Name

FuriosaAI

Legal name

주식회사 퓨리오사에이아이

Website

https://furiosa.ai

Company type

Private

Founded year

2017

Operating status

Operating

Headcount range

101–250 employees

Ownership category

akta.pro rank

Product category

AI Inference Accelerators

NAICS

Computer and Peripheral Equipment Manufacturing (3341), Computer Systems Design and Related Services (54151)

akta.pro primary industry

AI Accelerators (GPUs/TPUs/NPUs/ASICs) (HDAAAAAA)

akta.pro secondary industries

AI Server Systems & HGX/Accelerator Platforms (HDAAAAAB), AI/Neural Processing Units (NPUs) & ML Accelerators (HDAHAJAF), AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers) (HDAAAAAI)

Keywords

• AI inference accelerators
• Data center AI chips
• Neural processing units
• AI semiconductors
• LLM inference hardware

Headquarters

HQ city

Seoul

HQ country

South Korea

HQ region

Asia

Offices6 records

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Supply Chain, Personnel, Operations, Marketing or Sales, Infrastructure

Revenue model

1. AI Accelerator Hardware Sales: Sale of RNGD PCIe accelerator cards and NXT RNGD Server systems to enterprise customers, cloud providers, and data center operators. Products sold directly and through channel partners. Hardware constitutes the primary revenue stream with first large-scale commercial shipments beginning Q1 2026.
2. NPU-as-a-Service (NPUaaS): Samsung SDS offers Furiosa RNGD as a subscription NPU service on Samsung Cloud Platform starting July 2026. Provides enterprise customers with on-demand access to RNGD compute without purchasing hardware outright.
3. Software Licenses & Developer Tools: Furiosa SDK, Furiosa Docs, Furiosa Compiler, and runtime software provided as part of the hardware solution. Software stack includes Tensor Contraction Language, vLLM-compatible serving, and OpenAI API compatibility.

Pricing tiers

Go-to-market motion5 records

Distribution channels8 records

Marketing channels8 records

Core offering

FuriosaAI designs and manufactures data center AI inference accelerators built on its proprietary Tensor Contraction Processor (TCP) architecture. Its flagship RNGD (Renegade) PCIe card delivers 512 TFLOPS FP8 at 180W TDP, and is bundled with the NXT RNGD turnkey server and the Furiosa SDK for LLM and multimodal inference. The company also sells the Warboy Gen 1 Vision NPU for computer vision workloads and co-developed the Sovereign AI Appliance with LG U+ for on-premise enterprise AI.

Product overview

FuriosaAI offers a portfolio of AI accelerator products built on its proprietary Tensor Contraction Processor (TCP) architecture. The product lineup includes the RNGD (Renegade) PCIe card as the core second-generation inference accelerator, the NXT RNGD Server as a turnkey multi-card deployment solution, and the Warboy Gen 1 Vision NPU for computer vision workloads. These hardware products are supported by the Furiosa SDK software stack providing PyTorch integration, OpenAI API compatibility, and Kubernetes deployment support. The company also offers the Furiosa Access Program for evaluation and the Sovereign AI Appliance (co-developed with LG U+) as a ready-to-run enterprise solution.

Differentiator

Problem solved

Functional benefit

Products and services

• RNGD (Renegade) PCIe Card Second-generation AI inference accelerator delivering 512 TFLOPS FP8 / 256 TFLOPS BF16 performance with 48GB HBM3 memory and 256MB on-chip SRAM at 180W TDP. Designed for air-cooled data centers and targeting LLM and multimodal AI inference workloads. Supports BF16, FP8, INT8, and INT4 precision.
• NXT RNGD Server Turnkey 4U rackmount server housing 8x RNGD cards delivering 4,096 TFLOPS FP8 (512 TFLOPS per card). Features 384GB HBM3 capacity, 12 TB/s memory bandwidth, and 3kW power consumption. Designed for air-cooled data centers with up to 5 servers per 15kW rack.
• Warboy (Gen 1 Vision NPU) First-generation Vision NPU for computer vision deployment in data centers and edge servers. Delivers 64 TOPS INT8 with 16GB LPDDR4X at 40-60W TDP. Achieved MLPerf benchmark scores outperforming Nvidia on image classification and object detection tasks.
• Furiosa SDK Comprehensive software toolchain for LLM inference and agentic workloads including compiler, quantization, and runtime. Features PyTorch 2.x integration, OpenAI API compatibility, vLLM drop-in replacement, and Kubernetes support.
• Sovereign AI Appliance Ready-to-run on-premise AI solution co-developed with LG U+ combining Furiosa RNGD accelerator with LG's EXAONE 4.0 model and ixi-Enterprise platform. Delivers 30% TCO reduction versus GPU clusters with no external cloud requirements.

Quantifiable outcome

• 2.25x better LLM inference performance per watt vs comparable GPUs — validated by LG AI Research using EXAONE model
• +8 more outcomes

Named customers15 records

Segments5 records

Ideal customer profiles4 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration4 records

AI capability8 records

Feature6 records

Partnerships

24 partnerships are on record, tiered core, minor and secondary.

