FuriosaAI is seeking to win inference customers away from Nvidia with RNGD, its second-generation AI processor. The South Korean startup says the chip consumes less power than comparable GPUs while operating in conventional air-cooled data centres — a combination it believes will appeal to operators constrained by both energy budgets and existing facilities.

The company, founded in 2017 by June Paik, a former Samsung Electronics and AMD engineer, began shipping RNGD in January 2026 and opened a Singapore subsidiary last month. It plans to more than double production next year. FuriosaAI has raised total funding of US$246 million, with backers including Korea Development Bank, Industrial Bank of Korea and Kakao Investment, and is considering a public listing as early as 2027.

Rather than chasing the far costlier market for training chips, FuriosaAI is targeting inference — the stage at which trained models generate outputs such as text, images and recommendations. As generative AI applications and agent-based systems proliferate, inference volumes have grown alongside scrutiny of power consumption, operating costs and infrastructure requirements. The company's bet is that processors designed specifically for inference can compete on efficiency and cost.

Early signs are encouraging. LG AI Research, an early customer, has run RNGD with its EXAONE models and reported 2.25 times the inference performance per watt of GPU-based systems. FuriosaAI is also set to supply Samsung SDS for cloud deployments from July, while Velox DCI Europe and I/ONX HPC plan to deploy more than 7,000 RNGD accelerators at a Stockholm data centre.

Background and analysis. Nvidia's dominance rests as much on its CUDA software ecosystem as on its silicon, which is why every inference challenger faces a migration hurdle as well as a hardware comparison. The energy-efficiency pitch is well timed: data centre operators are increasingly boxed in by power availability, and a chip that runs in conventional air-cooled facilities lowers the barrier to adoption without waiting on liquid-cooling retrofits. Southeast Asia, meanwhile, has become contested ground for AI infrastructure as cloud providers and enterprises build out regional capacity — Singapore is the obvious beachhead, which explains the new subsidiary.

Our take. A narrow focus on inference is a sensible wedge: compete on efficiency and cost where Nvidia is strongest on versatility. The Stockholm deployment will be a useful test of whether the performance-per-watt claims hold at scale. For Southeast Asian operators, a credible second source of inference silicon is welcome — competition, not just capacity, is what will keep regional AI infrastructure costs in check.

Source: iTnews Asia, "FuriosaAI pitches energy-efficient inference chip as an alternative to Nvidia", 15 October 2026. https://www.itnews.asia/news/furiosaai-pitches-energy-efficient-inference-chip-as-an-alternative-to-nvidia-629673