What should I vote for the generative AI in 2024？

Image source @ vision china

Text | Read Finance and Economics

With the evolution of the AI story, an important question is gradually placed in front of investors: What should the generative AI invest in? To answer this question, we might as well look at the development history of the mobile Internet.

In the eyes of many people, the opportunity of generative AI is a bit like the mobile Internet. Looking back at the development of the mobile Internet, there are probably three key stages: in 2007, the iPhone 1 was released, and the mobile Internet officially set sail; In 2010, iPhone 4 was released, which laid the basic framework of mobile Internet. In 2012, the application of mobile Internet began to explode, and Bytes, Didi and Little Red Books were established one after another.

Logically speaking,The release of GPT 3.5 is more like the iPhone 1 moment in the AI ​ ​ world. The development direction has been clear, but the industrial framework is still unclear.Such chaos has also increased the difficulty of investment to some extent, and even many investment "traps" will appear.

This kind of thing also happens in the era of mobile Internet. In the early days of the mobile Internet, tools such as "flashlights" and application stores once attracted many investors’ bets. Later, it turns out that this is just an investment "trap" of the mobile Internet.

How to find investment opportunities for mobile Internet? The method of Wuyuan Capital may be worth your reference.

According to Wuyuan Capital Liu Qin’s previous analysis of the mobile Internet:The mobile phone has the characteristics of a PC. However, there are three very important things about mobile phones that are not available on PC:1) There are location parameters on the mobile phone; 2) There is an address book in the mobile phone; 3) The mobile phone has a camera and an external device. Following this logic, Wuyuan Capital draws a conclusion that the next generation of killer applications are mobile, social and rich in media.

Along this logic, when we are looking for investment opportunities for generative AI, we might as well ask ourselves a few more questions:What ability is unique to generative AI that was not available before?

2023 is the year when the big model really broke. Looking back, the breakthrough of this round of generative AI comes from the continuous evolution of the big model at the bottom. As the most powerful language model in the world, GPT has made a qualitative leap in the performance of the model in just five years, from the initial version of GPT in May 2018 to GPT-4 in March 2023.

At present, there are two main reasons for the rapid evolution of GPT model:

One is the continuous iteration of training methods, from semi-supervised learning of GPT-1 to abandoning the fine-tuning stage of GPT-2, to In-context learning and massive parameters of GPT-3, and ChatGPT after introducing reinforcement learning based on artificial feedback.

Second, behind the expansion of model parameters, OpenAI’s continuous high investment in research and development and computing power has supported the rapid expansion of model parameters and training data through the way of "making great efforts to make miracles".

With the birth of big models and a series of "killer" applications such as ChatGPT, generative AI has shown great capabilities in the fields of text, image, code, audio, video and 3D models.

In March of this year, Microsoft released Office Copy, an AI office assistant based on GPT-4. Since then, AI applications including enterprise services, marketing, low code, security, education, medical care, finance and other fields have been released one after another. In July, Microsoft 365 Copilot announced the pricing of $30/month for each user. At the same time, Salesforce, the global CRM leader, announced the official opening of AI products to all users, and gave the pricing of $50 per user per month for a single product. With the release of AI function pricing of the two software giants, AI applications will officially enter the stage of commercialization.

The application of generative AI is not limited to B-end, and the landing speed of C-end is also considerable. On November 29th, Pika, an AI startup founded only half a year ago, officially launched Pika1.0, an AI video generation tool. On the same day, the company announced that it had obtained financing of 55 million US dollars, with a current valuation of 250 million US dollars.

The characteristic of Pika is that it can fully cover movie scenes and 3D animations from ordinary 2D animation to real-life photography, and it can also support real-time editing and modification of videos, in which the generated videos don’t even lose Hollywood animation movie level in terms of light and shadow, motion fluency and so on.

There are indications that AI application is entering the era of big explosion, driven by model, computing power and ecology.

The rapid growth of generative AI has also detonated investment in related fields.

From the primary market, as of the end of August, the number of AI open source projects on GitHub reached 910,000, an increase of 264% compared with last year. According to Replit’s data, the growth rate of AI projects in the second quarter of 2013 reached 80%, which was 34 times higher than that in the same period last year.

