At the 28th Beijing International High-Tech Expo - Future Industries Promotion Conference held from May 8 to 9, 2026, in Beijing, Xin Zhiyun, Vice President and Chief Information Officer of Gf Securities Co.,Ltd., delivered a keynote speech.
The following is a summary of the speech:
Xin Zhiyun: Friends and distinguished guests, my presentation today is titled "The Application Practice and Reflections on Large Language Models in the Securities Industry." Due to time constraints, I will focus on two main areas: first, the development trends and competitive landscape of large model technology, and second, the practical applications of AI within the securities industry.
First, let's examine the development of large model technology. Currently, large models are in a phase of rapid iteration and acceleration, with competition among leading companies becoming increasingly intense. The industry's iteration cycle has accelerated from a yearly pace to a monthly one, with new models being released frequently and capability curves continuing to rise steeply. Simultaneously, the market share of large models is diversifying, with non-ChatGPT models gaining ground, and models like Google's Gemini experiencing rapid growth, reflecting the dynamic shifts in the competitive landscape.
Leading global AI companies are leveraging their unique resources to build differentiated competitive advantages, accelerating their strategic layouts across key areas such as models, computing power, and ecosystems, thereby forging distinct competitive paths. Google has established a trinity of advantages through its search data flywheel, proprietary TPU chips, and multimodal data. Meta relies on its massive social media user base, employing a dual-track strategy of open-source and closed-source development. OpenAI possesses the world's largest user base, giving it significant advantages in general-purpose applications. Anthropic focuses on high security and trustworthiness, with deep penetration in heavily regulated industries. Tesla holds unique leadership in connecting the physical world, humanoid robotics, and proprietary chips, with its Optimus robot currently serving as the only scalable gateway to the physical world.
According to insights from Meta's senior experts, the AI value chain can be broken down into six key segments: on the left are computing power/chips and foundational models, followed by tool usage and distribution channels; on the right are enterprise workflows and physical world execution. The further to the right, the closer one gets to value realization and the physical world. We can observe that large model technology is gradually moving beyond apps and into the physical world. For instance, Google is strong in computing power, chips, and foundational models; OpenAI excels in foundational models and agents; Anthropic has advantages in deeply integrating with enterprise workflows; and Tesla is a leader in connecting the physical world. Two key characteristics are evident: First, the focus of AI value is shifting from the model layer to the model ecosystem and task execution layer, with the industry placing greater emphasis on practical implementation and the deep integration of large models into business workflows. Second, leading U.S. companies have established comprehensive layouts across all six segments. AI competition has evolved from a simple contest of models and computing power to a comprehensive battle of strength across the entire value chain and process.
Leading domestic companies, including traditional internet giants and AI newcomers, are also accelerating their strategic deployments in models, computing power, and ecosystems, leveraging their respective strengths in traffic, cloud services, and technological foundations. ByteDance leverages its Doubao large model and Douyin traffic ecosystem, giving it a prominent advantage in consumer-facing (C-end) channels while also driving business-to-business (B-end) adoption. Tencent utilizes its comprehensive WeChat ecosystem and social traffic channels to build its presence across both B-end and C-end markets. DeepSeek has shown remarkable performance in foundational models and hardware-software optimization. Kimi has developed distinctive capabilities in long-text processing and agent tools.
Analyzing the domestic model value chain across these six segments, ByteDance demonstrates strong overall capabilities, with a slight relative weakness only in the physical world segment. Tencent, leveraging its C-end traffic and social media channels, holds advantages in agent tool applications and default distribution channels. DeepSeek's foundational model ranks first in usage and API calls among open-source models. Moonshot AI also exhibits strong capabilities in agent tool usage. Overall, China's large model sector is developing rapidly, driving the intelligent transformation across numerous industries.
Since 2026, the industry has reached a consensus on four major trends in large model development: First, trillion parameters have become the entry-level threshold for foundational models, serving as a crucial basis for agent applications. Second, the industry has transitioned from the pre-training Bot era to the post-training dominated Agent era, with OpenClaw being a landmark event driving this shift. Third, a new 3:1:1 pattern has emerged in computing power allocation. Previously, computing power was primarily focused on pre-training. Now, due to surging demands for model development, experimentation, and inference, GPU distribution is approximately three parts for research, one for pre-training, and one for post-training. Consequently, GPUs have become the primary bottleneck constraining research progress, which explains why many Wall Street firms are investing heavily in building their own computing platforms—a key application is to allocate more resources to research, experimentation, and post-training. Fourth, the AgentOS framework has become mainstream, with models shifting from passive Q&A to active interaction, and continuous innovation in multi-agent collaboration and memory management technologies.
Regarding application penetration, the global adoption rate of AI by enterprises is steadily increasing. Data from the U.S. Census Bureau in March 2026 shows an overall AI adoption rate of 18.9% among U.S. businesses. By industry, information services, professional services, educational services, and finance/insurance are leading. By company size, large enterprises with over 250 employees have the highest adoption rate at 35.3%, while small and medium-sized enterprises (SMEs) with 20-49 employees have shown the greatest increase, indicating that AI adoption is spreading from leading firms to a broader market.
