FuriosaAI
Category: AI Chips / Semiconductors
A high-performance AI semiconductor company designing specialized NPU (Neural Processing Unit) accelerators optimized for data center inference and Generative AI workloads. FuriosaAI was founded in 2017. The company is led by Juneho Baek. Based in Seoul, South Korea. Team size: 200+. Total funding raised: $246.0M. Latest round: Series C Bridge ($125.0M, Jul 2025). Key investors include Naver D2SF, Korea Development Bank (KDB), DSC Investment, Keistone Partners, Kakao Investment, Industrial Bank of Korea, PI Partners, National Growth Fund (국민성장펀드), Export-Import Bank of Korea.
AMW Analysis
FuriosaAI is a South Korean fabless semiconductor company that designs NPU accelerators for data center inference and generative AI workloads, emphasizing performance, programmability, and power efficiency to lower TCO versus GPUs. Recent news flow shows a rapid escalation in scale and strategic positioning. The company rejected an $800 million acquisition offer from Meta in early 2026, signaling a commitment to independence. It subsequently secured approximately $830 million from a South Korean government advanced industry fund and is nearing completion of an $800 billion won Series D round at a $4 trillion won valuation. FuriosaAI has also entered a partnership with Broadcom to co-develop a 2nm AI inference chip targeting agentic workloads, with sampling planned for 2028. This partnership, combined with the ramp-up of its RNGD chip production, indicates a shift from early-stage development toward mass production and long-term ecosystem building, supported by substantial government and venture capital backing.
AMW analysis, generated from 7 tracked news signals.
- Founded
- 2017
- Headquarters
- Seoul, South Korea
- Team size
- 200+
- Total funding
- $246.0M
Value proposition
Delivers a 'trifecta' of high performance, programmability, and superior power efficiency (Perf/Watt) for LLMs compared to traditional GPUs, significantly lowering the Total Cost of Ownership (TCO) for AI inference.
Products and solutions
RNGD (Renegade): 2nd Gen AI Accelerator, Warboy: 1st Gen NPU, Furiosa SDK, NXT RNGD Server: 4U rackmount server housing 8 RNGD cards
Unique value
Pioneered the Tensor Contraction Processor (TCP) architecture, a non-GPU hardware design that treats tensor contraction as a primitive operation.
Target customer
Hyperscale Data Centers, Cloud Service Providers (CSPs), Enterprise AI Adopters, and Edge AI Infrastructure Providers
Industries served
Cloud Computing & Data Centers, Generative AI (LLMs), Computer Vision / Smart City, Enterprise Software, Scientific Computing
Technology advantage
The TCP architecture eliminates the inefficiencies of adapting matrix multiplication on GPUs, allowing FuriosaAI to achieve higher performance-per-watt and deterministic latency for massive models.
How they differentiate
Proprietary 'Tensor Contraction Processor' (TCP) architecture that abandons the GPU legacy to optimize for matrix operations, delivering 2.25x better performance-per-watt for LLM inference compared to traditional GPUs.
Main competitors
Rebellions (merged with Sapeon), Groq, Nvidia
Key partnerships
LG AI Research, Kakao Enterprise, ASUS, Hugging Face, Gcore, TSMC, OpenAI, SK hynix, Broadcom, Equinix, Samsung SDS
Notable customers
LG AI Research, Kakao Enterprise, Samsung SDS
Major milestones
Launched 2nd Gen AI Chip 'RNGD' (Renegade) in 2024, Secured $125M Series C funding in July 2025, Reportedly rejected an $800M acquisition offer from Meta, RNGD enters mass production: 4,000 units shipped by TSMC (Jan 2026), Established European flagship office in Portugal (Apr 2026), Broadcom partnership for 3rd-gen agentic inference platform (May 2026), Pre-IPO / National Growth Fund tranche (~₩4000억) begins closing (May 2026), Equinix Lisbon RNGD deployment (Jul 2026), Samsung SDS launches Korea's first RNGD NPUaaS (Jul 2026), Singapore APAC hub established (Sep 2026)
Growth metrics
Unicorn (~$735–770M) after Jul 2025 Series C bridge; Pre-IPO targeting ~₩3T (~$2.3B) pre-money valuation with 2027 IPO plans; RNGD in mass production with Samsung SDS NPUaaS and Equinix Lisbon deployments; ~214 employees (Jun 2026).
Market positioning
Leading non-GPU AI accelerator challenger in the Asia-Pacific region, targeting the global data center inference market.
Geographic focus
South Korea (HQ), United States (Bay Area), Portugal (Lisbon / European HQ), Singapore (APAC hub), Malaysia (Johor Bahru), Global Enterprise Market
Patents and IP
Holds core patents on Tensor Contraction Processor (TCP) architecture, NPU memory management, and compiler-hardware co-design.
About Juneho Baek
Semiconductor engineer with expertise in memory and GPU architecture. Previously served as a Senior Engineer at Samsung Electronics (Memory Division, 2013–2016) and Staff Engineer at AMD (GPU Infrastructure, 2010–2013). MS from Georgia Tech.
Latest news about FuriosaAI
- FuriosaAI nears completion of 800 billion won Series D, valuation at 4 trillion won
- FuriosaAI, Nuvacore, D-Matrix Join Chip Funding Spree Amid AI Inference Demand
- FuriosaAI receives ~$830M from South Korea government advanced industry fund
- FuriosaAI and Broadcom co-develop 2nm AI accelerator targeting 2028 sampling
- FuriosaAI partners with Broadcom to build 2nm AI inference chip for agentic workloads
More AI Chips / Semiconductors companies
Official website: https://furiosa.ai/