South Korea's AI Semiconductor Fabless Companies — From NPU to LPU, the K-Chip Era Begins

With NVIDIA dominating the AI semiconductor market today, South Korea is mounting a quiet but determined challenge. From datacenter NPUs and edge AI chips to automotive AI accelerators and semiconductor IP — Korean fabless companies are racing toward the global stage across a wide range of sectors. Today we break down South Korea's major AI semiconductor fabless companies by category.

---

🏢 Datacenter AI Inference / High-Performance NPU

1. Rebellions

Founded: 2020

The flagship of South Korean AI semiconductors, with strategic investment from Samsung Electronics and ARM. Designing NPUs optimized for large language model (LLM) inference, the company merged with SAPEON Korea in December 2024 to become South Korea's first AI semiconductor unicorn.

Key Products:

• ATOM / ATOM Max: 1st-generation NPU. Deployed in SK Telecom and KT Cloud datacenters.
• REBEL / REBEL-Quad: 2nd-generation NPU. Chiplet architecture + HBM3e 144GB.

Foundry & Process:

• ATOM: Samsung Foundry 7nm
• REBEL / REBEL-Quad: TSMC 4nm (N4P). Chiplet design executed in partnership with SEMIFIVE.

Recent News: In March 2026, closed a pre-IPO round led by Mirae Asset and others, raising $400 million (~₩570 billion). Cumulative fundraising of $850 million, valuation approximately ₩2.3 trillion. IPO targeted for second half of 2026.

---

2. FuriosaAI

Founded: 2017 (CEO Joo-Ho Baek, former AMD and Samsung engineer)

A datacenter NPU specialist that positions "performance per watt" as its core differentiator. Develops its own software stack and compiler alongside hardware for a highly polished integrated solution.

Key Products:

• WARBOY: 1st-generation vision NPU. Adopted in Kakao and ASUS datacenter servers.
• RNGD (Renegade): 2nd-generation NPU. TSMC 5nm, HBM3 48GB, FP8 512TFLOPS, 180W. Demonstrated running OpenAI's 120B parameter model.

Foundry & Process:

• WARBOY: Samsung Foundry 14nm
• RNGD: TSMC 5nm (N5)

Recent News: Became a unicorn in 2025 (cumulative investment of $246 million). Secured major domestic AI companies including Upstage and LG AI Research as customers.

---

3. HyperAccel

Founded: 2023

The company that unveiled the world's first LPU (Latency Processing Unit) architecture for LLM inference. Achieves approximately 90% memory bandwidth utilization using low-cost LPDDR memory instead of HBM.

Key Products:

• Bertha: TSMC 4nm LPU. Co-designed by SEMIFIVE. Claims ~5x tokens/second vs. H100 GPU, and 19x price-performance advantage.

Foundry & Process:

• Bertha: TSMC 4nm (N4), leveraging SEMIFIVE's ASIC design platform.

Recent News: Raised ₩55 billion in Series A. Signed mass production contract with SEMIFIVE in Q1 2026.

---

📡 Edge AI / On-Device NPU

4. DeepX

Founded: 2018 (CEO Nok-won Kim, former designer of Apple A11 Bionic chip)

Founded by the engineer behind the Apple A11 Bionic chip. An edge AI NPU specialist with ultra-low power consumption (average under 3W) as its core strength, covering 7 physical AI domains including smart factories, robotics, autonomous driving, and surveillance cameras.

Key Products:

• DX-M1: Samsung 5nm, 200 eTOPS, 1–5W ultra-low power edge AI SoC.
• DX-H1 V-NPU: 1,600 eTOPS PCIe card.
• DX-M2: Next-gen chip targeting on-device 20B parameter LLM inference, 2nm process.

Foundry & Process:

• DX-M1: Samsung Foundry 5nm
• DX-H1: Samsung Foundry (specific node undisclosed)
• DX-M2: Targeting Samsung Foundry 2nm (SF2), currently in development

Recent News: 27 commercial contracts across 8 countries within 7 months of mass production launch. Integrated supply contract with Baidu PaddlePaddle platform (40,000 units). Partnerships with Hyundai Kia Robotics Lab, POSCO DX, and LGU+.

---

5. Mobilint

Founded: 2019

An NPU specialist covering both edge and on-premise environments. Developed a proprietary hardware-software integrated SDK ('qb') capable of deploying 400+ AI models without additional tuning.

Key Products:

• ARIES: Samsung 14nm, 80 TOPS NPU.
• MLX-A1: Industrial AI PC, 70W.
• REGULUS: Next-gen chip under development.

Foundry & Process:

• ARIES: Samsung Foundry 14nm (14LPP)
• REGULUS: Advanced node under evaluation

Recent News: ARIES mass production began in 2024. Participated in CES 2026. Achieved 80% power reduction compared to conventional GPUs.

---

🚗 Automotive AI Semiconductors

6. Telechips

Founded: 1999 (KOSDAQ listed)

A veteran in automotive multimedia SoCs, recently launching an automotive AI accelerator targeting the SDV (Software Defined Vehicle) era. Operates global offices in Japan, China, the US, Germany, and Singapore.

Key Products:

• Dolphin5: Arm-based automotive SoC.
• A2X (TCA2000/TCA1000): Korea's first automotive AI accelerator. 200 TOPS NPU. Adopts SAPEON's X300-based NPU IP.

