OpenAI calls out NVIDIA, GPU seriously insufficient! Hardware requirements resonate; Alphabet-C brings StyleDrop to explode the AI graphics circle; Microsoft will fully AI-ize Windows by the end of the year | Jianzhi Research

Weekly Highlights

GPUs are in severe shortage, and there is a resonance in demand for optical modules and storage PCBs.

Overseas:

1. Google's StyleDrop is back to blow up the scene with OpenAI graphics, which are more creative and have more controllable styles;

2. Apple is launching the OpenAIGC+MR strategy, and OpenAI+XR will become the next generation of mobile battlefield;

3. NVIDIA team launches 3D video OpenAI large model, making virtual reality more realistic;

4. Microsoft will launch Teams 2.0 at the end of the year, and the operating system will launch a total attack on OpenAI;

Domestic:

1. Many regions have favorable policies, and Beijing, Shanghai, and Shenzhen have successively introduced OpenAI development plans;

2. TENCENT invests in large models for the first time! OpenAI start-ups are at the forefront of the trend;

3. Alibaba Cloud's OpenAI assistant "Tongyi Tingwu" is in public beta, and the landing speed on mobile terminals will exceed expectations;

4. The domestically produced self-developed database, Toerther Haibei, has strong demand in the financial and government fields;

5. The Chinese OpenAI large model initiates open source governance, completing the offensive and defensive between "poisoning" and "detoxification" with OpenAI;

Insights from Jianzhi Research

OpenOpenAI and Supermicro are calling out NVIDIA for insufficient GPUs!

The biggest complaint from OpenOpenAI's customers at present is the reliability and speed of the API. OpenOpenAI's CEO Sam Altman admits that GPUs are currently in very short supply, which has delayed many short-term plans. The adjustment of the API and dedicated capacity products are all limited by the availability of GPUs. However, OpenOpenAI will still provide dedicated capacity and private copies of models for customers. But if customers want to access this service, they must commit to paying $1 million in advance.

Liang Jianhou, founder and CEO of Supermicro, said that the market demand for OpenAI is strong, and the company is expanding production capacity in the United States, the Netherlands, Malaysia, and Japan. It is expected to increase the production capacity of 4,000 cabinets to 5,000 cabinets by the end of the year. He also told Huang Renxun to provide more chips, even if they have already been provided but are not enough.

Jianzhi Research believes:

Driven by the demand for generative OpenAI, GPU products will face sustained shortages and price increases. NVIDIA's current delivery cycle is still getting longer, from one month before to now basically requiring three months or more, and even some orders can only be delivered by the end of the year.

In addition, NVIDIA has also released the powerful OpenAI computing platform GH200, which is used for large model training, not only faster but also more cost-effective. Google Cloud, Meta, Microsoft, and software have all announced that they will be used for generative OpenAI work. For the industry chain, it has become a common consensus that the amount of optical modules has increased, and at the same time, the growth of storage and PCB needs has slowly been realized. Nvidia's high-end GPU has driven the continuous surge in demand for HBM storage chips for two consecutive years. The orders for HBM from Samsung Electronics and Hynix, two major storage factories, are rapidly increasing.

Overseas

StyleDrop can capture the subtle differences in texture, shadow, and structure of various styles, with just one image as a reference. It can deconstruct and replicate even the most complex artistic styles. Even Nvidia scientists have called it a "phenomenal" achievement.

Jianzhi Research believes that:

Compared with the previously popular MidJourney tool for generating images, StyleDrop can better control the style of the generated images, and the generated content will be more in line with the needs of designers. MidJourney, on the other hand, avoids the plain camera effects in daily life, and increases the overall realism when generating ultra-clear images. In addition, it tends to lean towards content and aesthetic preferences.

However, both can draw inspiration from other artistic media and painting styles for creation.

The market has high expectations for MR, and considering Apple's significance as a trendsetter in consumer electronics, the MR that has been awaited for 7 years may drop a bomb on the XR industry. Everyone is looking forward to the new changes that MR will bring in terms of technology and user experience.

Jianzhi Research believes that:

The rapid development of generative OpenAI combined with MR will bring about a comprehensive upgrade of mobile products, especially in terms of innovative application content, which will break through the previous development methods and greatly improve the problem of the lack of popular XR game categories at the current stage.

This will also become an important factor for the sinking of the MR market. The difficulty in breaking through the bottleneck of XR game penetration after entering the growth stage lies in the niche of the application ecology. However, the number of loyal fans in the Apple ecosystem is extremely large. Under the all-round high-quality integration of content + terminal + ecology, it will help to quickly sell MR and drive a new round of development cycle for the XR industry chain. In the OpenAI Daily, we also analyzed the impact of Meta, the VR market giant, announcing the early sale of Oculus 3 in the autumn.

