As the AI inference market grows, a South Korean inference chipmaker startup has raised $400 million as it prepares to go public and compete with Nvidia.
Semiconductor vendor Rebellions said on March 30 that its new fundraising round values it at $2.34 billion. Mirae Asset Financial Group and the Korea National Growth Fund led the funding round. Korea National Growth Fund is the South Korean government’s investment entity. The vendor aims to use the fund to expand into the U.S. market and snag customers such as Meta and xAI.
Other financial backers not part of the latest fundraising round include Samsung, SK Hynix and Arm.
Founded in 2020, Rebellions has grown over the past six years for a range of reasons. The vendor is a key part of South Korea’s push for a key chipmaker rivaling Nvidia. However, with the AI inference market expanding, Nvidia appears to be paying attention. In December, the AI hardware giant paid $20 billion in cash to acquire intellectual property and inference technology from Groq, a startup specializing in high-speed AI inference chips. While Groq is a separate entity with a new CEO, the deal gives Nvidia the ability to integrate Groq’s language processing unit technology into its AI Factory architecture.
The focus on inference not only provides new opportunities for established vendors, such as Nvidia, but also opens the door for startups such as Rebellions and other vendors, notably AMD and Cerebras.
“It’s a different profile of chips that’s needed, and this is where other non-GPU architectures will become more important,” said Nick Patience, an analyst at Futurum Group. “It’s not about massive parallel processing. It’s about continuous iteration and continuous processing.”
The Enterprise Advantage
For enterprises, the opportunities for more vendors to enter the AI inference mean silicon diversity, with a resurgence of CPUs and other Processors, Patience added.
While it is too early to say whether Rebellions will be successful, enterprises that are paying attention to the startup should focus on the software stack that sits on top of Rebellions’ hardware, Patience said.
Many enterprise developers are locked into the Nvidia CUDA stack, which has been the go-to for decades. However, Rebellions appears more open source-focused because it is a member of the PyTorch Foundation and has used other open source engines such as vLLM, an open source engine for large language models, he said.
“Rebellions is not locking people in,” Patience said. “That can be seen as an advantage for those who want to buy systems that use these [open] processes.”
Another benefit for Rebellions is that Samsung and SK Hynix, vendors focused on manufacturing high-bandwidth memory (HBM) chips, are key investors. With the scarcity of HBM, they could help Rebellions build these AI inference chips.
Despite all its advantages, Rebellions still faces a significant challenge: it has an extraordinarily strong competitor in not only Nvidia but also AMD and others.