Foundry & Process:

• Dolphin5: TSMC 16nm FFC (FinFET Compact)
• A2X: TSMC 12nm class (AEC-Q100 automotive-grade certified process)

---

🔧 Semiconductor IP / ASIC Platform

7. OPENEDGES Technology

Founded: 2017 (KOSDAQ listed)

The world's only provider of a unified "Total Memory Subsystem + AI Platform IP" solution. Supports AI semiconductor design for edge environments including autonomous driving and security cameras with integrated NPU and high-performance memory systems.

Key Products:

• Enlight: AI accelerator IP + network model compiler optimization.
• DDR memory controller IP, DDR PHY IP, NoC interconnect IP.

Foundry & Process:

• As an IP provider, OPENEDGES does not manufacture directly. IP is ported and validated for the foundry and process node chosen by each customer. Supports TSMC 5nm–28nm, Samsung Foundry, and other major processes.

---

8. SEMIFIVE

Founded: 2019 (KOSDAQ listed, December 2025)

An AI ASIC partner providing end-to-end solutions from custom chip design to mass production. Specializes in advanced process nodes, large-die design, and 3D-IC technology.

Foundry & Process:

• TSMC-focused: Core expertise in leading-edge nodes — N5 (5nm), N4 (4nm), N3 (3nm).
• Also supports Samsung Foundry processes, including chiplet and 3D-IC heterogeneous integration.
• Has executed designs for Rebellions (REBEL, TSMC 4nm), HyperAccel (Bertha, TSMC 4nm), and FuriosaAI (RNGD, TSMC 5nm).

Key Customers: Design partner for major domestic AI chip startups including Rebellions, FuriosaAI, and HyperAccel. Secured large enterprise customers including Hanwha Vision.

Growth Metrics: New orders grew from ₩57.2 billion (2022) to ₩123.9 billion (2024), more than doubling.

---

💾 AI Infrastructure (SSD Controllers)

9. FADU

Founded: 2015 (KOSDAQ listed, Korea's first fabless unicorn)

A fabless specialist in enterprise NVMe SSD controllers — a critical component of AI datacenter infrastructure. Focused on AI server storage solutions, supplying to global hyperscalers.

Key Products:

• ECHO / Gen5: PCIe Gen5 NVMe SSD controller. NVIDIA GB200 NVL72 certified.

Foundry & Process:

• Gen5 controller: TSMC 5nm (N5). Adopts leading-edge process to meet the low-power, high-density demands of high-performance SSD controllers.

Recent News: Confirmed mass production supply to 2 of the 4 major hyperscalers (AWS, Google, Microsoft, Meta). Partnered with SK Hynix to supply SSDs to Meta. First-half 2025 revenue reached 99% of the full prior year.

---

🧠 AI Software / Model Optimization

10. Nota AI

Founded: 2015 (KOSDAQ listed, November 2025)

A specialist in AI model optimization and compression platforms for edge and on-device environments. Set the record as the first AI optimization company to list on KOSDAQ in South Korea.

Key Products:

• NetsPresso®: Automated AI model optimization and compression platform.

Foundry & Process:

• Software specialist; does not produce silicon directly. Provides software optimized to run across diverse hardware platforms including NVIDIA, Qualcomm, and Renesas.

Key Customers: NVIDIA, Samsung Electronics, Qualcomm, Sony, Renesas and other global semiconductor and AI companies.

Recent News: 2024 revenue grew 300% YoY. Raised approximately ₩258 billion in Series C.

---

11. ENERZAi

A specialist in edge AI software that enables high-performance AI on low-power, small-form-factor devices without dedicated AI accelerators.

Core Technologies:

• Optimium: MLIR-based AI compiler.
• 1.58-bit ultra-low bitwidth quantization: Less than 0.39% accuracy loss, 2.46x speed improvement, 77.3% memory reduction.

Foundry & Process:

• Software specialist; does not use a foundry directly. Delivers a software stack that runs on diverse edge processors including Arm Cortex and RISC-V.

Recent News: Selected for Arm AI Partner Program. Global edge AI partnerships with Synaptics and Advantech. Participated in CES 2026.

---

📊 South Korea's AI Fabless Companies at a Glance

---

🏭 Foundry Partner Landscape

There is a clear pattern in the foundry choices made by South Korean AI fabless companies.

TSMC-centric (High-performance datacenter chips) High-performance datacenter chips — including Rebellions' REBEL, FuriosaAI's RNGD, and HyperAccel's Bertha — are all manufactured on TSMC's 4–5nm leading-edge processes. The consensus is that TSMC's process technology leads in both performance and power efficiency at these nodes.

Samsung Foundry (Edge AI chips) Edge AI NPU companies such as DeepX (5nm) and Mobilint (14nm) primarily leverage Samsung Foundry. Geographic proximity, existing collaborative relationships, and integration with the domestic ecosystem are key advantages.

SEMIFIVE — The gateway for Korean fabless companies to TSMC Through its close partnership with TSMC, SEMIFIVE plays a critical intermediary role — providing leading-edge process design services to domestic startups that would otherwise struggle to access TSMC directly.

---

Closing Thoughts

South Korea's AI fabless ecosystem is now growing at its fastest pace ever. Rebellions and FuriosaAI are competing with global players in the datacenter NPU market, while DeepX has entered the Chinese market through its Baidu partnership. FADU has secured global hyperscalers as customers, and both Nota AI and SEMIFIVE have successfully listed on KOSDAQ.

South Korea's challenge to crack NVIDIA's GPU dominance — the road ahead is still long, but the direction is clear. The new chapter of K-semiconductors is beginning right now.

---

This article was written based on information as of April 2026. Given the rapidly evolving nature of the AI semiconductor market, some details may have changed.