• South Korean startup FuriosaAI rejected an $800 takeover from Meta
• Deal collapsed over post-acquisition plans despite above market valuation
• FuriosaAI chose to stay independent and will launch its RNGD chip this year

As demand for AI infrastructure continues to grow, hyperscalers like Meta, Microsoft, and Google are actively working to reduce their reliance on Nvidia’s pricey hardware by developing their own custom silicon. Meta has reportedly begun testing its first in-house AI training chip, as part of a broader push to gain more control over its AI stack.

In addition to its internal efforts, Meta has also been exploring acquisitions of promising AI chipmakers to bolster its hardware capabilities.

One such target was Korean startup FuriosaAI, which the tech giant saw as a strong candidate to help accelerate its AI infrastructure ambitions. Meta made a buyout offer worth $800 million (approximately 1.2 trillion won) for the company, but despite the offer coming in about 400 billion won (roughly $300 million) above FuriosaAI’s estimated market value, it was rebuffed.

A difference of opinion

According to Maeil Business Newspaper, negotiations broke down due to conflicting visions about the company’s future. FuriosaAI’s leadership chose to walk away from the deal and continue operating independently in the increasingly competitive AI semiconductor space.

“Since October of last year, Meta had been looking at several AI semiconductor companies in the U.S. and Israel, and finally chose FuriosaAI as a strong acquisition target and entered into negotiations at the beginning of the year,” an insider familiar with the company told Maeil Business Newspaper.

“I understand that the negotiations broke down because the two sides could not narrow their differences over the direction of the business and organizational structure after the acquisition, rather than the price.”

Founded in 2017 by CEO June Paik, FuriosaAI specializes in AI chip design and currently employs around 140 people - over 90 percent of whom are developers, including engineers from Google, Qualcomm, and Samsung.

The company’s flagship product, RNGD (pronounced 'Renegade'), unveiled at Hot Chips 2024, is a high-performance AI inference chip built on TSMC’s 5nm process and equipped with dual HBM3 memory.

According to FuriosaAI, it delivers twice the efficiency of traditional GPUs while consuming only a quarter of the power, so it’s easy to see the appeal to Meta.

Mass production of RNGD is expected to begin in the second half of 2025 and Maeil Business Newspaper reports that a number of major organizations, including LG AI Research and Saudi Aramco, have signed up to test the chip’s performance.

FuriosaAI has raised around 70 billion won (approximately $52 million) in funding to support its chip production and operational costs.

You might also like

• Meta is taking a RISC-y route for AI training and inference acceleration
• Meta reveals hardware behind recommendations for Facebook and Instagram
• Meta purportedly trained its AI on more than 80TB of pirated content