[Are NVIDIA and OpenAI really popular?]

MZ investors are unaware, but around 2000, investors were enthusiastic about internet stocks. Internet bookstores, internet supermarkets, and internet schools were enthusiastic about the colorful and wonderful news of BM and stock-rich people that the world would change quickly.

The internet and WWW services were truly new and shocking. I was able to buy books and things without having to go to the store, and I was able to enjoy magazines and newspapers on the computer. The existing industry seemed to collapse quickly, and people were thrilled by this amazing service and technology.

More than we expected 24 years ago, 2024 Today, we live in a world where the Internet is as natural as air. However, many internet companies that people were enthusiastic about and invested in at that time are not left. Amazon survived, but many of the H/W and S/W companies that dominated the time, such as Cisco, have either gone bankrupt, been absorbed, or remain in name only.

As Bill Gates said, people usually tend to over-expect changes over a year and under-expect changes over a decade. However, when a company runs out of cash, it’s hard to last six months.

Nvidia’s high stock price and OpenAI’s high valuation should eventually be supported by its ability to generate cash. Both depend on how much value LLM technology creates, creates markets, and provides consumers with high cost-effectiveness.

Including SORA, innovative and interesting services are emerging, and various BMs that are destructive are pouring out. However, let’s think calmly. In order for a BM to actually make money and break down an existing industry, it must create more value than the input resource exponentially. At this time, competition with existing industries must also be considered.

SORA can shoot plausible advertising videos. It could reduce the cost of producing existing advertisements to 1/10. However, will SORA really produce quality videos that people will be impressed with? How much resources should we devote to producing such videos?

The reason why the Internet economy has changed the world by combining mobile since 2010 is that digital data costs almost zero won. Once the infrastructure is in place, ordering delivery on a smartphone and buying a train ticket are much cheaper than making a phone call and going to the counter. But will the cost be reduced if LLM learns multiple times to achieve the desired results?

If you make a one-minute ad with 8K videos, it’s 24 frames per second, so you have to create 1440 images. Let’s say the director goes through 100 detailed revisions to reach the level of quality he wants. In other words, please make it with ABC… Make it ABD… Make it ABDc… Through the process, the modifications and generations are repeated.

How much computation and energy will this take? Nvidia’s new B200 graphics card is expected to consume 1 kW of power, so how many B200s will it take and how many times will it take to generate 1440 8K images 100 times? How much load will it take if millions of users demand a variety of “create” services?

Recently, in the United States, we have begun to talk about the power consumption of [AI-only processing data centers] in GW units. For your understanding, let’s look at a rather extreme hypothetical case. The so-called hyperscale data center has a size of at least 100,000 servers (a little smaller than gaming PCs). If eight B200s are installed on a single server, 800,000 B200s are turned on in the AI data center. Consider turning on 800,000 microwaves or irons at the same time inside a building because 1 kW of power is the amount of power you need to turn on a microwave or an iron.

This data center requires at least 0.8 GW of electricity. Suppose that 0.2 GW of electricity is used to cool the tremendous amount of heat, and the cooling system is operated. Since one center requires 1 GW of electricity, and one APR1400 nuclear power plant can generate 1.4 GW of electricity, one data center and one nuclear power plant must be connected.

Then, how expensive is the B200? The underperforming H100 costs about 50 million won per unit. How much money should we spend to build one of the latest AI data centers? The price of 800,000 B200s alone, excluding facility construction costs, is 40 trillion won. Suppose that the LLM business is booming, so it runs for 8,760 hours a year with 100% of the load. Then, if this AI-only processing data center uses at least 10 TWh of electricity a year, how much is it? If electricity is applied these days, it is 1.8 trillion won.

How much annual profit do I have to generate to make a profit by building this facility and operating it for 10 years? If the life of the B200 chip is 10 years, you have to make a profit of approximately 5 trillion won per year to recover your investment. If the profit margin is 20%, you have to make 25 trillion won in sales every year. It is based on the AI data center building 1.

To sum up, Nvidia’s customers are companies that invest W40 trillion in one AI data center, spend W1.8 trillion on electricity every year and W25 trillion in sales. OpenAI intends to make money by providing LLM services with such facilities. Except for AGI, it will eventually do this structure of business, will it be sustainable? Really?

Of course, it is a very extreme case. Meta and Microsoft bought the most H100s last year, with only 150,000 units each. There was also a shortage of supply. AI processing servers can be inserted into the server rack of existing data centers, or AI data centers can be built in several small locations. However, it is not too late to consider whether an investor is an investor in 2024, the climate crisis and the energy transition period, when the LLM business itself must manage the demand for electricity and water well and earn more than the investment funds.

(Photo is a server with 8 H100s in it)

(Farewell. I am overwhelmed when many people share and comment on my embarrassment that I wrote while waiting for the meeting in the morning. You made such comments as the advancement of steps, the limitation of 3 nanometers of width, the proprietary structure of N company, the structure of DC, neuromorphic, quantum computer, etc. They are all correct, and it is just my personal opinion that I lack. This is not a serious and serious article, but a “crush,” so I can’t believe it… I would appreciate it if you could let me know.

However, regarding the development of the steps, there is a part that I deleted when I was writing the original text and was told to come into the conference room, and when we supply products and services in earnest through LLM service, we have to consider consumer demand for high quality. This is because I myself have experienced errors that continue to remain in the results while working with ChatGPT, so I stopped subscribing. I hope many researchers will develop better AI services and servers to enrich our lives. I am also attaching the result of asking Claude about this topic. We will find out whether LLM is successful or not in time, right?)