Wen Shidao.When it comes to growth returns, the "winner-takes-all" economy of scale and network effects has been a killer for the past two decades.
Economies of scale (economies of scale) originally refers to the economic phenomenon that the average cost of an enterprise gradually decreases as the scale of production expands.
But if you follow the original meaning given by Adam Smith, you will find that "economies of scale" not only do not show an increase in growth, but rather a decrease in growth.
For example, a training institution recruits 1 teacher and 5 students. As the enrollment expanded, the number of students became 20. Suppose 1 teacher has limited energy and can only teach a maximum of 20 students. At this point, the training institution reaches the limit of "scale". If the enrollment continues to expand, the training institution will have to hire one more teacher, thus bearing more costs.
In this case, growth goes from "economies of scale" to "diseconomies of scale". In the end, the market competition has reached a dynamic equilibrium, and each enterprise has obtained the corresponding market share according to its own competitiveness.
But this seems to be different from the "winner" world as we know it.
The reason is that in a knowledge-based economic world, where products are knowledge-intensive and the proportion of natural resources is small, incremental returns can be realized. For example, Microsoft spent $50 million to develop the first floppy disk for Windows, while the second and subsequent floppy disks cost only $3. Moreover, unit costs decrease as sales volume increases.
Whoever can occupy the high ground first and iterate quickly through the positive feedback of early users may eat the entire market by a "slight advantage" and achieve "winner takes all".
When this phenomenon is superimposed on the "network effect" of the Internet age, the "winner-takes-all" is further amplified.
A network effect is a product (or service) in which each additional user generates new value for other users of that product. The most typical example is **, the more people in the world who pretend, the more people who can talk, and the greater the value of **.
But this is only the first generation of network effects - "direct network effects".
The Internet era is the world of "cross-network effects" – at least one group of users on a platform who are dependent on each other, but have different interests. For example, buyers and sellers of a treasure; hotels and tenants of a certain trip; drivers and passengers on a drop; A blogger and fan of a certain red book.
The role of the platform is to help bilateral (multilateral) users draw red lines and weave an inextricable network. The more the "red line" of which platform is in place, the more fascinating it is for users. For example, algorithmic push of bytes is based on this logic.
At the same time, because the cost of app replication is almost negligible, the marginal cost is close to zero, the economies of scale are further amplified, and the winner eats up the entire market.
Thanks to the double buff of economies of scale + network effects, almost all of the most spectacular entrepreneurial stories we have heard have taken place in the Internet era.
Taking advantage of the holiday, Shidao re-read Brian Arthur's classic 1996 book "Increasing Returns and the New World of Business". * The "economies of scale" of knowledge-intensive products and the nascent theory of "network effects" at that time were introduced. It was so significant that it was even a game-changer in Silicon Valley.
But as Arthur questioned the law of diminishing returns proposed by Alfred Marshall in the 1890s. Twenty-eight years have passed, but is Arthur's 1996 "network effects and economies of scale in the new world of business" still relevant in the AI era?
01 Weakness: Economies of Scale and Network Effects: At least for now, these two growth theories don't seem to be working so well.
Let's start with the "failure" of economies of scale.
Every time ChatGPT answers a question, it consumes computing power. Microsoft's GitHub Copilot, for example, used to pay an average of $20 per user per month, and some users up to $80. Even the more users, the more you lose. Unless your product is good enough for users to pay for a better experience. However, even in this case, high costs remain, and profits will not be as huge as in the Internet era.
Let's talk about the "weakening" of network effects.
The first reason, from the most intuitive experience, when you use AI software, you are faced with robots, not many people, and the connection between users is no longer complicated. Of course, software developers can also build AI communities and stack "network effect" buffs, such as Midjourney;
The second reason is that while "data" is crucial, its role may be exaggerated. Let's think about the question, will the product model be better because the more user data there is? Or is there an S-curve?
Let's not talk about chatgpt40 "become lazy". In fact, it may not take more data to train the model, a certain level is sufficient. For example, Everlaw, the company in which A16Z invests, has developed a legal software. After analyzing a million emails, the software doesn't need to be trained.
In addition, the data used to train the model must be very suitable for the actual problem. The data barriers of Internet giants are more valuable in quantity than quality.
02 Change: What are the other "reliable" growth theories in the era of Moore's Law?
In addition to economies of scale and network effects. The main independent reasons for the return to growth are learning by doing and recombination of ideas.
Learning by doing refers to the accumulation of experience and knowledge in production by workers, and they will find ways to improve their skills, so as to form technological progress at the overall economic level, get rid of the constraints of the law of diminishing returns, and promote long-term economic growth.
