FuriosaAI unveils AI chip to challenge Nvidia in inference
FuriosaAI unveiled its second-generation artificial intelligence chip, Renegade, or RNGD, targeting a fast shift in the industry from model training to cost-heavy inference, Thursday.
The new chip, which kicked off mass production in January, is designed to handle high volumes of real-time AI queries at lower power and cost.
FuriosaAI CEO Paik June-ho at the Renegade 2026 Summit in Seoul, said:
By 2030, we expect roughly 100 gigawatts of AI data center capacity to be added globally, with about 70 percent allocated to inference,
“The core of future data center design will be how efficiently operators can execute repetitive inference workloads at the lowest total cost.”
FuriosaAI said benchmarking with overseas clients showed RNGD can handle up to 7.4 times more concurrent users than Nvidia’s RTX Pro 6000 at the same power level. The chip’s thermal design power is about 180 watts, enabling roughly a 40 percent reduction in data center total cost of ownership, according to the company.
Paik said,
The AI ecosystem can’t scale if computing costs keep rising faster than revenues,
“Our mission is to make AI computing sustainable and accessible.”
FuriosaAI began silicon development in 2022, with early samples emerging two years later. The company said it prioritized power efficiency as a core design constraint from day one.
Paik said,
Rather than maximizing peak performance, we set a strict power envelope — under 200 watts — and optimized how to run diverse and rapidly evolving AI models efficiently within that limit,
The company has started mass production of about 4,000 RNGD units this year — a relatively rare case of a high-performance NPU with high-bandwidth memory reaching commercial scale.
The flagship RNGD chip packs about 40 billion transistors and is manufactured on a 5-nanometer process by TSMC. It is paired with fourth-generation high-bandwidth memory, known as HBM3, from SK hynix.
Each chip delivers up to 512 teraflops of compute performance and roughly 1.5 terabytes per second of memory bandwidth, the company said.
Beyond hardware, FuriosaAI outlined plans to expand its software stack, including compilers and SDKs tailored for scalable inference architectures. Partners including LG AI Research, LG Uplus, Samsung SDS and MegazoneCloud presented deployment use cases.
In a notable step, Samsung SDS plans to roll out a subscription-based neural processing unit service in July using RNGD — the first such offering by a domestic cloud provider.
The service will allow customers to access the chips through the Samsung Cloud Platform in flexible configurations, integrating them with storage, networking and compute services. The move also diversifies Samsung SDS’ existing graphics processing units-based offerings and supports South Korea’s push for “sovereign AI” infrastructure.
Vice Minister Ryu Je-myung said at the event,
FuriosaAI is one of the strongest contenders capable of surviving in the global AI battlefield,
“RNGD could secure an irreplaceable position in the global supply chain in the era of physical AI.”
READ the latest news shaping the Nvidia market at Newsvidia
FuriosaAI unveils AI chip to challenge Nvidia in inference, source
