AI Weekly: ByteDance Denies TSMC Chip Deal; Runway Enters Hollywood

Microsoft, BlackRock, and others announce the establishment of a $30 billion AI investment fund.

On September 17th, local time, Microsoft and BlackRock, among other companies, announced a partnership in AI investment, jointly investing in data centers and other AI-supporting infrastructure.

This strategic cooperation, named the "Global AI Infrastructure Investment Partnership" (GAIIP), aims to make the fund the world's largest AI investment fund, attracting $30 billion in private equity investment, including debt financing, with a potential total investment of up to $100 billion.

In addition, AI chip leader NVIDIA will support GAIIP, including leveraging its expertise in AI data centers and AI factories to jointly build an AI ecosystem.

This cooperation shows that Wall Street still has a huge investment interest and confidence in the field of AI.

It is introduced that GAIIP will support open architecture and a broad ecosystem, and infrastructure investment, including energy projects, will be mainly deployed in the United States, with some funds deployed in its partner countries.

Advertisement

Saudi Aramco partners with Groq to build the world's largest inference data center.

AI chip design company Groq officially stated that it has signed a memorandum of understanding with Aramco Digital, the digital and technology subsidiary of oil giant Saudi Aramco, to build the world's largest inference data center in Saudi Arabia.

It is reported that the advantage of the Aramco Digital-Groq large data center in Saudi Arabia lies in the low local energy prices and sufficient construction space, and at the same time, the network time distance is not more than 100ms ping with nearly half of the global population.

It is reported that the data center will be completed and put into operation by the end of this year, initially including 19,000 Groq's LPU language processors, which can process billions of tokens per day, and by 2025, its processing capacity will be increased to hundreds of billions of tokens per day, serving millions of developers.

The data center in Saudi Arabia is expected to reach a scale of 200,000 LPUs in the future.

Saudi Aramco will provide financial support for the project, with a scale estimated to be "about nine figures" in US dollars.

Runway enters Hollywood and cooperates with Lionsgate to develop customized AI video models.

On September 18th, AI video generation platform Runway announced a cooperation with Lionsgate.

The core of this cooperation is that Runway will use AI technology to help Lionsgate's filmmakers, directors, and creative teams complete the complex links in movies more efficiently.

Specifically, AI models will first be applied to storyboard production, background generation, and special effects design, especially for films involving a large number of action scenes.

These scenes are traditionally not only time-consuming and costly but also come with certain safety risks.

AI technology can automatically generate realistic scenes, special effects, and backgrounds, enabling production teams to quickly preview and iterate, significantly reducing production cycles and costs.

The application of this technology may completely change the key links in film and television production, bringing huge efficiency improvements to the industry.

Comment: It is worth mentioning that this is the first generative AI company to publicly cooperate with a major Hollywood studio.

The day before Lionsgate reached a deal with Runway, California in the United States signed a bill restricting the use of AI digital replicas in movies and television projects.

Runway is still facing a lawsuit, and the company is accused of training models with copyrighted works without permission.

AI search startup Glean raises $260 million, with business growth three times in one year.

AI search startup Glean founder Arvind Jain recently announced on his blog that he has successfully completed an E-round financing of more than $260 million, with a valuation of $4.6 billion.

This round of financing was led by Altimeter and DST Global, with Craft Ventures and SoftBank Vision Fund II, among others.

The founder also revealed in the blog that Glean's valuation has more than doubled in six months, and the company has more than $550 million in cash after financing.

Along with the financing, Glean launched several new features of the Work AI platform.

Public information shows that Glean is a company that provides enterprise AI search services, and since its establishment in 2019, Glean's demand has surged, especially in the deployment of generative AI.

Currently, Glean's Work AI platform has been adopted by technology companies such as Reddit, Pinterest, and Duolingo, as well as leading enterprises in various industries such as telecommunications, banking, and retail.

Microsoft seeks US permission for AI chip exports to the Middle East.

Microsoft said it hopes the US government will provide more "clarity and consistency" in export controls, as these controls have delayed the delivery of cutting-edge AI chips to the Middle East.

This year, Microsoft invested $1.5 billion in G42, hoping to use this UAE's largest AI company as a gateway to enter the African and Asian markets, and Microsoft believes that these regions have unsatisfied and growing demand for AI.

However, the United States has restricted the export of AI-specific chips to the Middle East, and Microsoft has not yet obtained the component export license required for its cooperation plan with G42.

Comment: AI chip sales are not entirely a free-flowing market.

Last year, the United States expanded the sales restrictions of NVIDIA and AMD's AI chips to the Middle East, and there were also reports that the United States slowed down the issuance of export licenses for chip manufacturers to export such chips to the Middle East.

For the global layout and global business cooperation of AI companies, chip sales restrictions may cause obstacles.

ByteDance denies cooperation with TSMC to make chips.

In response to previous reports of "ByteDance plans to cooperate with TSMC on AI chips," ByteDance responded that the report is untrue.

The company does have some exploration in the field of chips, but it is still in the early stage, mainly focusing on cost optimization of business such as recommendations and advertisements, and all projects are also in full compliance with relevant trade control regulations.

Comment: In the past two years, there have been continuous rumors about ByteDance making chips in the market.

In June this year, there were reports that ByteDance cooperated with US Broadcom to develop advanced AI processors to ensure the stable supply of high-end AI chips.

At that time, ByteDance responded that the news was untrue.

In addition to ByteDance, some top technology factories have also tried to make chips in recent years to reduce procurement costs and gain more autonomy.

Alibaba Cloud launches open-source large model Qwen2.5.

On September 19th, Alibaba Cloud released the new generation of open-source models Qwen2.5, among which, the flagship model Qwen2.5-72B's performance exceeds Llama 405B.

Qwen2.5 covers a large language model of multiple sizes, multimodal models, mathematical models, and code models, each size has a basic version, instruction following version, and quantization version, totaling more than 100 models.

In terms of multimodal models, Alibaba Cloud also announced the open-source of the visual language model Qwen2-VL-72B.

Alibaba continues to layout open-source models, and as of mid-September, the cumulative download volume of Tongyi Qianwen open-source models has exceeded 40 million.

At the same time, the calling price of Tongyi Qianwen API on Alibaba Cloud Bai Lian has decreased by 97% over the past year.

Although the application scenarios of large models are still limited, open-source and price reduction are still beneficial to the expansion of applications.

Alibaba Cloud Intelligence Group Chief Technology Officer Zhou Jingren said that open-source and price reduction are all a logic, that is, how to develop the ecosystem.

Humanoid robot startup 1X launches the world model.

The humanoid robot company 1X, invested by OpenAI, has launched its own world model.

1X said that based on the progress of video generation and world models for autonomous driving cars, the company has trained a world model as a virtual simulator for robots.

The model can predict object interactions, such as the effects of rigid bodies and object drops, but the model may not maintain the shape and color of objects during the interaction process.

The model required by robots is not a large language model, but a world model that can recognize, understand the physical world, and give feedback.

This is also why humanoid robots carrying large language models at this stage cannot reflect real intelligence.

At present, many humanoid robot companies only focus on the body and do not participate in the construction of large models themselves.

1X's business model has a certain uniqueness, and the effect of its world model is worth looking forward to in the future.