To put it simply, it is that the king of the volume first rolls himself to death and then rolls others.
Under this "King of Volumes" theory, there is a very famous law.
Moore's Law: When ** is constant, the number of components that can fit on a single chip doubles approximately every 18-24 months, and the performance will also double.
Bill Gates once said, "I like Moore's Law as a way to future." It helps us understand the speed at which technology is evolving, and what we should be going to do in the future. ”
However, Moore's Law is not a law of mathematics or physics, but a law based on the development experience of the semiconductor industry. As the number of transistors increases, Moore's Law now hits the ceiling. There are experts** that in 2025, Moore's Law will be invalidated.
Is that really the case? According to the ARK Invest report, since 2014, AI chip performance has increased at a rate of 93% per year, which means that the cost has fallen by 48% per year, faster than the 30% per year of Moore's Law. For example, in 2020, the cost of a single training session for GPT-3 was $4.6 million, and now** it is $1.4 million, a decrease of about 70%. At the beginning of 2023, the API services** provided by OpenAI fell by 90%. If the trend continues, the cost of hardware to accomplish the same quality task will fall from $11,000 in 2014 to 5 cents in 2030.
In addition, since 2012, the amount of computation required for AI models to train neural networks in ImageNet classification to achieve the same performance has also been reduced by a factor of 2 every 16 months.
In other words, the traditional Moore's Law may not work, but the AI era has its own "Intelligent Moore's Law".
1) Huang's Law: GPUs will drive AI performance to double year on year.
In 2020, Huang proposed "Huang's Law" to replace Moore's Law. In 2023, Lao Huang also said: NVIDIA's GPUs have increased AI processing performance by no less than 1 million times in the past 10 years, and will make AI performance 1 million times more powerful in the next 10 years. And Moore's Law, in its best days, grew 100-fold (only) in ten years.
2) The law of large models: the parameters of large models and the amount of training data are growing rapidly.
According to OpenAI's calculations, the global demand for training computing power for head AI models has doubled in 3-4 months, and the computing power required for head training models has increased by up to 10 times every year.
If you follow the traditional Moore's Law, chip computing performance only doubles roughly every 18-24 months. In this case, the pace of chip performance improvement cannot keep up with the appetite of AI training models. However, if the above-mentioned "Huang's law" increases by 1 million times in 10 years, the result is a different story.
In addition, high-quality data does not seem to be sufficient. "MIT Technology Review" once published an article saying that large models are like a "network black hole" that is constantly absorbed, which ultimately leads to insufficient data for training.
The AI research institute Epochai's ** gives a precise time frame: in 2026, large model training will exhaust high-quality data; From 2030 to 2050, all low-quality data will be exhausted; From 2030 to 2060, all image training data will be exhausted. This means that without significant improvements in data efficiency or the availability of new data sources, model size growth will slow by 2040.
3) Altman's Law: The number of intelligence in the universe will double after every 18 months.
Gary Marcus bluntly said that the amount of AI hype doubles every 18 months. In fact, it is no wonder that Marcus is yin and yang, because Altman does not have a clear definition of "amount of intelligence". This sentence is more like a blurted out feeling.
However, back in 2021, Sam Altman wrote an article on "Moore's Law of Everything."'S Law for Everything), which proposes that Moore's Law applies to all things and that AI will reduce the cost of goods and services. A utopian world unfolds: wealth or technology is growing rapidly, and humans can get what they want with cheaper money.
All in all, whether it is the traditional Moore's Law that "peaks" or the new Moore's Law, it shows the explosive exponential growth and rapid iteration characteristics of information technology. This point is probably the eternal "law of growth" that runs through the development of human technology.
03 Constant: Another independent reason for the growth of endogenous technology through patents: recombination of idea – any innovation is the recombination or splicing of different materials.
If a company's product is an idea (algorithm, formula, design), its development process is difficult and long. However, once an idea is developed, its subsequent output becomes the exclusive property of the company (e.g. patents). As a result, the company gets a product with an incremental cost close to zero.
Keen friends may have discovered that this example belongs to one of the three basic premises of Nobel laureate economist - Romer, the pioneering work of endogenous economic growth theory, "increasing returns and long-run growth".
First, technological progress is at the heart of economic growth;
Second, technological progress is largely a conscious act of people. In other words, a reaction to market incentives;
Third, developing a new technology incurs a fixed cost, but the cost of using it afterwards is zero.
Romer postulates that there are three sectors of the economy: the sector that produces the final product, the R&D sector, and the sector that produces intermediate goods.
The R&D department is responsible for producing ideas and selling them to the intermediate goods department; The intermediate goods sector, on the other hand, produces durable capital equipment and leases it to the final goods production sector, which is responsible for producing the final goods of economic weight.