• BroadcomcoreStrategic or Co-development Partner · 27 May 2026Strategic partnership to co-develop third-generation AI accelerator based on Furiosa's Tensor Contraction Processor (TCP) architecture into a multi-die chiplet system. Broadcom contributes XPU Technology, Ethernet scale-up networking, and fabric switches for rack-scale AI inference infrastructure. Platform targets 2nm compute die with HBM4/4E memory. Sampling expected in 1H 2028. Contract value estimated at KRW 200 billion range.
• CMC Corporation (Vietnam)minorStrategic or Co-development Partner · 24 April 2026Vietnam's CMC Corporation signed MOU with FuriosaAI to jointly develop AI solutions, expanding business opportunities in Vietnam and South Korea. CMC Global and FuriosaAI collaborate on AI application software leveraging their respective technological strengths.
• Samsung SDScoreChannel Partner/ Reseller/ Distributor · 2 April 2026Samsung SDS announced launch of new RNGD-powered NPU-as-a-Service (NPUaaS) on Samsung Cloud Platform starting July 2026 — the first NPUaaS by a domestic Korean cloud provider. Provides enterprise customers with on-demand access to RNGD compute fully integrated with Samsung Cloud Platform's virtualization, storage, and networking layers. Samsung SDS also serves as a validated enterprise customer for RNGD deployment.
• LG U+coreStrategic or Co-development Partner · 4 March 2026Joint development of Sovereign AI Appliance — the first 'ready-to-run' AI inference solution combining LG AI Research's EXAONE 4.0 language model, LG's ixi-Enterprise platform, and FuriosaAI's RNGD NPU in a single server package. Enables on-premise deployment for defense, healthcare, finance, manufacturing, and research sectors. Reference configuration uses 14 RNGD cards across 2 servers delivering 7,168 TOPS (FP8) with 32K context length support.
• HelikaisecondaryTechnology or Integration · 26 February 2026Partnership to deliver secure, production-ready enterprise AI automation stack. Helikai's full software stack certified on FuriosaAI NXT RNGD inference servers. Joint solution enables enterprises in regulated sectors to run high-performance, domain-specific AI workflows with complete data sovereignty. Targets bioinformatics, clinical trial matching, and legal discovery use cases.
• Nota AIsecondaryTechnology or Integration · 3 February 2026Nota AI signed AI model optimization technology supply contract with FuriosaAI. Using NetsPresso quantization platform to optimize AI models for RNGD NPU, reducing memory requirements for Llama 3.1 70B from ~140GB to 35-40GB via INT4 precision conversion. Collaboration targets security, medical, retail, and smart building sectors.
• TSMCcoreOEM/ Whitelabel/ Licensing PartnerTSMC manufactures RNGD AI accelerators at 5nm process technology. First delivery of 4,000 RNGD units shipped and assembled. TSMC producing 2nm third-generation chips in partnership with Broadcom. Longstanding partnership enabling FuriosaAI's silicon production capabilities.
• SK hynixcoreTechnology or IntegrationSK hynix supplies HBM3 memory for RNGD accelerators. HBM3E upgraded variants planned for H2 2026. SK hynix partnership enables RNGD's 48GB HBM3 memory capacity at 1.5TB/s bandwidth critical for LLM inference workloads.
• ASUScoreOEM/ Whitelabel/ Licensing PartnerASUS assembles RNGD accelerator cards. Q&A partnership discussing AI server trends and FuriosaAI collaboration in system integration.
• AnexiasecondaryChannel Partner/ Reseller/ DistributorEuropean cloud infrastructure provider integrating RNGD into global service offerings, expanding FuriosaAI's footprint in EMEA market.
• UpstagesecondaryTechnology or IntegrationKorean AI lab partnering to deploy Solar family of foundation models on RNGD for a wide range of AI applications.
• MegazoneCloudsecondaryChannel Partner/ Reseller/ DistributorKorean cloud managed service provider launching new cloud compute project powered by RNGD to meet surging regional demand.
• LG CNSsecondaryChannel Partner/ Reseller/ DistributorLG CNS partners with FuriosaAI bringing South Korea's NPU to enterprise AI services, implementing NPU-based AI solutions for industrial applications.
• MangoBoostsecondaryTechnology or IntegrationKorean silicon company partnered to develop world's first NPU-DPU scale-out architecture, leveraging 400G RDMA fabric to enable linear multi-node scaling for large-scale LLM inference.
• Baro AIsecondaryTechnology or IntegrationKorean AI infrastructure company showcasing AI server integrating RNGD with POSEIDON liquid-cooling architecture for maximum computational throughput and absolute stability.
• CreverseminorChannel Partner/ Reseller/ DistributorKorean education technology company deploying RNGD to power its AI-driven HUMMINGo platform.
• flexgrid.cloudminorChannel Partner/ Reseller/ DistributorDelivers sovereign AI compute at the edge, standardized on RNGD to power air-cooled metro AI data centers built for inference in power-constrained urban environments worldwide.
• OpenAIcoreTechnology or IntegrationOpenAI showcased FuriosaAI RNGD at an event in Seoul. OpenAI demonstrated gpt-oss 120B model running live on RNGD accelerator — marking the first time OpenAI models ran on Korean silicon. Live demo validated RNGD's enterprise inference capabilities.
• Korea Southern Power (KOSPO)secondaryTechnology or IntegrationKOSPO signed MOU with FuriosaAI to build energy-efficiency-focused domestic AI infrastructure ecosystem using domestic NPUs. Partnership cuts KOSPO's operating costs for power generation facilities while enabling RNGD validation in real industrial settings.
• POSCO DXminorChannel Partner/ Reseller/ DistributorPOSCO DX signed agreement with Mobilint for NPU-based AI solutions for industrial applications. Note: FuriosaAI also referenced as alternative domestic NPU option for Korean industrial AI transformation.
• AramcominorChannel Partner/ Reseller/ DistributorAramco listed as partner customer. Saudi Aramco's digital development initiatives incorporating FuriosaAI technology for AI infrastructure.
• GUC (Global Unichip Corp)secondaryTechnology or IntegrationGUC partnered on RNGD as the Most Efficient AI Accelerator for LLMs. GUC provides advanced ASIC design services supporting RNGD development.
• hosted.aisecondaryChannel Partner/ Reseller/ Distributorhosted.ai partnered to deliver industry-leading AI infrastructure powered by Tensor Contraction Processors, bringing FuriosaAI RNGD to Microsoft's Azure Marketplace.
• LG AI ResearchcoreTechnology or IntegrationLG AI Research serves as flagship enterprise customer, validating RNGD in EXAONE model deployments. Confirmed 2.25x better LLM inference performance per watt vs GPUs. Jointly developed EXAONE 4.0 integration in Sovereign AI Appliance with LG U+.

