FuriosaAI

퓨리오사AI / 퓨리오사에이아이

A South Korean fabless semiconductor startup designing data-center inference accelerators for large language models and multimodal AI workloads.

Private company
Profile as of September 13, 2026
Not ranked: no published revenue or market capitalisation
How the ranking works →

FuriosaAI is one of the more technically credible South Korean AI-chip startups, having moved a large HBM-equipped inference chip into production. However, its revenue scale, margins, and long-term software adoption remain undisclosed, leaving it dependent on external manufacturing and a small number of early deployment partners.

Key figures

Revenue: Not disclosed
Revenue growth: Not disclosed
Operating margin: Not disclosed
Net income: Not disclosed
R&D spend: Not disclosed
Capital expenditure: Not disclosed
Private valuation Jul 2025: US$735 million
Total funding Jul 2025: US$246 million
RNGD memory bandwidth: 1.5 TB/s
Initial manufacturing batch Jan 2026: 4,000 units

Overview

FuriosaAI is a privately held South Korean fabless AI-chip designer founded in 2017. Its principal product is the RNGD data-center inference accelerator, designed for large language models, multimodal models, and agentic AI workloads. The company supplies RNGD as a PCIe accelerator card and as the NXT RNGD Server, a 4U system containing eight cards.

The chip uses FuriosaAI's Tensor Contraction Processor architecture and runs through the company's compiler, SDK, and LLM runtime. RNGD is manufactured on TSMC's 5-nanometer process and incorporates two HBM3 memory modules. This makes FuriosaAI dependent on external foundry capacity, HBM supply, advanced package assembly, and server integration. Its commercial ecosystem is currently narrow, with early deployments centered on South Korean partners such as LG AI Research, Samsung SDS, and LG Uplus.

The AI angle

FuriosaAI generates revenue directly through sales of RNGD accelerator cards, bundled servers, and associated software support for data-center inference. It is a direct supplier of AI inference compute, positioned to address power density and air-cooling constraints in data centers.

RNGD delivers up to 512 teraflops of FP8 performance and includes 48 gigabytes of HBM3 memory with 1.5 terabytes per second of bandwidth. The NXT RNGD Server incorporates eight cards in a 4U rack system, which the company states can produce up to 20 petaflops of INT8 inference performance per rack at approximately 3 kilowatts of total system power. LG AI Research adopts RNGD for EXAONE model deployments, and FuriosaAI works with partners like Samsung SDS and Hancom to develop enterprise and public-sector AI appliances. The scale of commercially recognized revenue from these AI deployments is not disclosed, but the company's entire product portfolio is dedicated to AI acceleration.

Technology and moat

FuriosaAI's central technical differentiation is its Tensor Contraction Processor architecture. Rather than organizing the accelerator around a conventional GPU-style matrix multiplication pipeline, it uses tensor-contraction execution and compiler co-design to fuse model operations more broadly. This aims to reduce data movement and improve inference efficiency for transformer and multimodal workloads.

The hardware design combines 48 gigabytes of HBM3, PCIe Gen5 x16 connectivity, multi-instance support, and SR-IOV virtualization, making it suitable for partitioned data-center inference. While the technology requires significant effort to replicate—including chip architecture, HBM integration, and TSMC qualification—it is not protected by a disclosed dominant market share, software standard, or exclusive foundry allocation. Nvidia remains the principal competitive benchmark due to its CUDA ecosystem lock-in and broad cloud availability. FuriosaAI's advantage is positioned around lower power consumption and inference density rather than equivalent general-purpose training breadth.

Five-pillar assessment

Scale and market position: A private startup with an undisclosed revenue scale, currently moving its first major data-center inference accelerator into initial mass production.
Technology and R&D: Differentiated by its Tensor Contraction Processor architecture and a 5-nanometer TSMC design featuring 48 gigabytes of HBM3 memory for efficient data-center inference.
Supply-chain centrality: Dependent on TSMC, Global Unichip, and Asus for manufacturing, with early commercial deployments concentrated among South Korean partners like LG AI Research and Samsung SDS.
Financial momentum: Secured US$125 million in 2025 bridge financing and transitioned to mass production in early 2026, though recurring commercial revenue remains unproven.
Governance and quality: Independent founder-led startup with limited public disclosure regarding financial results, board composition, and customer concentration.

Supply chain and relationships

FuriosaAI operates as a fabless designer and depends entirely on external manufacturing and assembly. TSMC fabricates the RNGD chip on its 5-nanometer process, while Global Unichip supports SoC and engineering work. Asus produces the RNGD cards for the reported initial production batch. The company also relies on external suppliers for HBM3 memory, advanced packaging, and substrates.

On the customer side, publicly named deployments center on a small number of South Korean partners. LG AI Research adopts RNGD for EXAONE model deployments, while Samsung SDS receives the chip for commercial services. LG Uplus and Hancom are partnering with FuriosaAI to develop AI appliances. Customer revenue shares and concentration are not disclosed, but the current ecosystem relies heavily on domestic enterprise infrastructure providers.

