Meta pursues South Korean AI chip maker to address Nvidia supply crunch

Meta is reportedly in “advanced talks” to acquire South Korean AI chip startup FuriosaAI, according to industry insiders. The potential deal, which could close as early as this month, comes as Meta and other tech giants seek alternatives to Nvidia’s chips.

According to reports, Nvidia is struggling to meet the market’s demand.

A Forbes exclusive published Wednesday, Meta, led by CEO Mark Zuckerberg, is among several companies interested in acquiring FuriosaAI. It is a Seoul-based startup specializing in AI chip design. 

Founded in 2017 by former Samsung Electronics and AMD engineer June Paik, FuriosaAI has received significant backing from investors. These include South Korean internet giant Naver and venture capital firm DSC Investment.

Last week, FuriosaAI announced it had secured 2 billion won ($1.4 million) from CRIT Ventures, a firm led by former Com2us CEO Jaejoon Song. In total, the startup has raised about 170 billion won ($115 million) in venture funding. However, its current market valuation is unknown.

The acquisition, if completed, will help Meta become less reliant on Nvidia’s AI hardware. On that same thought, major cloud and AI companies, including Google and Amazon, have accelerated their custom chip development.

FuriosaAI claims RNGD chip is ‘better’ than Nvidia’s top products

FuriosaAI made headlines last August with the launch of its RNGD chip, which was developed in collaboration with the Taiwanese semiconductor firm Global Unichip Corp. Designed for AI inference tasks, RNGD’s specifications are more efficient compared to Nvidia’s H100 graphics processing units (GPUs).

According to the company, RNGD operates at a thermal design power (TDP) of just 150 watts, compared to high-end GPUs that can consume up to 1,200 watts. FuriosaAI claims the chip delivers three times better performance per watt than Nvidia’s H100 when running advanced AI models.

Additionally, the chip reportedly has HBM3, a high-bandwidth memory from SK Hynix, and is designed for large-scale AI deployments. Because its Llama 2 and upcoming Llama 3 AI models require advanced processing, Meta has a vested interest in such hardware. Mass production of RNGD is planned for the second half of 2024.

Several organizations, including LG’s AI research lab and Saudi Arabian Oil Co., are expressing interest in the AI product. 

Last September, Saudi Aramco penned a memorandum of understanding with FuriosaAI and Cerebras Systems, another AI chip company backed by OpenAI CEO Sam Altman, to explore collaboration in AI and supercomputing.

OpenAI races to develop Its own AI chips

A Monday Reuters report revealed that OpenAI, the company behind ChatGPT, is finalizing the design of its first custom AI chip. The entity plans to send it for fabrication at Taiwan Semiconductor Manufacturing Co. (TSMC) in the coming months.

On January 27, tech startups were surprised when DeepSeek, a Chinese-made AI chatbot, entered the US market and caused Nvidia stocks to plummet by 18% in one day. OpenAI now wants to stay ahead of such market routs, so it is planning ways to alleviate itself from the overreliance the tech market has on Nvidia. 

Sources indicate that OpenAI’s goal is to mass-produce the chip by 2026. The design process, known as “taping out,” typically costs tens of millions of dollars and can take months to complete. If the initial batch encounters technical issues, the company may need to refine and redo the process, potentially delaying production. 

Internally, OpenAI sees the custom chip as a strategic move to gain leverage in negotiations with existing suppliers. The company plans to iterate on its design, creating increasingly advanced processors over time. If its first tape-out is successful, the tech giant could begin testing a viable alternative to Nvidia’s AI hardware later this year.

Cryptopolitan Academy: How to Write a Web3 Resume That Lands Interviews - FREE Cheat Sheet