AI penetration in China is occurring even more rapidly. Data from CNNIC shows that as of the end of 2025, China's generative AI user base reached 612 million, with a penetration rate of 47.8%, representing a year-on-year growth of 141.7%. IDC assessments indicate that the top five industries for AI penetration are internet, finance, telecommunications operators, manufacturing, and government, with the finance industry reaching 78%, ranking second. McKinsey research shows that 88% of enterprises have normalized AI use in at least one functional area, with the generative AI usage rate among Chinese enterprises reaching 83%, the highest globally. Despite the rapid acceleration in technological penetration, the depth of application remains insufficient—this is the current state of AI penetration in China.
In the second part, I will introduce the practical applications of AI in the securities industry.
Research from Stanford University categorizes work scenarios into four quadrants based on two dimensions: "human willingness to apply" and "AI capability": the green zone, the R&D opportunity zone, the red zone, and the low-priority zone. The green zone represents scenarios where there is both human willingness and sufficient AI capability. The R&D opportunity zone represents scenarios where there is human willingness to apply AI, but AI capability is still relatively lacking, yet holds application potential. Research indicates that approximately 41% of AI startups are concentrated in the low-priority and red zones, reflecting a significant mismatch between current AI agent research, investment, and practical needs. This suggests that AI deployment should prioritize scenarios where employees are willing to use it and where the technology can be effectively implemented, avoiding inefficient investments.
Considering the characteristics of the securities industry, we have proposed the CODE model as a comprehensive framework for identifying AI application value: C (Coding) targets headquarters IT, risk control, and compliance personnel, focusing on assisted programming, intelligent risk control, and compliance efficiency. O (Operation) targets frontline business and investment advisor personnel, aiming to enhance operational efficiency and service productivity. D (Decision) targets management personnel, providing data insights, risk warnings, and decision support. E (Experience) targets external clients, optimizing service experience, innovating products, and increasing revenue. This model provides a clear pathway for identifying AI value and planning its implementation.
Overseas investment banks have comprehensively advanced their AI strategies. In April of this year, I conducted research on Goldman Sachs. Goldman Sachs launched its One GS 3.0 strategy, fully embracing AI with primary goals of enhancing client experience, improving profitability, boosting productivity and efficiency, and increasing resilience. Where is the focus? Prioritizing process implementation, it is reengineering and reshaping processes in areas directly impacting client experience, such as sales enablement, client onboarding, loan processing, regulatory reporting, supplier management, and intelligent automation, to improve efficiency. It has also developed numerous typical applications in sales and trading AI, employee assistance, and development support. For example, sales and trading AI optimizes sales processes, provides deep investment insights, and supports loan approvals, increasing loan application assessment accuracy by 23% and boosting intraday trading profitability by 27%. AI engineers have enhanced productivity for over 17,000 developers at Goldman Sachs.
Morgan Stanley, as the first major Wall Street firm to deeply embrace OpenAI's GPT-4 technology, systematically advances its AI strategy through an exclusive strategic partnership with OpenAI, establishing a dedicated group-wide AI department and steering committee. It has implemented benchmark applications in wealth management, investment research, and code modernization, significantly improving efficiency and reducing costs.
Charles Schwab empowers client services, investment advisory operations, wealth planning, and risk management with AI, consolidating its leading position in retail wealth management.
Within China's asset management industry, represented by E Fund, there has been long-term dedication to financial technology. E Fund independently developed its EFund large model, covering the entire process of investment research, operations, marketing, and office work. Applications like the intelligent fixed-income trading robot and AI macro researcher have shown outstanding results, making it a benchmark for AI implementation in the fund industry.
Domestic securities companies primarily follow two paths for client-facing AI applications: one is creating standalone AI-native apps to enhance immersive client experiences; the other involves iterative evolution on existing apps. At Gf Securities Co.,Ltd., we adopted a strategy of "small, quick steps and rapid iteration," launching the "AI Lens" version of our Etaojin app. It focuses on core investment scenarios like stock selection, market monitoring, trading, and ETF investment assistance. It is currently in internal trial operation and will be fully launched for client services after completing regulatory filing procedures.
Simultaneously, we have built the "Tianji Zhirong" one-stop financial AI portal, integrating nearly 50 professional agents to achieve cross-business, one-stop intelligent responses. It serves over 300 of our business branches, thousands of investment advisors, and headquarters and branch employees. It has shown significant results in fund analysis, client matching, marketing scripts, and business conversion, effectively enhancing service experience and business conversion efficiency.
Based on OpenClaw technology, we have developed an AI employee assistant, with a particular focus on upgrading security controls and undergoing third-party security assessment certification. The AI assistant supports tasks in general office and professional scenarios such as schedule management, meeting bookings, expense reimbursement, and business inquiries, enabling an upgrade in work modes. As an example, our research analysts using the OpenClaw employee assistant can automatically read financial reports and generate research reports and roadshow materials, significantly freeing up analyst bandwidth. A single analyst can gain capabilities equivalent to 3-5 research assistants.
In the R&D domain, we are utilizing AI not only for assisted programming but also striving to empower the entire process—from requirement design and preliminary design to testing and operations—with large models. We are also exploring the potential impacts of AI-assisted programming on development, organization, and architecture, conducting research to improve overall IT R&D efficiency.
Finally, AI security governance is of paramount importance for financial institutions. As AI applications rapidly proliferate, security risks have garnered global attention. China has released the "Artificial Intelligence Security Governance Framework," and the U.S. Treasury Department has specifically convened financial institutions to discuss AI security. Therefore, this is an area that requires our utmost attention.
That concludes my report. Thank you all.
Comments