China’s economic transition and the development of artificial intelligence (AI) infrastructure are increasingly being discussed together rather than as separate themes. At the Shanghai International AI Finance Summit 2026, Zhang Jun, dean of the School of Economics at Fudan University, addressed the structural forces shaping China’s economic outlook, while Lin Yonghua, vice president and chief engineer at the Beijing Academy of Artificial Intelligence (BAAI), examined the technological systems required to make AI usable at scale across industries. Zhang set the macroeconomic context. Nominal GDP grew only 3.9% in 2025, well below the headline real figure of 5%, because constant-price accounting masks the impact of deflation, which ran at minus 1.1%. Fixed asset investment fell sharply: real estate contracted 17.2%, infrastructure fell 2.2%, and even eastern provinces saw investment decline more than 8%. Exports grew 6.1% but imports rose only 0.5%, reflecting weak domestic demand and a record trade surplus of $1.1 to $1.2 trillion. Zhang’s analysis placed China’s current economic phase within a longer historical trajectory. For several decades, China’s rapid expansion was supported by large-scale capital investment, infrastructure construction and industrial capacity growth. These drivers enabled the country to build extensive manufacturing capability and integrate deeply into global trade networks. However, as Zhang noted, the economic environment has gradually shifted. Urbanisation has matured, industrial capacity in many sectors has reached high levels and the scope for growth based purely on scale expansion has narrowed. These changes do not necessarily imply economic stagnation, but they do require a different set of drivers to sustain long-term development. Hu Yong, deputy district mayor of Jing'an District, noted that the summit coincided with a significant policy moment: building an intelligent economy had been written into the government work report for the first time at the recently concluded National People's Congress session. For financial institutions, he argued, embracing AI as the core driver of transformation in business models and management systems had become an unavoidable industry imperative. AI is evolving from models to industrial systems Lin from BAAI examined the technological systems required to support that transition. Her keynote, titled "From ChatGPT to OpenClaw: Making AI Compute Affordable and Accessible at Scale," focused on the infrastructure necessary for deploying AI technologies across industries. Lin urged caution about what "AI" actually means. The term encompasses statistics, machine learning, optimisation and data science — disciplines that have evolved over decades. Conflating them, she warned, produces conceptual confusion about what the current wave can actually deliver. She traced the recent acceleration through four inflection points: ChatGPT in 2023, multimodal models in 2024, DeepSeek's open-source breakthrough in early 2025, and the emergence of agentic systems by year-end. Each shift changed not just capability but the logic of how AI interacts with human work. In the agentic model, AI moves from supporting role to protagonist — executing sequences of tasks autonomously rather than responding to individual prompts. That shift amplifies both the potential and the risk. Lin's work at BAAI centres on foundational technologies supporting large models, including computing infrastructure, system architecture and open-source ecosystems — a perspective that shifts attention away from individual models towards the broader stack required for deployment at scale. AI systems require reliable compute infrastructure, efficient deployment frameworks and cost structures that allow organisations to integrate them into operational environments — conditions that are particularly demanding for financial services, where technologies must be embedded into regulated environments, integrated with existing data infrastructure and deployed across large organisations with consistent governance. The missing ingredient in enterprise AI deployment, Lin argued, is not raw model capability but certified, domain-specific skills — structured knowledge that agents can reliably call upon to solve professional problems. Hundreds of thousands of open-source skills now exist globally, but what the industry lacks are verified skills capable of solving specialist problems efficiently. Banks and financial institutions, she suggested, are well-placed to develop and own that layer. Technological ecosystems determine the scalability of AI Lin's perspective also emphasised technological ecosystems. AI becomes economically meaningful only when infrastructure lowers the cost of entry enough for large numbers of organisations to deploy it. Open-source collaboration, shared research platforms and interoperable standards are central to that process. Underpinning all of this is a compute infrastructure problem the industry has not yet resolved. By 2026, with agentic systems making tens or hundreds of model calls per task and running continuously, more than 70% of AI compute demand has shifted to inference — and supply has not kept pace. Incompatible hardware architectures compound the problem across both global and domestic markets. Lin described this as "one chip, one stack": a model deployed on one architecture cannot readily migrate to another, raising costs and slowing enterprise adoption. BAAI has led a three-year collaboration with Peking University, Tsinghua University, the Institute of Computing Technology, the Chinese Academy of Sciences and more than ten chip manufacturers to build FlagOS, an open-source software stack that now supports more than twenty chip types. The goal is universal compatibility: any model, any chip, one codebase. Open-source collaboration of this kind, she concluded, remains the fastest path to making compute genuinely accessible — a prerequisite for enterprise-scale AI adoption. Data integrity is an equally consequential challenge. As agentic systems take on greater autonomy — executing workflows, triggering transactions, operating continuously without human review — the accuracy of the data they act upon becomes critical. Lin cited cases from China's supply chain finance sector involving fictitious goods, false trade documentation and multiple pledging of the same