Finding the Investment Thread in Generative AI: What to Invest in Generative AI in 2024?
-
As the AI narrative evolves, an important question has emerged for investors: What should we invest in generative AI? To answer this question, we might look at the history of mobile internet.
In the eyes of many, the opportunities in generative AI somewhat resemble those of mobile internet. Looking back at the development of mobile internet, there were roughly three key phases: In 2007, the iPhone 1 was released, marking the official launch of mobile internet; in 2010, the iPhone 4 was released, establishing the basic framework of mobile internet; and in 2012, mobile internet applications began to explode, with companies like ByteDance, Didi, and Xiaohongshu being founded.
Logically speaking, the release of GPT-3.5 is more like the iPhone 1 moment in the AI industry—the direction is clear, but the industrial framework remains ambiguous. Such chaos has, to some extent, increased the difficulty of investment, even leading to numerous investment "traps."
This phenomenon also occurred during the mobile internet era. In its early stages, utility apps like "flashlight" tools and app stores attracted significant investor bets. However, history has shown these were merely investment "traps" in the mobile internet landscape.
How to find investment opportunities in the mobile internet? Five Capital's approach may be worth considering.
According to Liu Qin's previous analysis of the mobile internet at Five Capital: Mobile phones possess the characteristics of PCs. However, mobile phones also have three crucial features that PCs lack: 1) Mobile phones have location parameters; 2) Mobile phones contain contact lists; 3) Mobile phones have cameras and external audio devices. Following this logic, Five Capital concluded that the next generation of killer applications will be mobile, social, and rich-media oriented.
Following this logic, when we are looking for investment opportunities in generative AI, it's better to ask ourselves: What capabilities are unique to generative AI that were previously unavailable?
2023 marked the year when large models truly broke through. Reflecting on the breakthroughs in this round of generative AI, they stem from the continuous evolution of underlying large models. GPT, currently the most powerful language model globally, has seen a qualitative leap in performance in just five years—from the first version of GPT in May 2018 to GPT-4 in March 2023.
Currently, there are two main reasons for the rapid evolution of GPT models:
First is the continuous iteration of training methods - from GPT-1's semi-supervised learning, to GPT-2 abandoning the fine-tuning phase, then to GPT-3's In-context learning with massive parameters, and finally ChatGPT which introduced reinforcement learning based on human feedback.
Second, behind the expansion of model parameters lies OpenAI's sustained high investment in R&D and computing power. Through the 'brute force approach,' they have supported the rapid expansion of model parameters and training data.
With the emergence of large models and a series of "killer" applications like ChatGPT, generative AI has demonstrated powerful capabilities in fields such as text, images, code, audio, video, and 3D models.
In March this year, Microsoft released the GPT-4-based AI office assistant, Office Copilot. Since then, AI applications in enterprise services, marketing, low-code development, security, education, healthcare, finance, and other fields have been successively launched. In July, Microsoft 365 Copilot announced its pricing at $30 per user per month, while Salesforce, the global CRM leader, announced the official release of its AI products to all users, with pricing set at $50 per user per month for a single product. With the pricing announcements from these two software giants, AI applications are now officially entering the commercialization phase.
The applications of generative AI are not limited to the B2B sector, as its adoption in the consumer market is also progressing at a remarkable pace. On November 29th, Pika, an AI startup founded just six months ago, officially launched its AI video generation tool Pika1.0. On the same day, the company announced securing $55 million in funding, currently valuing the company at $250 million.
Pika's standout feature lies in its ability to cover everything from ordinary 2D animations to cinematic live-action scenes and 3D animations. It also supports real-time video editing and modifications, with the generated videos rivaling Hollywood animated films in aspects like lighting and motion fluidity.
Various signs indicate that AI applications are entering an era of explosive growth, driven by advancements in models, computing power, and ecosystems.
The rapid growth of generative AI has also ignited investment in related fields.
From the perspective of the primary market, as of the end of August, the number of AI open-source projects on GitHub reached 910,000, marking a 264% increase compared to the previous year. According to Replit's data, the quarter-on-quarter growth rate of AI projects in Q2 2023 reached 80%, with a year-on-year increase of 34 times compared to the same period last year.
