack (supporting PPO, DPO, GRPO, etc.) to "refine" general models into specialized ones. The Training Factory has passed the China Academy of Information and Communications Technology (CAICT) standard evaluation, with training efficiency improved by 100% and GPU utilization increased by 50%.
• Packaging Side: The Token Factory packages specialized models into callable, metered, and operable professional tokens, divided into three tiers: consumer-grade, professional-grade, and frontier-grade. Professional-grade tokens target scenarios like financial risk control and quality inspection, offering a defined ROI.
• Output Side: After consuming tokens, enterprises return business data, driving continuous model iteration, forming a flywheel of "stronger with use, cheaper with strength."

Reactions and Data

• Industry Pain Points: Traditional enterprise AI construction cycles last 6-12 months, requiring self-built clusters and operations teams with high upfront capital expenditure. Under the AI Factory model, enterprises can start from any stage—training or inference. Large model companies begin with the Training Factory, while industry clients directly call services from the Token Factory, achieving proof of concept in as little as two weeks.
• Technical Validation: Jiuzhang Yunji has become the first to pass the CAICT "Large Model Computing Resource Scheduling Platform" standard evaluation, covering all 81 capability assessments.
• Market Data: As of March 2026, China's daily token call volume exceeds 140 trillion, but companies like Amazon have found that surging token usage does not bring efficiency gains, highlighting the need for professional tokens.

Impact and Outlook

The launch of the AI Factory marks a shift in the business logic of intelligent computing clouds: from simply providing computing power to delivering quantifiable intelligence. Jiuzhang Yunji positions itself upstream of the application layer, providing ISVs, integrators, and enterprises with an underlying intelligent production and delivery system. If the targets are met, 100,000 PetaFLOPS of computing power and 10 trillion tokens daily output will significantly lower the barrier to enterprise AI adoption, accelerating industrial intelligence. However, large-scale industrialization of reinforcement learning still faces engineering challenges such as stability of ten-thousand-card clusters and automation of reward functions. Whether the Training Factory can consistently deliver on its promises remains to be tested by the market.