LOGO

On the 5th, the market capitalization of the US semiconductor giant NVIDIA surpassed $3 trillion (approximately 468 trillion yen) for the first time.

NVIDIA’s market capitalization first broke the $2 trillion mark in February this year. The company’s performance has been rapidly expanding, supported by semiconductors for generative AI. It has increased its presence as a driving force in the US stock market, surpassing $3 trillion just over three months after exceeding $2 trillion.

Sales to data centers, including AI, are boosting performance. In the financial results for February-April 2024 announced in May, net profit increased to $14.881 billion, about 7.3 times the same period last year, and sales increased to $26.044 billion, about 3.6 times.

The Strength of NVIDIA: The Background

NVIDIA manufactures GPUs (graphics processing units). GPUs are chips designed for display functions such as video, image, and animation display and have been used to display images smoothly in gaming PCs and other devices.

Now, the demand for GPUs is expanding. The catalysts for this are “data centers” and “generative AI” represented by ChatGPT.

Until now, it was common for data centers to be equipped only with CPUs (Central Processing Units), but with the spread of AI, the trend is for data centers to be equipped with GPUs in addition to CPUs. However, only about 10-20% of data centers are currently equipped with GPUs.

This situation will change with the spread of generative AI.

Generative AI, such as image generation and natural language generation, requires a process called “inference” to run an AI model created by learning and reach a conclusion. When you ask ChatGPT a question, the answer you get is the result of “inference.” The inference process requires more calculations than the learning process. Therefore, it is necessary to also install a GPU, which is suitable for large-scale calculations.

In the future, it is clear that generative AI will become the main task of generating information in most data centers worldwide, and it is said that within 10 years, most data centers in the world will be equipped with GPUs.

NVIDIA’s “CUDA”: The Standard for AI Developers

Unlike CPUs, GPUs are good at performing large amounts of calculations simultaneously in parallel, but to bring out their capabilities, a development environment for GPUs is required.

NVIDIA’s “CUDA” is one such “development environment for GPUs,” and since CUDA has become the de facto standard among neural network researchers, many libraries have been created on top of it. Now, at least when it comes to the learning process, there is almost no choice but to use CUDA.

CUDA is a development environment created by NVIDIA for its own GPUs, so as a result, it has become the standard to “use NVIDIA in the neuron learning process.”

CUDA protects NVIDIA as a competitive “moat” that rivals cannot easily overcome.

Will NVIDIA Have a Near Monopoly for the Time Being?

According to the British research company Omdia, NVIDIA will have a 77% share (2023) of the global market for AI semiconductors for data centers. The company’s cutting-edge GPUs are highly sought after by companies developing AI.

For the time being, NVIDIA is expected to have a near monopoly on the GPU market for data centers. However, we will also be keeping an eye on the movements of other companies that deal with GPUs.

In May 2024, OpenAI released the latest model of ChatGPT, “GPT-4o”.

It is a cutting-edge multimodal AI that can process text, voice, and images in an integrated manner, and it is attracting attention because it will also be implemented in the free version of ChatGPT.

Generative AI has a major impact on the data centers that are being newly built by GAFA and other Japanese companies, and we would like to take a look at how OpenAI’s latest version of ChatGPT, “4o”, which is a representative example of generative AI, is different from the previous version.

What is GPT-4o?

ChatGPT-4o (Omni) is the latest model of ChatGPT announced by OpenAI in May 2024. Omni means “all” in Latin, and represents the ability to handle all information, including not only text but also images and voice, and perform any task.

Compared to the conventional model GPT-4 Turbo, the answer accuracy and speed have been overwhelmingly improved, and it has been upgraded in every respect, such as being able to have emotionally rich voice conversations like humans and reading the fine details of images.

What are the features of GPT-4o and how does it differ from other models?

The GPT series is a large-scale language model developed by OpenAI, and its performance improvement is remarkable.

GPT-3, announced in 2020, attracted attention as a large-scale model with 175B parameters. In 2022, GPT-3.5 was implemented in ChatGPT, widely publicizing the potential of language generation AI through dialogue with general users. And in 2023, GPT-4 showed the first step toward multimodalization.

GPT-4o is positioned as an extension of the evolution of this GPT series. However, it stands out from conventional GPTs in that it does not just improve performance, but also achieves smooth integrated processing of voice, images, and text.

The main evaluation points that have been significantly improved compared to conventional models are introduced below.

① Text accuracy

It boasts high accuracy in understanding and generating complex sentences. This allows for more natural and consistent text generation.

You can also easily create article structure plans, which are essential for writing.

② Text and voice response speed

New algorithms have improved text and voice response speeds, making real-time dialogue even smoother. In addition, the voice has intonation, making it feel like you are talking to a person.

③ Voice recognition and translation function

The accuracy of the voice recognition function has been improved, and the multilingual translation function has also been enhanced. This makes global communication more efficient.

