In this digital era, the digital economy has become a crucial engine for global economic growth. With the rapid development of technologies such as the Internet, big data, and artificial intelligence (AI), the digital economy industry is undergoing an unprecedented upgrade. AIGC—referring to Artificial Intelligence, the Internet of Things (IoT), Big Data, and Cloud Computing—are the four key technologies driving the upgrade of the digital economy industry. The integration and innovation of these four technologies have brought tremendous transformations and development space to the digital economy industry. Nanjing Innovation Investment Group has invited Ms. Yang Jinyie, Founding Partner of Shenzhen Xinpao Investment Consulting Co., Ltd., to discuss and share the latest dynamics and development trends of the digital economy industry around the theme of "Digital Economy Industry Upgrade Driven by AIGC".

Could you talk about the application scenarios and investment opportunities of pre-trained large models in the fintech field (such as insurance, credit, commercial banking, etc.)?

Pre-trained large models, on one hand, provide tools for financial language processing and understanding; on the other hand, they enable more accurate Q&A for financial-related information, and better realize information and decision recommendation. Traditional IT vendors in the financial field, with their accumulated rich vertical knowledge and existing AI product layouts, can build domain-specific models in the financial sector based on high-quality financial data and large language models as the foundation. Correspondingly, fields such as intelligent customer service, intelligent investment consulting, and anti-money laundering will witness significant upgrades and iterations of their functions.

Regarding the new business models brought by the development of AIGC, in which directions lie the relatively certain opportunities?

Overall, the current observation is that AIGC has brought substantial transformations to the content field—such as text generation, code editing, graphic and image generation and editing, as well as the fields of voice, video, and 3D. These transformations can free up a large number of human resources engaged in basic content production, and many existing content production tools and plugins for streaming editing will also undergo changes. Additionally, there are new opportunities to reduce corporate production costs and improve efficiency in vertical application fields. AI has shortened the time from the proposal of theories to their verification, and also reduced the time from the raising of demands to their realization. The trends we previously mentioned, such as digitalization, networking, and intellectualization, will be further deepened in this wave.

Take code generation as an example: Generative AI translates natural language into code, which greatly enhances the intelligence and automation of computer programming. This allows programmers to write more code within the same time while improving the efficiency of debugging. Currently, companies such as OpenAI, Microsoft, Google, Amazon, and Huawei have all laid out plans in the field of AI code generation. For instance, GitHub Copilot—launched in June 2021 by GitHub (a subsidiary of Microsoft) and powered by OpenAI Codex—can be integrated into editors such as Neovim, JetBrains IDEs, Visual Studio, and Visual Studio Code, supporting programming languages including Python, JavaScript, TypeScript, Java, Ruby, and Go. It can automatically write code based on context, including docstrings, comments, function names, and code; as long as the user provides a prompt, it can generate a complete function. GitHub Copilot can save developers’ costs by up to tens of billions of US dollars annually. Between June 2021 and June 2022, over 1.2 million developers registered to use the preview version of GitHub Copilot, and nearly 40% of the code in enabled files was written by Copilot. Based on 1.2 million users and a developer hourly wage of 20-200 US dollars, Copilot can save development costs of 4.8-192 billion US dollars per year.

From the perspective of the development trends of pre-trained large models and AIGC, in terms of computing power, algorithms, data, and application fields, what opportunities do domestic startups have to stand out?

First, there is the opportunity in the understanding and processing of multi-modal data. Second, the "pre-training + fine-tuning" large model effectively addresses the insufficient generalization ability of traditional AI; the combination of large models and small models can effectively reduce the marginal cost of AI implementation. The new-generation AI technology is expected to initiate a brand-new cycle of technological innovation.

At the current stage, building and developing super large-scale generalized models imposes extremely high requirements on the total investment cost, organizational capabilities, engineering barriers, and underlying frameworks. As models expand, their training and maintenance costs also increase continuously. In terms of training costs, for example, the cost of training GPT-3 once is as high as 84 million RMB. In terms of computing power, the computing power consumed for ChatGPT’s training is approximately 3,640 PF-days—meaning that if the computing speed is 100 trillion operations per second, it would take 3,640 consecutive days to complete the training. Additionally, there are costs related to data acquisition and labeling, data barriers, and social cognitive acceptance.

