Beijing Launches AI Factory: Targeting 100,000 P Flops and 10 Trillion Daily Tokens, DataCanvas Unveils New Strategy

In April 2026, DataCanvas officially launched its "AI Factory" strategy at the 2026 Global Intelligent Computing Technology Summit and Strategy Conference, aiming to restructure intelligent production and delivery through standardized, industrialized pipelines. The factory comprises two core components: the Training Factory (targeting 100,000 P flops) and the Token Factory (targeting 10 trillion daily tokens), and introduces the "Degree of Computing" (DCU) as a standardized unit for computing power.

Background: From "Answering" to "Executing" — The Engineering Gap

By March 2025, China's daily token calls had exceeded 140 trillion, growing over a thousandfold in two years. However, enterprises integrating large models face an "execution gap": general models excel in chat scenarios but frequently fail in real business systems (e.g., approvals, quality inspection, risk control). DataCanvas founder Fang Lei noted that the core of AI competition has shifted from technical excellence to industrial productivity, with the key being to "smelt" general models into specialized models capable of reliably executing tasks through reinforcement learning (RL).

Key Details: Four Components of the AI Factory

• Input Side: Degree of Computing (DCU)
  Definition: 312 TFlops×hour (312 trillion floating-point operations per second × 1 hour). It unifies heterogeneous chips, networks, and storage into a single metric, enabling computing power procurement to be billed by the "degree" like electricity.

• Production Side: Training Factory
  Target computing power scale: 100,000 P flops. It leverages five engineering capabilities (thousand-to-ten-thousand-card clusters, hybrid scheduling, network optimization, storage optimization, multi-tenancy) and a reinforcement learning training stack (supporting PPO, DPO, GRPO, etc.) to transform general models into specialized models for finance, manufacturing, government, and scientific research. DataCanvas has passed the China Academy of Information and Communications Technology (CAICT) standard evaluation for "Large Model Computing Resource Scheduling Platform," achieving a 100% improvement in training efficiency and a 50% increase in GPU utilization.

• Packaging Side: Token Factory
  Target daily production capacity: 10 trillion tokens. It packages specialized models into callable, measurable, and billable professional tokens. Tokens are defined as intelligent value units for business tasks, divided into three tiers: Consumer (daily AI applications), Professional (deep industry know-how), and Frontier (complex task automation and scientific breakthroughs).

• Output Side: Data Flywheel
  DCU measures input → Training Factory smelts models → Token Factory packages → Enterprises consume and return business data → Models iterate continuously, forming a virtuous cycle of "stronger with use, cheaper with strength."

Reactions and Data

• Enterprise Pain Points: Traditional AI development paths take 6–12 months with high upfront capital expenditure. Under the AI Factory model, enterprises can flexibly start from the Training Factory or Token Factory: large model companies can begin with training, while industry clients can start with inference, achieving proof of concept in as little as two weeks.
• Industry Impact: DataCanvas positions itself upstream of the application layer, providing the underlying intelligent production and delivery system for ISVs, integrators, and enterprise development teams, rather than directly developing agent applications.
• Target Benefits: 100,000 P flops, 10 trillion daily tokens, 1,000x comprehensive cost reduction, and plans to incubate 1,000 high-value models and intelligent applications.

Impact and Outlook

The launch of the AI Factory marks a shift in the business logic of intelligent computing: from "Is computing power sufficient?" to "Can computing power be transformed into deliverable intelligence?" Through standardized metering (DCU) and industrialized production (Training Factory + Token Factory), DataCanvas aims to lower the barrier for enterprise AI adoption, moving AI from the lab to the production line. In the future, the tiered pricing of professional tokens may reshape the pricing and delivery models of industrial intelligence.