It is also possible to translate in real time by recognizing and processing voice.

④ Improved image recognition function

Image recognition capabilities have also been improved, allowing the content of images to be analyzed with high accuracy and related information to be provided.

It is also possible to extract characters from image data. For characters that are difficult to read, characters can be inferred from other image data and extracted.

⑤ Security function

A new tokenizer has been introduced in 20 languages, including Japanese, and significant improvements have been made in terms of security. This has improved data security and processing efficiency, and has enabled fast and secure data processing while protecting user privacy.

Evolving ChatGPT

ChatGPT has added many surprising features in this update, such as improved image processing capabilities and the addition of a voice recognition function.

In the future, it will be possible to converse via real-time video, and a new voice mode is planned to be released that will allow the contents of the loaded video to be explained in voice.

The development of ChatGPT, which is leading the generative AI, will have a major impact on future data centers, so we will continue to watch the situation from time to time.

Meanwhile, as expectations grow for new functions to be developed in the future, power consumption is expected to increase several times over.

In Japan, how will the power shortage of newly opened data centers be resolved?

We will also be keeping a close eye on this.

On April 18, US Oracle announced it would invest $8 billion (approximately 1.2 trillion yen) in data centers in Japan over the next 10 years. Additionally, US OpenAI also announced its entry into the Japanese market.

Along with other major US cloud companies like Microsoft, the total amount of announced investments in Japan this year is approaching 4 trillion yen. What is behind these US cloud giants’ emphasis on data centers in Japan?

The Spread of Generative AI and Response to Security Risks

One underlying factor is the rapid spread of generative AI (artificial intelligence). Among user companies, the demand for cloud services to train and operate large language models, which serve as the foundation, is increasing. German research firm Statista predicts that the market size of data centers in Japan will reach approximately $24 billion by 2028, expanding to 1.4 times the size of 2023.

However, security risks associated with cloud services are emerging. Surveys by sources like the Nikkei indicate that about half of companies lack sufficient regulations regarding disclosure requests from authorities in various countries. Many Japanese companies depend on storing data overseas, with some even placing data in countries like China and Russia, where concerns about censorship exist, making immediate action necessary.

With heightened awareness of security and privacy, regulatory authorities in various countries and regions are increasingly emphasizing data sovereignty, managing their own data domestically. The Japanese government also restricts cross-border transfers of personal data under the Personal Information Protection Law. Japanese companies are being urged to manage sensitive data domestically.

To meet these needs, major US cloud companies are announcing large-scale investments in Japan one after another. The focus on Japan extends beyond the AI field. The world’s largest semiconductor foundry, Taiwan Semiconductor Manufacturing Company (TSMC), is investing about 1.3 trillion yen to mass-produce computing semiconductors at a factory built in Kumamoto Prefecture by the end of 2024. They have also decided to invest about 2 trillion yen in constructing a second factory aimed at commencing operations in 2027.

Previously, TSMC concentrated its production bases in Taiwan, but considering the risk of Chinese aggression, it is diversifying production bases to countries like Japan, the US, and Germany. The construction of the Kumamoto factory is part of this strategy, and the importance of Japan, where related industries gather and semiconductor demand is high, may further increase. This situation is likely to continue influencing the actions of major cloud companies.

AI as an Essential Element for Japan’s Economic Growth

In Japan, a country prone to earthquakes and high electricity costs, data center costs are considered higher compared to overseas. Nonetheless, US Amazon Web Services (AWS) and Google, which compete with Microsoft in cloud services, are also embarking on large-scale data center investments domestically.

Microsoft President Brad Smith stated about Japan, “With an aging and declining population, AI is an essential element for sustainable economic growth.”

Going forward, we should continue to pay attention to the potential of AI and the trends of major cloud companies for Japan’s economic growth.

In March 2024, NVIDIA, a leading semiconductor company, announced the launch of the next-generation AI semiconductor “Blackwell B200” GPU.

What is B200?

“Blackwell B200” is a combination of two chips of the same size as the company’s previous products, and the number of transistors that greatly affect performance is 208 billion, compared to 80 billion in the previous main product “H100”. Approximately 2.6 times.

What is the price of B200?

The B200 is expected to ship in the latter half of 2024, and while pricing is still unclear, it is expected to be higher than the “H100” and “H200.”

However, the announcement that major cloud service providers such as Amazon, Google, Meta, Microsoft, Open AI, and Oracle are expected to adopt this new semiconductor underscores the significant impact this technology will have on the industry.

NVIDIA is not only accelerating the advancement of AI technology but is also solidifying its position as a key player in the industrial sector.

As the processing power of GPUs improves, there is no doubt that AI will be increasingly utilized. The development of generative AI is also expected to increase the production of content that requires substantial data volumes, such as videos and music, which were previously difficult to generate. Thus, the demand for data centers is expected to continue rising.

