South Korean startup FuriosaAI, which last year introduced the RNGD AI accelerator for working with large language models (LLM) and multimodal AI, has rejected a takeover offer from Meta✴ Platforms, which was ready to pay $800 million for it. FuriosaAI intends to continue developing its business as an independent company, Bloomberg reported, citing an informed source.
Meta✴ had been in talks to acquire Seoul-based FuriosaAI since early this year, according to a Bloomberg source. Shares of its largest investor, South Korean venture capital firm DSC Investment, began to rise after rumors of a possible takeover emerged in February, but fell 16% on Monday after news of the deal’s rejection.
The RNGD accelerator, built on TSMC’s 5nm process and using HBM3 memory chips from SK Hynix, is capable of competing with chips from NVIDIA and other AI startups including Groq, SambaNova, and Cerebras. FuriosaAI has about 150 employees, including 15 people based in its Silicon Valley office. Late last year, the startup began testing the RNGD and sending out sample chips to customers including LG AI Research (LG Group’s AI arm) and Saudi Aramco. More than a dozen potential customers are expected to test the RNGD samples in the first half of this year.
Image source: FuriosaAI
FuriosaAI plans to raise additional capital before pursuing an initial public offering (IPO), the people said. The company is expected to close an expanded Series C round in about a month, which will raise the target amount. FuriosaAI, founded in 2017, has raised $115 million in four rounds of funding, Data Center Dynamics reported.
Meta✴, meanwhile, is investing heavily in AI infrastructure to stay ahead of competitors including OpenAI and Google. In January 2025, Meta✴ CEO Mark Zuckerberg said the company would spend up to $65 billion on AI in 2025, including building a massive data center and hiring AI talent. Just a week later, Zuckerberg told investors that Meta✴ planned to spend hundreds of billions of dollars on AI infrastructure.
Meta✴ is also working on its own chips. In 2023, it introduced the Meta✴ Training and Inference Accelerator (MTIA) chip for AI workloads, and in 2024, the second-generation MTIA AI accelerator was announced, which is three times more powerful than its predecessor. The company is currently testing its own accelerator for training AI models.