I. AI Investment Logic
Primary Market
It is difficult to determine today whether AI startups can disrupt giants. Many AI startups find it hard to achieve a complete product commercial cycle, often serving as prototypes for established companies, so the early value of AI is mostly captured by large companies.
Secondary Market
AI is driving the growth of tech stocks, such as the Mag 7 companies, all of which have AI layouts, with NVIDIA benefiting the most. The AGIX Index launched by Picking Elephants tracks companies that truly benefit from AI and performs better than other tech indices.
Investment logic includes:
Believe in the Power Law: Over the past 10 years, the top 1% of US stock market companies have contributed to 99% of the returns. The same will happen in the AI field, where the next generation of leading tech companies will emerge.
AI-native is the tech investment paradigm for the next 10 years: AI is the core driver of future tech investment. The performance of related indices after the release of ChatGPT shows that the higher the AI content, the better the growth. However, current AI startups have limitations, and there are more opportunities in mature tech companies in the secondary market, and the AGIX Index can quantify this trend.
AI will bring a reconstruction of business models: Many first-tier AI companies in Silicon Valley are innovating in AI features, but few have a complete business cycle. For example, OpenAI is limited by NVIDIA, Microsoft, etc., and the ToC fruits in the AGI chain may be picked by Apple, while the ToB fruits may be picked by Microsoft. Companies like Adobe and Hikvision have successfully changed their business models in technological transformations, and AI will bring similar stories.
II. Key Judgments in AI Investment
Uncertainty of AGI
There is currently no consensus on the definition and understanding of AGI, which has both scientific discovery attributes. In the future, the competition will be about the efficiency and capabilities of intelligent output. At the same time, it is unknown how big AGI will be. The pricing of software products may change from being based on the number of people to being based on the results produced. AI may double the global GDP, but currently, the GDP directly related to AGI is small.
The role of AI in global growth
AI is a key force in global growth over the next 15 years and may contribute at least $10 trillion in GDP growth. It can replace at least 30% of white-collar jobs. Large models empower or disrupt the knowledge worker group, and its progress may usher in a new era of culture and creativity.
Stage of LLM
LLM is in the early stage of infrastructure construction, and the lack of computing power limits the outbreak of AGI applications. In 2023, the global GPU consumption time is limited, but it may increase significantly in the future. Currently, many large model companies in Silicon Valley have a large number of GPU clusters, and some companies may build supercomputers by 2027-2028. The competition for LLM is a competition for resources, and the entry threshold is locked. Players capable of deploying ultra-large clusters converge to 4-5 companies.
Nature of AGI Infrastructure
AGI infrastructure is an engineering problem that can be solved through investment and time. For example, the training cost of GPT-4 is high, and there are technical and energy issues with GPU clusters, but these can be improved over time and with investment.
Unlocking AGI capabilities
AGI capabilities are unlocked gradually. Search is the early killer app of LLM, and in the next 2-3 years, we may see AGI in the coding field. The software logic is clear and has feedback from 0 to 1. In the future, long-tail needs may be solved by AI combined agents.
III. Capturing AI Alpha in the Secondary Market
Current Investment Status and Trends
We are currently in the early stage of the AI revolution, which is a good time for investment. Although the hardware sector has recently adjusted, AI is still developing from a fundamental perspective. Enterprises have bottlenecks in adopting LLM, mainly considering issues such as accuracy, data security, and compliance with application scenarios. However, in reality, some enterprises have already deployed LLM internally. The model capabilities will further develop in the second half of the year, and more applications are expected to appear next year.
Main Investment Threads
Computing Power: The world needs more computing power, and cheap computing power is scarce. Model iteration is constrained by computing power. In the medium and long term, factors such as the early scaling law, the development of multimodal models, and the assumption of defense expenditure will drive the demand for computing power. NVIDIA is the preferred investment for computing power. Although the stock price has risen, the valuation is not saturated, and there are reasons such as the expansion of market size, three-in-one competitive advantages, and a complete product matrix.
Cloud Infrastructure: The profit pull of AI on cloud vendors has begun to emerge, such as increasing marginal contributions to Microsoft. The willingness of leading enterprises to invest in the cloud is increasing. Amazon is the preferred investment for cloud vendors, AWS is an important distribution channel for LLM, and its retail and advertising businesses are entering a growth period.
End Side: End-side applications are beginning to progress. On the ToC side, this year is the first year of AI phones, and executives value end-side AI on the enterprise side. Apple is the preferred layout for the end side because its comprehensive software and hardware layout can control the traffic entrance, and the growth potential of iCloud is underestimated.
Interconnection and Software: The super cycle of storage and interconnection is still unfolding. The next generation of large model training requires an increase in the scale of clusters and data. Broadcom is the preferred choice in the interconnection and storage link because it has strategic significance in interconnection and is the main producer of customized chips.
Software Model Transformation: The SaaS model may change from being based on seats to being based on the amount of computing Token-as-a-Service model. ServiceNow is the preferred choice in the software aspect, which can help enterprises deploy AI and provide good services in aspects such as enterprise internal search. The rapid growth of Now Assist shows that the demand for enterprises to deploy LLM is strong.