NVIDIA Research has developed a new OpenAI large-scale model Neuralangelo, which is an OpenAI model that uses neural networks to perform 2D reconstruction of 3D video clips. The new model can convert the video of any device into a detailed 3D structure.

According to Jianzhi Research:

Although 3D generation technology has long existed, it is worth noting that Neuralangelo, this OpenAI large model, significantly surpasses all previous methods in the ability to convert 2D videos into 3D objects. The model selects images taken from different angles from the 2D video to obtain details of the 3D object representation, and finally renders to improve detail clarity. The feature of this model is that it uses Nvidia's solution to better construct video details, making the content look clearer and suitable for both small sculptures and large constructions.

Especially pay attention to the fields that can be widely applied in the future: such as virtual reality, digital twins, robot development, and industrial digitalization that use 3D object construction in large-scale scenes.

Microsoft plans to start using Teams 2.0 by default on the Win10 and Win11 platforms by the end of 2023; release the Teams 2.0 preview version to Mac, VDI, and web users, and further promote it to other customer groups such as education and government.

The new version of Teams promises to increase installation speed by 3 times, start-up time by 2 times, and increase the switching speed between chat and channels by 1.7 times. The speed of joining meetings should also be 2 times faster; memory resource usage is reduced by 50%, and disk space is reduced by 70%.

According to Jianzhi Research:

Teams 2.0 is embedded in Microsoft, and the impact on the operating system will be earth-shaking. This will greatly accelerate the process of OpenAI on the PC side, including the convenience and intelligence of video conferencing, OpenAI chat assistants, Office365, and many other tools, which will completely change users' usage habits. It is particularly noteworthy that the upgraded Teams 2.0 has a smaller memory footprint and faster speed, making multi-threading and high-frequency use not particularly laggy.

Domestic OpenAI

According to Jianzhi Research:

Local governments will introduce policies to encourage the development of the OpenAI industry, from the construction of underlying hardware computing to the research and development of intelligent robots with application ends, which will enter the policy dividend period in order to create a better and more open environment to promote the rapid development of the OpenAI industry. Yesterday, Beijing and Shanghai also introduced new policy plans for OpenAI; including the implementation of the computing power partner plan, strengthening cooperation with cloud vendors, providing diversified and high-quality affordable computing power; supporting private investment in major projects, participating in data, computing power, and other artificial intelligence infrastructure construction. Overall, the progress of large-scale model development in China is very fast, and open source large-scale models now have security databases. The development of OpenAI's application side, such as media IP and games, is also rapidly landing. In the future, we should focus on the development of embodied intelligence. This field is still in the early stages, and innovative development opportunities are worth looking forward to.

Jianzhi Research believes:

Due to the wave of large models driven by ChatGPT and OpenAI's development, many star start-up companies have emerged one after another. MiniMax, which has only been established for more than a year and a half, has become the most attractive player in the venture capital field. In November of last year, it released the virtual chat software product Glow; in March of this year, it launched the generative dialogue OpenAI assistant Inspo; it also launched an API open platform for enterprise users, supporting text and voice model service calls. Since its inception, MiniMax has grown rapidly, with an overall valuation of more than 1.2 billion US dollars. This is the first time TENCENT has started investing in large-scale model start-ups. With the recognition and pursuit of capital, the entrepreneurial atmosphere of OpenAI will become more active.

Jianzhi Research believes:

The landing progress of large-scale models in China's application field is very rapid. Tongyi Tingwu is mainly used in the audio and video fields, bringing users a brand-new experience of recording and reading audio and video content. The user stickiness of traditional software will be quickly broken. It is worth noting that in terms of content summarization, Feishu Miaoji can only provide keywords; while Tingwu can provide corresponding speech summaries for different guests' speeches. At the same time, attention should be paid to the application progress of voice large-scale models on mobile terminals, such as smart speakers, which are very good ports.

Jianzhi Research believes:

Haibei is a pure domestic search engine database that is completely self-developed from the underlying word segmentation algorithm to the core engine and upper-layer system. It has higher-level security, compatibility, and high-performance retrieval features, not only can achieve full-field indexing, support arbitrary dimension combination queries, but also can automatically partition cold and hot data, and support multiple storage hybrid use.

In terms of application layer, especially for fields with strong specialization and high security, such as banking, government affairs, and military industry, it has shown strong competitiveness.

Jianzhi Research believes:

Data annotation is a crucial step in the large-scale model process. Using the annotated "safe data set" for model training can achieve training results that approach the ideal. However, data standards have always been accompanied by subjective, religious, and personal preferences. Therefore, if foreign data sets are used for training, they will be somewhat "unaccustomed". Therefore, building a local training data set is very important. The first Chinese OpenAI anti-discrimination project has gathered many industry experts and will become one of the high-standard data sets for open source large-scale model training in China.

Next week's focus

Apple's WWDC conference, can MR live up to expectations and lead the XR industry into a new era.