At the recently concluded 'China Wealth Management 50 Forum 2025 Annual Conference', the convergence of AI and finance emerged as a central topic of discussion.

Zhang Chi, Vice President of Alibaba Cloud Intelligence Group, articulated during a forward-looking dialogue on 'Emerging AI Trends and Industrial Synergy in 2026' that AI is propelling global financial institutions towards a digital and intelligent transformation, transitioning from an efficiency-driven revolution to a fundamental restructuring of business models.

He forecasted a pivotal transformation in the landscape of AI computing power requirements: inference computing power is poised to overtake training computing power, becoming the predominant force in the near future.

In response to this trend, Zhang Chi advocated for an evolution in the core paradigm of AI applications within the financial sector, shifting from isolated tools to a comprehensive ecosystem. He emphasized that successful financial AI agents are not merely about traffic acquisition or external integrations; the crux lies in establishing a synergy system where a macro flywheel drives intent comprehension and a micro flywheel executes tasks.

This 'dual-wheel' architecture is designed to facilitate deep integration and collaboration between AI and professional workflows, transitioning AI from a supportive role to a central one, ultimately enabling the widespread implementation of production-grade scenarios through holistic solutions. This insight offers financial institutions a clear roadmap for strategizing their AI deployments during the '15th Five-Year Plan' period.

Zhang Chi's assertion that 'inference computing power will surpass training computing power' underscores the inevitable trajectory as AI implementation progresses into more mature phases.

This shift is underpinned by the growing maturity of foundational large model technologies and their entry into widespread application deployment. As industries, particularly high-value, high-concurrency sectors like finance, begin to extensively integrate large models into their operational frameworks, the inference computing demands stemming from continuous online interactions, real-time analytics, and decision-making support will vastly outweigh the periodic, phased model training requirements.

For financial institutions, this transformation carries significant practical implications. It necessitates a proactive realignment of infrastructure investment priorities and resource allocation strategies. The computing power infrastructure must be capable of supporting high-concurrency, low-latency, and highly reliable real-time inference services while maintaining cost-effectiveness.

Zhang Chi had previously highlighted the computing power engineering challenges faced by the financial AI sector, noting that top-tier institutions grapple with high-concurrency inference demands, while smaller entities are hindered by fragmented computing power resources. With inference computing power taking center stage, the industry will be compelled to prioritize refined computing power management and universal access.

The comprehensive systems offered by vendors like Alibaba Cloud, which provide a 'full-stack AI cloud' encompassing everything from underlying chips to model services, are tailored to address this new computing power demand that is shifting from centralized training to distributed, deep applications.

Confronted with the massive, fragmented, and highly real-time demands arising from the widespread adoption of inference, the paramount challenge is enabling AI to genuinely comprehend complex business intents and execute tasks reliably. Zhang Chi's proposed 'large and small flywheel' synergy system represents Alibaba Cloud's architectural response to this challenge.

The 'large flywheel' is dedicated to macro-level intent comprehension and intricate task planning. Leveraging deep learning of financial domain expertise and enterprise-wide data, it can accurately decipher complex and ambiguous instructions from users or business systems, breaking them down into structured, executable task sequences.

Conversely, the 'small flywheel' specializes in ultra-efficient execution and rapid adaptation within vertical scenarios. It comprises numerous specialized small models or agents embedded in specific business processes, such as credit approval, compliance review, data governance, or customer service.

Operating under the coordination of the 'large flywheel', these 'small flywheels' leverage their domain-specific knowledge to execute tasks efficiently, providing real-time feedback on execution results and insights, thereby driving continuous optimization of the entire system.

A prime example of this approach's success is the data security classification project undertaken by Nanjing Bank in collaboration with Alibaba Cloud. Leveraging the 'model flywheel' paradigm, the project achieved a model accuracy rate of 94.6% with just 6,000 training data points.

At its core, the project utilized a large model to interpret lengthy compliance standard documents and generate knowledge, which then powered lightweight small models to accurately perform data identification and classification tasks, forming an efficient data governance 'small flywheel'.

This success story corroborates Zhang Chi's perspective: financial AI must transition from 'localized experiments' to 'systematic deployment', constructing a self-evolving agent ecosystem rather than relying on isolated tools.

References:

https://finance.sina.cn/bank/yhgd/2026-01-04/detail-inhfcpht8954843.d.html?vt=4

https://www.sohu.com/a/886036491_411876?scm=10001.325_13-109000.0.0.5_32&spm=smpc.channel_248.block3_308_NDdFbm_1_fd.4.1744944938936JJNjYHY_324

https://www.cnfin.com/hb-lb/detail/20251028/4324822_1.html