Getty Images
FuriosaAI to fuel LG Exaone LLM: Is it a challenge to Nvidia?
The South Korean semiconductor startup scored LG as its first major customer, as companies compete to take advantage of the AI boom.
Seoul-based chip startup FuriosaAI is looking to threaten Nvidia's data center AI chip dominance and has scored its first major contract -- powering LG's Exaone large language model.
LG AI Research will offer FuriosaAI's RNGD chip-powered servers to enterprise customers across electronics, finance, telecommunications and biotechnology for LLMs.
FuriosaAI's RNGD chip, a neural processing unit, is designed to support LLMs and other deep learning models on the inference side, rather than training. Inference uses trained models to make predictions.
RNGD is based on tensor contraction processor architecture and is manufactured using Taiwan Semiconductor Manufacturing Co.'s 5 nm process node. The company claims that RNGD achieves 2.25 times better per-watt performance for inference over GPUs, which have become the most popular processors for AI workloads.
"I think FuriosaAI is smart to go after the inference market as its primary target ... It is far more open and lucrative than the training market," said Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, in an interview.
Nvidia's hold on data center
Toppling GPU juggernaut Nvidia is easier said than done. Nvidia enjoyed a 98% data center GPU market share in 2023, according to Silicon Analysts. Competitors claiming performance wins over Nvidia's hardware still have a mountain to climb -- Nvidia's popular CUDA software helps the company maintain its comfortable grip on AI workloads. GPU challenges from AMD and Intel have done little to loosen that grip.
"Training is dominated by Nvidia GPUs and CUDA, with AMD showing potential to eat some of that share on the enterprise side," Kimball said. "Does [FuriosaAI's RNGD] pose a threat to Nvidia? I don't think so. ... RNGD seems to have scored in its niche area."
Nvidia leads in the inference market with its A100, H100, L4, L40 and Blackwell GPUs. But the inference side is more diverse, with competition from Google using custom-built TPU v4i chips, Amazon with in-house Inferentia chips, Intel CPUs, AMD, Qualcomm and several startup chipmakers.
Finding a niche in AI inference
Inference is gaining momentum and is expected to be a larger market than training in the coming years. The global AI inference market was estimated at $106 billion in 2025 and projected to grow to $255 billion by 2030, according to a MarketsandMarkets report.
Kimball said FuriosaAI's technology is aimed at a very specific low-latency inference, and that Nvidia's GPUs will still be valuable on the inference side, especially for extremely large models.
"RNGD has an architecture that lends itself to high-query-per-second and lower-precision kind of inference. ... Think recommendation engines, customer service bots, programmatic advertising, etc.," Kimball said.
Other big tech players clearly see potential in the growing inference market. In March, FuriosaAI turned down Meta's $800 million bid to acquire the company.
"I believe the inference market is wide open. Partly because of its size, and partly because of the diversity of use cases and deployment scenarios," Kimball said.
Shane Snider, a veteran journalist with more than 20 years of experience, covers IT infrastructure at Informa TechTarget.