From the perspective of investment direction, most generative AI projects are still in their early stages. The majority of funding has been directed towards the AI infrastructure layer, including large model development, while the application layer accounts for only 30% of the total funding.
Among these, the investment concentration in the infrastructure layer is relatively high. Since the third quarter of 2022, the AI infrastructure layer, despite accounting for only 10% of the total number of investments, has captured over 70% of the total funding in generative AI. This reflects the capital-intensive nature of the infrastructure layer.
At the application layer, general AI applications hold absolute dominance, accounting for 65%. In contrast, vertical industry applications currently lag far behind general applications in both the quantity and amount of investments.
From the perspective of the secondary market, AI computing infrastructure companies are the first to benefit from the AI industry wave. NVIDIA is the core beneficiary of the "gold rush shovel-selling" logic in AI, followed by leading cloud service providers and large model manufacturers such as Microsoft, Google, AWS, and Oracle.
The reason lies in the fact that within the current generative AI industry chain, the infrastructure layer is the most certain segment. According to rough estimates by overseas venture capital firm Andressen Horowitz, application manufacturers on average need to allocate 20-40% of their revenue to cloud service providers or large model manufacturers, while large model manufacturers typically spend nearly half of their income on cloud infrastructure. This means that currently, 10-20% of the total revenue from generative AI flows to cloud service providers.
At the hardware level, NVIDIA is the biggest beneficiary, with its flagship AI chips A100 and H100 handling the majority of AI model training and development, accounting for nearly 90% of AI server hardware costs.
Although AI applications are still in their early stages, and the commercialization and realization timeline of the application layer will lag behind the infrastructure layer by several quarters, the stock prices of leading application manufacturers have also seen significant movement this year. From the beginning of this year to date, application-layer companies such as Palantir, Salesforce, and Shopify have all experienced notable increases.
As more AI application manufacturers enter the substantive commercialization stage in the future, investments in the AI field will also enter a more complex phase.
In the context of information technology, software and hardware have exhibited entirely different developmental trajectories. In the software industry, only platform-based applications that directly control user and data assets emerge as the ultimate winners. In contrast, the semiconductor industry demonstrates even higher concentration than downstream consumer electronics companies, with Qualcomm, TSMC, and Intel wielding significant influence in their respective domains.
In other words, whether the evolutionary path of the AI world will more closely resemble the semiconductor or internet industry determines where the largest opportunities will emerge - at the model layer or the product layer? Before answering this question, let's first examine how AI differs from these two established fields.
Unlike many industries, the development logic of the semiconductor industry stems from the consumer market's insatiable demand for performance, which ultimately evolved into Moore's Law. This compels manufacturing giants to maintain technological leadership in each generation of products through massive R&D investments. In this process, performance improvements become increasingly difficult, and the resulting revenue growth often does not match the enormous R&D costs required, creating extremely high barriers in semiconductor manufacturing.
However, such a scenario does not apply to the model layer of generative AI, for two reasons:
First, software is more challenging than hardware to maintain long-term technological leadership. Currently, the primary cost of large models lies in training expenses, but these costs pale in comparison to the billions invested in chip manufacturing. This observation is evident in the development of large models both domestically and internationally. Although there remains a noticeable gap between domestic and foreign large models, this gap has significantly narrowed compared to the beginning of the year.
Second, unlike consumers' insatiable demand for the "performance" of mobile phones and PCs, users' demand for the intelligence level of generative AI exhibits diminishing marginal utility in many scenarios. In other words, not all scenarios require unlimited performance from generative AI.
While the model layer logic may not be viable, it doesn't necessarily mean application layer companies have great opportunities either. The reason is that in the context of large models, product experience is highly tied to model control, and data feedback is crucial for model improvement. In this scenario, it's hard to trust an application layer company that heavily relies on large model providers. For application layer companies, handing over their high-quality data entirely to large model providers for iteration is itself an extremely risky proposition.
From this perspective, both model layer and application layer companies have their own issues. Only full-stack companies that occupy both the model and application layers might have the potential to capture maximum value.