HyperAccel
HyperAccel is a Seoul-based fabless AI semiconductor company developing LPU (LLM Processing Unit) chips optimized for LLM inference. It serves cloud data centers, edge devices, and on-premise deployments via Bertha and Forte accelerators, the HX-F55X server, licensable LPU IP, and the HyperDex software stack.
- Company typePrivate
- Founded2023
- HeadquartersSeoul, South Korea
- Headcount51–100
- GTM typeB2B
- OfferingHardware or Manufacturing
What HyperAccel does
HyperAccel is a fabless AI semiconductor company founded in January 2023 in Seoul, South Korea, that designs and develops LPU (LLM Processing Unit) chips purpose-built for end-to-end LLM inference. The company's core technology is a coarse-grained, programmable processor architecture with a streamlined dataflow that directly connects external LPDDR5X or HBM2 memory to compute engines, eliminating traditional memory hierarchies and achieving roughly 90% memory bandwidth utilization compared with ~50% on general-purpose GPUs. Multi-chip scaling is enabled by a proprietary interconnect called ESLink (Expandable Synchronization Link), and the architecture is supported by peer-reviewed publications at Hot Chips 2023, IEEE Micro 2024, VLSI 2025, and ISCA 2025.
HyperAccel's product portfolio spans the LPU value chain: licensable LPU processor IP (in latency-optimized, high-performance, and balanced profiles); the Bertha 500 4nm ASIC accelerator (768 TFLOPS FP8, 546 GB/s, 250W TDP) for cloud data centers, with first samples in Q1 2026 and mass production targeted for early 2027; the Bertha 100 6nm edge SoC (32.768 TFLOPS FP8, M.2 form factor) co-developed with LG Electronics, with samples in Q4 2026; the Forte 55X FPGA-based accelerator built on AMD Alveo U55C for on-premise inference; and the HX-F55X server system integrating Forte 55X accelerators with ASUS chassis. The HyperDex software platform supports HuggingFace, vLLM, PyTorch, ONNX, and OpenAI-compatible APIs, and includes the Legato domain-specific language for kernel authoring. An LPU AFI and AMI are listed on AWS Marketplace for on-demand, usage-based access via Amazon EC2 F2 instances.
The business model combines hardware sales (Bertha and Forte accelerators, HX-F55X servers), LPU IP licensing, design services under joint development agreements, and cloud-based LPU-as-a-Service on AWS. HyperAccel serves hyperscalers and data center operators (Naver Cloud as the flagship PoC customer), Korean national research institutions (Seoul National University for Orion servers), consumer electronics OEMs (LG Electronics for the Bertha 100), and global self-serve cloud developers via AWS Marketplace. The company is pre-revenue, with cumulative funding of approximately $45M (Series A of KRW 55B / ~$40M USD led by Korea Investment Partners in September 2025), a Series B expected by June 2026, and an IPO targeted for around 2028.
HyperAccel firmographics
Firmographics- Name
- HyperAccel
- Legal name
- HyperAccel Inc.
- Website
- https://hyperaccel.ai
- Company type
- Private
- Founded year
- 2023
- Operating status
- Operating
- Headcount range
- 51–100 employees
- Short description
- HyperAccel is a Seoul-based fabless AI semiconductor company developing LPU (LLM Processing Unit) chips optimized for LLM inference. It serves cloud data centers, edge devices, and on-premise deployments via Bertha and Forte accelerators, the HX-F55X server, licensable LPU IP, and the HyperDex software stack.
- Ownership category
- akta.pro rank
HyperAccel industry classification
Industry- Product category
- AI Inference Semiconductors
- NAICS
- Computer Systems Design and Related Services (54151), Computer Systems Design Services (541512)
- SIC
- Services-Computer Integrated Systems Design (7373), Services-Computer Programming, Data Processing, Etc. (7370)
- akta.pro primary industry
- Compute & Acceleration ICs (CPU/GPU/AI/FPGA/ASIC) (HDAHABAC)
- akta.pro secondary industry
- AI/Neural Processing Units (NPUs) & ML Accelerators (HDAHAJAF)
Keywords
Where HyperAccel is headquartered
LocationHeadquarters
- HQ city
- Seoul
- HQ country
- South Korea
- HQ region
- Asia
Markets served
HyperAccel business model
Business model- GTM type
- B2B
- Offering type
- Hardware or Manufacturing
- Cost components
- Personnel, Technology or R&D, Supply Chain, Marketing or Sales, Operations
Revenue model
- Accelerator hardware sales: Sale of LPU-based accelerator products (Bertha 500 cloud accelerator, Bertha 100 edge accelerator, Forte 55X based on AMD Alveo U55C FPGA) packaged as PCIe cards or M.2 modules. Mass production planned for 2027 with initial data center customers.
- Server sales (Orion / HX-F55X): Turnkey inference servers (Orion / HX-F55X) combining the LPU accelerator with ASUS server hardware, sold to Korean national research institutes and other customers for on-premise, data-private AI inference.
- LPU IP licensing: Licensing of the LPU processor IP (configurable for latency-optimized, high-performance, or balanced profiles) to other semiconductor and SoC designers; supported via PCIe and UCIe protocols for seamless SoC integration.
- Design services and joint development: Design and engineering services delivered under joint development agreements (e.g., with LG Electronics for the Bertha 100 edge SoC, where HyperAccel handles LPU design and LG provides backend services).