assets. These were not AI failures — they were data failures. But as AI systems are trusted to act on such data without human oversight, she warned, the consequences of inaccuracy scale exponentially. For sectors such as banking, this ecosystem development determines whether AI remains an experimental tool or becomes a genuinely industrial capability. Emmanuel Daniel, founder of The Asian Banker, argued that AI raises deeper questions about the structure of financial intermediation itself. Digital banking largely represented the digitisation of existing processes and customer interfaces. AI, by contrast, introduces the possibility of transforming how financial services are organised and delivered at a more fundamental level. Operational gains versus enterprise impact Christian Kapfer, director of research at TABInsights, brought a sharper empirical lens to the question of where AI deployment actually stands. Drawing on the Global Excellence Retail Finance Programme — which analysed more than 260 bank submissions worldwide — he noted that AI deployment at scale has not grown meaningfully since 2023, with most banks still stuck at around 20% of use cases reaching production level. Banks, he said, have reorganised their leadership structures multiple times in search of the right governance model for AI, cycling through changes at CEO, CIO and data analytics levels. He added that despite the investment, institutions have been candid in acknowledging they do not know how to measure AI's impact at enterprise level. The industry benchmark most often cited — DBS's $1 billion in AI-generated economic value — derives from machine learning and A/B testing, not from generative AI, which remains far harder to quantify. Li Lin, vice-president of the SPDB Research Institute, reinforced the point from a practitioner angle. Deploying more AI tools or setting more AI-related KPIs does not, in itself, correlate with better outcomes. He was clear that the measure of AI adoption, or the absence of it, must ultimately show up in performance: in scale and in efficiency. Chak Wong, managing director of the machine learning centre of excellence at J.P. Morgan offered an equally measured view. He characterised AI as a continuation of a long technological evolution rather than a rupture, warned that expectations risk overshooting what the technology can currently deliver, and questioned whether the tendency of large language models to hallucinate can be overcome at all. The human dimension Despite the focus on technology, participants were emphatic that banking remains a service industry centred on trust — nowhere more visibly than in wealth management. Mary Lo, general manager of personal banking and wealth management at Bank of China (Hong Kong), was direct about what that means in practice, particularly in wealth management. "You still need a human touch," Lo said. "You don't want to talk to an AI or robot for a million-dollar investment." The bank has deployed an internal frontline AI capability that gives relationship managers faster access to product information, policies and procedural guidance — reducing preparation time and sharpening the quality of each client interaction. The relationship manager remains central; the technology restructures how that manager prepares and responds. David Lim, head of credit card and personal loan product development at Alliance Bank Malaysia, framed the human-AI boundary clearly. "Whatever AI has done, or whatever decision AI has made for you, it still keeps you informed and you make the final call," he said. Varun Sabhlok, a member of the International Council of Advisors to The Asian Banker's Excellence in Retail Finance programme, raised a broader question for the room: what does it mean to remain human in an AI-driven world, and not only in financial services? Retail banking diversifies; digital banks face scale constraints It is a question the retail banking landscape raises in its own way, where the gap between digital investment and measurable competitive advantage remains the defining issue. Lower interest rates since 2024 have weakened reliance on balance-sheet spreads, prompting institutions to strengthen fee-based revenue and deepen engagement across payments, deposits and wealth services. Bank of China (Hong Kong) emerged as the top-performing institution in TABInsights’ global retail bank rankings, rising from sixth to first. Lo attributed the result to revenue diversification, close integration with parent Bank of China, sustained development in wealth management and increasing focus on AI-driven process efficiency. Malaysia illustrated the broader regional picture. Lawrence Loh, co-chief executive of group commercial and transaction banking at CIMB Bank, noted that Malaysia's digital banks had followed a trajectory similar to Singapore's — high deposit rates, unsecured lending — and that incumbent banks with strong digital foundations had not been greatly disrupted. John Lau, head of business optimisation at Alliance Bank, was direct: "This is not a technology war — it is a price war." Digital banks are attracting customers primarily through higher deposit rates, and their cumulative customer base represents a meaningful threat to incumbent margins. Chammika Weerasinghe, head of digital business at Hatton National Bank, offered a Sri Lankan perspective. Where the average consumer holds relationships with three banks, becoming the preferred bank is the strategic imperative. The transition from digital to AI banking, he argued, hinges above all on data governance — specifically on the evolving regulatory framework for how data can be shared and used across the financial system. The competitive divide ahead David Gyori, advisor to the Asian Banker Excellence in Retail Finance programme, suggested that the future divide in banking may not lie primarily between traditional banks and digital institutions. He added that the real question is whether banks can convert existing digital capability into stronger profitability. The emerging divide, he argued, is not a digital divide but an AI divide. For now, the AI bank remains a concept in development. But the debate at Shanghai signalled clearly that the industry has entered a new phase of transformation. The question is no longer whether that transformation will happen, but which institutions have built the foundations to lead it.