Customers

LG AI ResearchAdopts RNGD for EXAONE model deployments
Samsung SDSReceives RNGD for commercial services

Suppliers

TSMCFabricates RNGD on its 5-nanometer process

Partners

Global UnichipSupports SoC and engineering work for RNGD
ASUSProduces RNGD cards for initial production batch
LG UplusDevelops a sovereign AI appliance with FuriosaAI
HancomSupplies RNGD for enterprise AI appliance
Meta PlatformsHistoricalReportedly proposed an acquisition in 2025, which FuriosaAI declined

Competitors

NVIDIAInferredDominant alternative for data-center AI inference
AMDInferredAlternative data-center AI compute products
RebellionsInferredSouth Korean AI-chip competitor
DEEPXInferredSouth Korean AI-chip competitor
MobilintInferredSouth Korean AI-chip competitor

Geopolitics and risk

FuriosaAI's primary geopolitical exposure stems from its reliance on TSMC for 5-nanometer manufacturing. Any disruption to Taiwanese semiconductor manufacturing, advanced packaging, or shipping routes would directly affect its flagship chip.

The company is also exposed to United States and allied-country export controls. Advanced-node fabrication, HBM, AI accelerators, and AI-server deployments sit in technology categories increasingly affected by these regulations, even though FuriosaAI itself is not publicly identified as sanctioned. A reported future transition to 2-nanometer technology and HBM4 would increase its dependence on scarce next-generation process and memory capacity, heightening its exposure to global supply-chain bottlenecks.

Risk matrix
Risk	Severity	Why it matters
TSMC fabrication dependence	High	RNGD depends on TSMC's 5-nanometer manufacturing capacity and advanced packaging.
HBM supply dependence	High	Constrained HBM supply can limit product availability and increase system cost.
Nvidia software dominance	High	Customers may prefer Nvidia's established CUDA ecosystem and broad cloud support.
Customer concentration	High	Publicly named deployments center on a small number of Korean partners.
Commercial-scale execution	High	Ability to convert technical benchmarks into repeat orders remains unproven publicly.
Taiwan Strait disruption	High	Disruption to Taiwanese manufacturing would directly affect its flagship chip.
Export-control exposure	Medium	AI accelerators and HBM are increasingly affected by U.S. and allied export controls.
Private-company disclosure limits	Medium	Limited visibility into margins, burn rate, backlog, and capitalization.

Governance and ownership

FuriosaAI is an independent private startup, not state-owned or part of a disclosed chaebol. It is led by founder and CEO June Paik. The company reportedly rejected an US$800 million acquisition proposal from Meta Platforms in March 2025 to remain independent.

Its July 2025 Series C bridge round included Korea Development Bank and Industrial Bank of Korea, but no public disclosure establishes state control. As a private company, FuriosaAI provides limited public financial, customer, ownership, production, and governance disclosure, reducing visibility into its margins, burn rate, and board composition.

What to watch

Whether FuriosaAI reports repeat commercial orders beyond its named South Korean partners.
Whether the reported 4,000-unit initial RNGD production batch converts into disclosed shipments and customer deployments.
Whether the company publishes revenue, gross margin, or cash consumption data after scaling production.
Whether the LG AI Research partnership results in identifiable enterprise deployments.
Whether the Singapore subsidiary produces named Asia-Pacific customers beyond South Korea.
Whether FuriosaAI confirms the architecture and schedule of its reported next-generation 2-nanometer accelerator.

Recent developments

July 23, 2026Announced a Korean AI-inference deployment using RNGD with Samsung SDS.
March 4, 2026Signed a memorandum of understanding with LG Uplus for a sovereign AI appliance.
January 28, 2026Press reporting stated that RNGD entered mass production with an initial batch of 4,000 units.
July 30, 2025Announced a US$125 million Series C bridge financing round at a US$735 million private valuation.
July 15, 2025Announced a partnership with LG AI Research to supply RNGD for EXAONE-based enterprise AI deployments.
March 24, 2025Reportedly rejected Meta Platforms' approximately US$800 million acquisition proposal.

Coverage on AsiaAI.FYI

We haven't written about FuriosaAI yet. Subscribe below to get it when we do.

Guides that cover FuriosaAI

Korea AI Industry Map

About this profile

Compiled with AI-assisted research from company filings, market data, and published reporting as of September 13, 2026, then reviewed by AsiaAI.FYI. Figures marked Estimate are not company-reported. Check primary filings before relying on any number.

Confidence: C. Product specifications, financing, and named partnerships have stronger support than revenue, profitability, ownership, and customer concentration.

Main sources: Company product documentation; Company funding and partnership announcements; Semiconductor-partner announcements; Credible technology and financial press.

All 55 sources

Position in the AI stack

Designs the RNGD data-center inference accelerator and packages it into deployable server hardware.

Legal name: FuriosaAI, Inc.
Headquarters: Seoul, Republic of Korea
Founded: 2017
Founders: June Paik / 백준호
Leadership: June Paik / 백준호, Founder and Chief Executive Officer
Employees: More than 240
Ownership: Private company
Business: Fabless AI-chip designer
Sector: Chip Design
Servers, PCs & ODMs
Website: furiosa.ai