Scale indicators8 records

Recent moves8 records

Expansion highlights7 records

Direct peers

• Rebellions: Korean AI chip startup (also part of the K-Nvidia cohort) developing data center AI accelerators with ATOM and upcoming REBEL chips. Most direct comparable to FuriosaAI: same national ecosystem, same customer overlap (LG, Samsung), and competing in the same MLPerf and sovereign-AI benchmarks.
• Groq: US-based AI inference accelerator company with LPU (Language Processing Unit) architecture. Direct comparable in target use case (LLM inference) and positioning (GPU alternative for cost-efficient inference at scale), although Groq uses a deterministic LPU systolic approach versus FuriosaAI's TCP tensor-contraction approach.
• Cerebras Systems: US-based AI accelerator company building wafer-scale chips for training and inference. Comparable in offering large-scale AI compute alternatives to NVIDIA GPUs, and competes in the same enterprise AI inference RFP pipeline that FuriosaAI is targeting.
• SambaNova Systems: US AI accelerator company offering RDU-based dataflow chips for enterprise AI. Comparable in selling AI inference systems to enterprises and governments as a CUDA-independent alternative, with overlapping customer profiles in regulated industries.
• Tenstorrent: AI accelerator startup building RISC-V-based Grayskull/Wormhole/Blackhole chips for training and inference. Comparable in competing for the same NVIDIA-displacement opportunity with custom silicon and a software-first developer story.

Broad incumbents

• NVIDIA: Dominant incumbent in AI accelerators (Hopper/Blackwell/Rubin) with the deepest CUDA software moat. FuriosaAI explicitly benchmarks against NVIDIA RTX Pro 6000 and H-series GPUs, and the entire value proposition is positioned as a TCO/performance alternative to NVIDIA.
• AMD: Major GPU and AI accelerator competitor with Instinct MI series (MI300X, MI325X, MI350) and ROCm software stack. Competes with FuriosaAI in the same AI data center RFPs and offers a CUDA-alternative software path with broader x86 integration.
• Intel (Habana Labs): Intel's AI accelerator subsidiary offering Gaudi 2 and Gaudi 3 training/inference chips as an NVIDIA alternative. Direct competition in the AI data center accelerator market with established enterprise sales motion, comparable to FuriosaAI's enterprise GTM.

Emerging players

• DEEPX: Korean AI chip startup focused on edge and on-device AI NPUs. Part of the same Korean AI chip cohort and listed alongside FuriosaAI in the K-Perf government benchmark program. Comparable in the Korean sovereign-AI ecosystem story, though focused on lower-power edge inference versus data center.
• Mobilint: Korean AI chip startup developing NPU silicon for edge and data center inference. Also part of the K-Nvidia cohort with overlapping investor base and customer profile (POSCO DX signed with Mobilint for industrial AI). Comparable as a fellow Korean sovereign-AI chip contender with mass production challenges.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

• YouTubeyoutube.com/channel/UCsaPzmRKx-1__hKQcTiiSlA
• LinkedInlinkedin.com/company/furiosaai
• Xx.com/FuriosaAI

Revenue estimate

Valuation estimate

Number of profiles

Profiles5 records

Funding overview

Funding rounds7 records

Investors27 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about FuriosaAI