From the perspective of whereabouts, most of the generative AI projects are still in the early stage, and most of the funds are invested in the AI infrastructure layer including large model development, while the application layer funds flow only accounts for 30%.

Among them, the investment concentration of infrastructure layer is relatively high. Since the third quarter of 22, the amount of investment and financing in the AI infrastructure layer has accounted for more than 70% of the total amount of generated AI financing, which also reflects the capital-intensive characteristics of the infrastructure layer.

At the application layer, general AI applications account for the majority, accounting for 65%. In contrast, vertical industry applications are far lower than general applications in terms of both the amount and amount of investment and financing.

From the perspective of the secondary market, AI computing infrastructure companies are the first to benefit from the wave of AI industry, among which NVIDIA is the core beneficiary of AI’s "Nuggets Buy Shovel" logic, followed by Microsoft, Google, AWS, Oracle and other head cloud service manufacturers and large model manufacturers.

The reason is that in the current generative AI industrial chain, the infrastructure layer is the most certain link. According to the rough estimation of Andressen Horowitz, an overseas venture capital institution, application vendors need to pay an average of 20-40% of their revenue to cloud service providers or large model vendors, and large model vendors usually pay nearly half of their revenue to cloud infrastructure. In other words, 10-20% of the total revenue of current generative AI flows to cloud service providers.

On the hardware level, NVIDIA is the most beneficial target, and its main AI chips A100 and H100 carry the vast majority of AI model training and development, accounting for nearly 90% of the hardware cost of the AI server.

Although AI application is still in its early stage, and the application layer will lag behind the infrastructure layer for several quarters from commercialization and redemption time, the share price of head application manufacturers has also been interpreted this year. Since the beginning of this year, application-layer companies such as Palantir, Salesforce and Shopify have also seen good growth.

As more and more AI application vendors enter the substantive commercialization stage, the investment in the AI field will also enter a more complicated stage.

In the context of information technology, software and hardware show completely different development trajectories. In the software industry, only platform-based applications that directly control users and data assets are the ultimate winners. In the semiconductor industry, it has a higher concentration than the downstream direct-to-user electronic products companies, and Qualcomm, TSMC and Intel all have a strong voice in their respective fields.

In other words, the evolution path of the AI ​ ​ world will be closer to the semiconductor industry or the Internet industry, which means that the biggest cake will appear in the model layer or the product layer? Before answering this question, let’s take a look at the difference between AI and the above two fields.

Different from many industries, the development logic of semiconductor industry comes from the endless demand for performance in consumer market, and finally evolves into Moore’s Law. This forced the manufacturing giants to maintain the technological leadership of each generation of products through huge R&D investment. In this process, it will become more and more difficult to improve performance, and the income growth it brings often does not match the huge R&D cost, which constitutes a very high barrier to semiconductor manufacturing.

But this situation does not appear in the model layer of generative AI for two reasons:

First, it is more difficult for software to keep ahead of technology for a long time than hardware. At present, the input cost of the big model is mainly the training cost, but the training cost is still a drop in the bucket compared with the investment of billions of dollars in making chips. This can also be seen from the development of large models at home and abroad. Although there is still a clear gap between domestic and foreign large models, the gap has been significantly shortened compared with the beginning of the year.

Second, compared with consumers’ unlimited demand for the "performance" of mobile phones and PCs, in many scenarios, users’ intelligence of generative AI is decreasing marginally. In other words, the requirements for the performance of generative AI in all scenarios are not infinite.

Since the model layer logic doesn’t work, it is not necessary to say that the application layer company has great opportunities. The reason is that under the big model logic, product experience and control model are highly bound, and data feedback is very important for model improvement. In this case, it is hard to believe an application-layer company that is highly dependent on big model manufacturers. For application layer companies, it is also a very risky thing to hand over their high-quality data to large model manufacturers for iteration.

From this point of view, both the model layer and the application layer companies have their own problems, and the full-stack companies that occupy both the model layer and the application layer may be able to capture the greatest value.