HyperAccel
Category: AI Chips / Semiconductors
A fabless AI semiconductor startup that develops 'Latency Processing Units' (LPUs) and high-performance servers specifically architected to accelerate Large Language Model (LLM) inference with superior cost and energy efficiency. HyperAccel was founded in 2023. The company is led by Joo-Young Kim. Based in Seoul, South Korea. Team size: 50-100. Total funding raised: ₩61B (~$45M). Latest round: Series A (₩55B, Dec 2024). Key investors include Korea Investment Partners, KDB Bank, KB Investment, Mirae Asset Venture Investment, SBVA, BonAngels Venture Partners, Company K Partners, KDB Capital, LB Investment, Vickers Venture Partners.
- Founded
- 2023
- Headquarters
- Seoul, South Korea
- Team size
- 50-100
- Total funding
- ₩61B (~$45M)
Value proposition
Delivers up to 2.4x higher cost-efficiency and 5x faster processing speeds than high-end GPUs for LLM tasks by maximizing memory bandwidth utilization (up to 90%) and using cost-effective LPDDR5X instead of HBM.
Products and solutions
LPU (LLM/Latency Processing Unit) - Proprietary AI chip architecture, Bertha 500 - Flagship 4nm LPDDR-based data-center LLM inference accelerator (samples/PoC 2026; mass production targeted ~2027), Bertha 100 - On-device / edge LPU variant, Forte 55X - LPU + AMD Alveo U55C acceleration card, HX-F55X (HyperAccel Xceleration; formerly Orion) - On-premise inference server, HyperAccel software stack - vLLM/HuggingFace/PyTorch toolchain, On-device AI IP - Silicon IP for home appliances and robotics
Unique value
Developed the world's first LPU architecture that bypasses the traditional memory bottleneck of GPUs by using a streamlined dataflow and LPDDR5X-based memory systems specialized for transformer models.
Target customer
Hyperscale data centers, Cloud Service Providers (CSPs), AI service developers (AIaaS), and Consumer Electronics/Robotics manufacturers.
Industries served
Artificial Intelligence, Cloud Computing, Semiconductors, Consumer Electronics, Robotics
Technology advantage
Achieves high-speed token generation (30-50% faster than standard services on Llama 3) through model parallelism, intra-layer parallel instructions, and optimized memory alignment without relying on expensive High-Bandwidth Memory (HBM).
How they differentiate
HyperAccel differentiates by utilizing cost-effective LPDDR5X memory instead of the industry-standard High-Bandwidth Memory (HBM), achieving up to 2.4x higher cost-efficiency and 5x faster processing speeds for LLM tasks compared to high-end GPUs.
Main competitors
Groq, Rebellions, SambaNova Systems, NVIDIA
Key partnerships
LG Electronics (On-device AI / robotics PoC), Naver Cloud (Business verification / HyperCLOVA PoC), Samsung Electronics (4nm foundry for Bertha), TSMC (On-device chip foundry per Apr 2026 CEO interview), Semifive (Bertha design/DSP partner), AMD (Forte 55X / Alveo), Advantech (AI infrastructure MOU, Feb 2026), MangoBoost (AI infra MOU, Dec 2025), AnyBridge (AI infra full-stack MOU, Aug 2026), Upstage consortium (Korea independent foundation-model infra, Feb 2026), Daewon CTS (Distribution / PoC)
Notable customers
LG Electronics (On-device AI / robotics PoC), Naver Cloud (HyperCLOVA / business verification PoC), Seoul National University (Orion server supply), Daewon CTS
Major milestones
Founded January 2023 by KAIST Prof. Joo-Young Kim, ₩6B Seed (Aug 2023); ₩55B Series A (Dec 2024), Launched Orion Gen-AI server (2024); rebranded line as HX-F55X, Bertha data-center LPU samples produced (Mar 2026); PoC with Naver Cloud & LG; mass production targeted ~2027, Pursuing Series B (~₩150–200B target; Krafton reported as ~₩50B SI, Jun 2026) — close not officially confirmed as of Aug 2026, Selected 14th Innovation Icon (KODIT, Nov 2025); CEO received MSIT Information & Communications merit award (Apr 2026)
Growth metrics
Headcount ~87–92 (THE VC / NPS mid-2026); Bertha 4nm samples in 2026; pursuing Series B at ~₩600B valuation target.
Market positioning
Specialized AI fabless challenger focused on high-efficiency Large Language Model (LLM) inference acceleration.
Geographic focus
South Korea, North America (targeting global data centers and consumer electronics)
Patents and IP
Disclosed 6 registered/filed patents regarding LPU architecture and LLM acceleration techniques.
About Joo-Young Kim
Joo-Young Kim is a former Associate Professor at KAIST (School of Electrical Engineering) and served as a Hardware Engineering Leader at Microsoft Azure in Redmond, where he spearheaded hardware acceleration for the Azure Data Lake platform. He is a globally recognized expert in AI semiconductor architecture, specifically specializing in Latency Processing Units (LPUs) for large-scale language model inference.
More AI Chips / Semiconductors companies
Official website: https://hyperaccel.ai/