In this model, the ideas produced by the R&D department are external, and the social benefits are not consistent with the private benefits they bring to the R&D department. In Romer's view, in order to encourage R&D, the difference between private and social benefits needs to be eliminated as much as possible, so it is necessary to introduce some incentives such as patents and copyrights.
When we shift our focus from the entire economic world to the micro-enterprise individual, the concept of conceptual restructuring is more like technology development, and building a "moat" with intellectual property is the top priority. Shidao has previously written: Some deep technology companies often face the problem of "chicken or egg first" in fundraising. That is, without a market close at hand, it is quite difficult for companies to raise funds; But without sufficient financial support, it is even more difficult to access the market.
Therefore, in the early stages, it is important to pay special attention to early monetization opportunities, and strategic partners and licensing agreements are "invaluable". For example, Halitus, a deep-tech company, licensed its own IP to an industry leader and partnered with an established startup to sell their products to generate a portion of the necessary revenue, as well as valuable early customer feedback through multiple partners.
04 Discernment: Why Apple didn't usher in a "BlackBerry moment" recently, famous economist Michael JMauboussin and Dan Callahan published an article titled "Increasing Returns: Identifying Forms of Increasing Returns and What Drives Them".
The article does not propose a completely new growth model. But by tracing economies of scale, network effects, learning by doing, concept restructuring, and internationalization, the growth of giants such as Google, Meta, NVIDIA, Microsoft, and Apple is distinguished.
Google and Meta rely heavily on network effects. It is not a restructuring of ideas that distinguishes them commercially. Because other companies can also make very similar products, but can't replicate their networks.
NVIDIA's growth stems from a restructuring of mindsets. Staying ahead of the curve in the future depends on the company's ability to make its technology the industry standard, just like Microsoft's PC operating system. Otherwise, NVIDIA will be replaced by a better or cheaper competitor. To borrow a phrase from Elon Musk: "NVIDIA will not have a monopoly in the large-scale training and inference chip market forever." ”
Microsoft, while benefiting from network effects, relies more on conceptual reorganization.
Just imagine, if a company legally owns all the technology behind the Facebook social network or Google's search engine, this company still can't compete with Meta, Google, because you can't move in a huge number of users overnight. But if a company can legally sell the exact same copy of Microsoft software, it can immediately compete with Microsoft.
And that's exactly why domestic pirated WIN is flying all over the place, but you can't find a second IG or Xiaohongshu.
While Amazon's retail business relies heavily on network effects, its growth returns are largely due to traditional economies of scale and learning-by-doing. As a result, Amazon is investing heavily in hardware infrastructure and is constantly trying new things.
At the same time, Amazon is also a huge beneficiary of international** growth returns. The company has created a technology and logistics layer to deliver cheap manufactured goods from Asian factories to the vast end markets in the West. However, from the current point of view, cross-border e-commerce unicorn SHEIN is challenging Amazon's dominance in this field.
Apple's impressive growth returns are due to the combination of these five independent growth factors.
Although Apple's incremental unit cost is not low: the iPhone 15 Pro Max costs $500-600 to produce. But Apple can protect its rate of return because they never seem to lose pricing power. This can be attributed to network effects and the high switching costs that come with them (consider the lock-in effect of iMessage in this context), restructuring of mindsets (software updates, patents), learning-by-doing and international**.
So when it was thought that one day the fate of Motorola, Nokia and BlackBerry would befall Apple, Apple continued to grow.
From a recent point of view, at a time when innovation in the field of smartphones is almost exhausted, at the end of 2023, Apple will publish two articles in a row** (concept reorganization). One of them proposes that Apple has successfully deployed large models on iPhones and other Apple devices with limited memory through an innovative flash memory utilization technology. And this route of "large model + hardware" may directly change the competitive landscape of AI mobile phones. The second part details a generative AI technology called Hugs (Human Gaussian Splats). Hugs only needs a raw ** of about 50-100 frames, which is equivalent to 24 fps in 2 to 4 seconds**, to generate a "digital human clone" in 30 minutes. And this technology is an inevitable requirement for the further development of VR headsets. The Vision Pro, which was launched not long ago, broke through the limitations of smartphone plane computing and kicked the door to the era of spatial computing.
Inspiration for investors: There is a fundamental difference between "BlackBerry" and Apple – is the return on growth sustainable? Do growth returns from different independent factors have different half-lives? Do companies have the ability to survive the half-life?
Going back to the beginning of the article, Shidao argues that, at least for now, the network effect may be weakened, but the power of concept restructuring (patented technology) will always shine.