We will continue to monitor industry trends.

In February 2024, OpenAI, famous for Chat GPT, announced a new video generation AI named “Sora”. As of April 2024, however, “Sora” is not yet available to the general public, and its pricing details remain undisclosed.

“Sora” is a next-generation AI system that builds upon OpenAI’s large-scale language model, GPT-4, with further enhancements.

With just a simple instruction, it can easily create high-quality videos up to one minute long.

While video generation AIs existed before, they could only produce videos of a few seconds, making “Sora” superior in both length and quality.

The launch of “Sora”, equipped with GPT-4, is undoubtedly set to significantly impact the world.

This article will provide a detailed introduction to the video generation AI “Sora”, including its overview, pricing, user experience, and applications.

Examples of “Sora” Creations

The following videos are actual examples created by “Sora”:

1. Balloon Head Prompt: Sunny, our balloon-headed boy, embodies the blue sky feeling of boundless potential that we felt when we first began using the tool. Our heads filled with so many ideas, it felt like they might POP.
    

1. Gold Record Bullying Prompt: Exploring space-time with Sora. This isn’t going to replace the filmmaking process, rather, it’s offering an entirely new way of thinking about it. Not restricted by time, money, or other people’s permission, I can ideate and experiment in bold and exciting ways.

Pricing of “Sora”

As of April 2024, the pricing structure for “Sora” after its public release has not yet been announced. However, based on previous patterns like Ghat-GPT, it is likely that both free and paid versions will be released, with more advanced features available only in the paid version. Initially, all features might be offered for free, with some becoming paid later on.

Applications of “Sora”

The potential future applications of “Sora” include:

1. Movies and animation videos

2. Virtual reality (VR) experiences

3. Educational explanatory videos and tutorials

4. Breaking news and weather forecasts

5. Product promotion videos

6. Short videos for social media

7. Personalized message videos

8. Virtual tour videos for real estate properties

9. Virtual event videos like fashion shows

10. Automated editing of sports highlights and commentary videos

11. Simulation videos in the medical and scientific fields

12. Documentary videos about historical events and figures

13. Music videos and live performance videos

14. Video manuals for automobiles and household appliances

15. Visualization and reporting videos of stock market and financial data

16. Presentation videos for architectural and interior design

17. Emergency evacuation routes and safety procedure videos

18. Language learning conversation scene videos

19. Customized videos for personal memories and anniversaries

These are just examples, but the ease of creating videos can cater to viewer demographics and needs, allowing for the creation of original videos. The ability to easily produce virtual imagery, which used to be costly, will likely be a major advantage.

Video Generation AI and Data Centers

Plans for new data centers dedicated to generation AI are being announced in succession. Given the large data capacity required for easily creating videos, the demand for data centers is expected to continue growing.

NVIDIA announced its fiscal year earnings for January 2024 on February 21st. The annual revenue reached $60.9 billion, marking a 126% increase from the previous year, with operating profit soaring to $33 billion, an increase by 7.8 times.

For the fourth quarter (November to January), the revenue was $22.1 billion, a 265% increase year-over-year, exceeding the company’s forecast of $20 billion. The operating profit for the quarter reached $13.6 billion, a tenfold increase, with an operating margin surpassing 61%.

In response to the strong performance, the after-hours stock price surged by over 7%. The sales forecast for the February to April period is set at $24 billion (±2%), with a confident outlook stating, “We will continue to grow in 2024, 2025, and beyond.”

Half of data center sales are for the cloud

Demand for data centers continues to drive NVIDIA’s performance.

Data center revenue has been increasing quarterly, with the third quarter reaching $14.51 billion, a 279% increase. The previous quarter saw revenue of $10.32 billion, a 171% increase, and the first quarter reported $4.28 billion, a 14% increase.

In the latest quarter, more than half of the data center revenue came from major cloud providers.

Jensen Huang, the founder and CEO of NVIDIA, stated, “Accelerated computing and generative AI are at a turning point. There is a surge in demand worldwide across corporations, industries, and nations.” He claims that the data center installation base amounts to approximately $1 trillion and predicts the emergence of data centers worth $2 trillion that will power global software over the next four to five years.

Will NVIDIA Continue Its Dominance?

The demand for data centers is expected to grow further. With the expansion of generative AI, the demand for high-performance semiconductors is surging, and NVIDIA is extending its lead in securing orders.

On March 19th, Hitachi, Ltd. announced a collaboration with NVIDIA in developing AI services. The partnership aims to digitize railways and infrastructure facilities in virtual spaces to improve maintenance efficiency.

Moreover, NVIDIA and Chinese EV companies have announced an expansion of their partnership to enhance autonomous driving technology.

While NVIDIA’s dominance seems likely to continue for the time being, we will keep a close eye on the company’s developments.