Therefore, different from the technical route of deep learning, the high barriers of large models and their "high initial investment, low marginal cost" model determine that the large model industry will inevitably move towards a concentrated pattern. However, there will be numerous opportunities involving industry-specific Know-How in key vertical niche scenarios, which will support many ISVs (Independent Software Vendors) to build applications based on large models or further optimize them. This also provides small models and refined models with comprehensive cost advantages and development opportunities in long-tail fields. It is worth noting that OpenAI has already launched investments in dozens of application software companies in various niche fields.

As pre-trained large models find more and more applications, what new investment opportunities will they bring to the traditional computing power chip investment track?

In terms of computing power, pre-trained large models have enormous demands for power consumption or energy consumption. Although the energy consumption of current chips is decreasing, training a large model requires accurate and continuous computing by high-performance digital computers, which incurs extremely high energy consumption. Studies have shown that the human brain only requires 30 watts of energy, while large AI systems need megawatt-level energy consumption to operate—a gap of 1,000 times. Therefore, from the perspective of energy utilization, a potential future development path for AI models is as follows: We train models on high-computing-power digital computers; after the models are fully trained, we can apply the trained models on low-power edge devices. These developments are already taking place. In the future, when we engage in intelligent interactions such as conversations with AI running on small household appliances, the chips equipped in these appliances may be low-cost, low-computing-power, and low-power edge computing chips that cost only a few US dollars. At the same time, when integrated with sensor systems, their conversation and interaction processing capabilities may still reach the level of large models. Therefore, both high-computing-power chips and edge-side chips present opportunities, but the time windows for startups to develop and the resource endowments they need are significantly different.

Computing power chips have high requirements for foundry processes. Under the current international environment of the semiconductor industry, what impact does foundry security have on domestic semiconductor track investments?

The most advanced logic chips (such as NVIDIA’s GPUs, Intel and AMD’s CPUs) have entered the 5nm "leading node" and will soon achieve large-scale mass production at the 3nm node. However, for discrete devices, optoelectronics, and sensor fields, even the 180nm traditional process (accounting for 19% of production) can still meet the needs. The manufacturing processes for memory chips are mainly in the 10-22nm and 28-45nm ranges. From this perspective, we have many foundry processes that can cover these needs. Of course, from a long-term perspective in the future, countries that can master advanced manufacturing processes below 10nm will gain core competitiveness, and we still need to catch up. In the short term, heterogeneous technologies such as Chiplet can achieve a balance among power consumption, performance, area, and cost, which is a potential solution.

Xinpao Capital is a private equity investment fund management institution focusing on the technology field. Its founding team has been involved in the venture capital field since 2014. Its main investment directions include advanced technology industries such as electronic technology, semiconductors, and new-generation information technology. Since 2016, Xinpao Capital has carried out early-stage investments in the AI field, covering natural language processing and computer vision; some of its portfolio companies rank among the top in the domestic industry. By 2017, Xinpao Capital launched early- and mid-stage investments in the semiconductor industry chain and the electronic information industry, gradually bringing more enterprises into the Xinpao resource pool. In addition, its portfolio companies can obtain in-depth resources and guidance in the semiconductor and electronic information industries, enhancing the ability to empower the industry. In the past five years, Xinpao Capital has further improved its industrial chain layout, successfully nurtured industry unicorns, and formed a network covering core industrial regions such as Shenzhen, Fujian, and the Yangtze River Delta. From 2022 to 2023, the first batch of portfolio companies of the fund went public, marking the initial realization of Xinpao Capital’s investment blueprint. Since 2021, Xinpao Capital has gradually expanded its industry reputation and influence, winning many awards: it was listed in Financing China’s 2021 Private Equity Investment List as "China’s Top 10 Most Potential Investment Institutions", FOFWEEKLY’s 2021 "Top 20 Investment Institutions in the Advanced Manufacturing Field Most Concerned by LPs", FOFWEEKLY’s 2022 "GP Top 100 in the Investment Institution Soft Power Ranking", Financing China’s 2022 "China’s Most Growing Investment Institutions", and the 2023 China IC Influential Ranking by the Semiconductor Investment Alliance (iJiwei) as "Annual Best Emerging Investment Institution".