- Cloud-based LPU-as-a-Service:
Pricing tiers
| Model | Billing | Price |
|---|---|---|
| Other | — | Quote-based pricing for Bertha 500 accelerator — request via 'Ask Inquiry' form |
| Other | — | Quote-based pricing for Bertha 100 edge accelerator — request via 'Ask Inquiry' form |
| Usage-based | Pay-as-you-go | Usage-based hourly pricing for LPU access on AWS EC2 F2 instances via AWS Marketplace |
| Other | — | Quote-based IP licensing for LPU IP (latency-optimized / high-performance / balanced profiles) |
Go-to-market motion6 records
Distribution channels7 records
Marketing channels8 records
HyperAccel product offering
Product offeringCore offering
HyperAccel designs and sells LPU (LLM Processing Unit) AI inference accelerators for cloud data centers and edge devices. Its product line spans the 4nm Bertha 500 cloud ASIC, the 6nm Bertha 100 edge SoC co-developed with LG Electronics, the AMD Alveo-based Forte 55X FPGA accelerator, the HX-F55X turnkey inference server, licensable LPU processor IP, and the HyperDex software stack supporting HuggingFace, vLLM, PyTorch, ONNX, and OpenAI-compatible APIs.
Product overview
HyperAccel offers a vertically integrated LPU (LLM Processing Unit)-based AI semiconductor platform spanning chip IP, data-center accelerators, edge SoCs, complete server systems, and a full software stack. The core processor technology — the LPU — is offered as licensable IP in three configurations (Latency Optimized for Edge Devices, High Performance for Cloud Servers, and Balanced for Edge Servers/Devices). On top of this IP, HyperAccel has built the Bertha 500 4nm ASIC accelerator for cloud data centers (targeting Naver Cloud and global Big Tech customers with mass production in early 2027), the smaller Bertha 100 6nm edge SoC jointly developed with LG Electronics for agentic AI on IoT and appliances (samples Q4 2026), and the FPGA-based Forte 55X accelerator built on AMD Alveo U55C. The Forte 55X is packaged together with an ASUS server chassis into the HX-F55X server system for on-premise datacenter inference. The HyperDex software platform ties everything together, providing HuggingFace/vLLM/PyTorch/ONNX/OpenAI-compatible APIs and the Legato DSL for kernel authoring on LPU hardware. Related research contributions include the Adelia 4nm accelerator paper at VLSI 2025 and the Hybe GPU-NPU hybrid system at ISCA 2025.
Differentiator
Problem solved
Functional benefit
Brands
- Bertha 500: Flagship 4nm LPU-based AI inference accelerator (B500) for cloud data centers, with 768 TFLOPS, 546 GB/s LPDDR5X bandwidth, up to 1024 batch size, and 250 W TDP.
- Bertha 100
- Forte 55X
- HX-F55X
- LPU (LLM Processing Unit)
- HyperDex
Products and services
- Bertha 500 Flagship 4nm ASIC LPU accelerator for cloud data center AI inference, delivering 768 TFLOPS (FP8) at 250W TDP with 546 GB/s LPDDR5X bandwidth and batch size up to 1024. Claims 2x throughput, 19x cost efficiency, and 12x power efficiency vs NVIDIA H100; initial supply targeted to Naver Cloud and global Big Tech customers.
- Bertha 100 Edge LPU SoC for agentic AI and multi-modal on-device inference, co-developed with LG Electronics and fabricated on TSMC 6nm. Delivers 32.768 TFLOPS (FP8) with 64 GB/s LPDDR5X bandwidth in M.2 form factor, supporting text, image, audio, and video for IoT, smart appliances, robotics, and on-premise workloads.
- Forte 55X FPGA-based accelerator for high-speed, low-latency on-premise AI inference, built on AMD Xilinx Alveo U55C with HBM2 (460 GB/s). Delivers 1.7x higher throughput and 1.42x higher power efficiency than NVIDIA L4, single-slot 75W operation, co-marketed with AMD.
- HX-F55X Turnkey on-premise inference server combining eight Forte 55X accelerators with an ASUS chassis, aggregating 3.68 TB/s DRAM bandwidth and 128 GB capacity at 576W TDP, delivering 1.33x higher throughput than 2x NVIDIA H100.
- LPU IP Licensable LPU processor IP for LLM inference, reconfigurable across memory types and compute resources. Three configurations (Latency Optimized for Edge, High Performance for Cloud, Balanced for Edge Servers) using standard PCIe and UCIe protocols for SoC integration.
- HyperDex Software Platform Full-stack software platform bridging AI applications and LPU hardware, including HyperDex Toolchain, Legato domain-specific language (Python-embedded eDSL), plug-and-play vLLM and HuggingFace support, ONNX/PyTorch compatibility, OpenAI-compatible APIs, and a model zoo for LLM deployment.
Quantifiable outcome
- ~2x greater throughput vs NVIDIA H100 on LLM inference
- +8 more outcomes
Companies that use HyperAccel
Customer profileNamed customers3 records
Ideal customer profiles4 records
HyperAccel technology and API
TechnologyTechnology focussed Yes
API detail
- Has API
- Yes
- API docs
- API detail
Core technology
AI maturity
App detail
Integration8 records
AI capability13 records
Feature8 records
HyperAccel partnerships and signals
Strategic signalPartnerships
14 partnerships are on record, tiered core, flagship and minor.
- SEMIFIVEcoreCollaboration presented at Samsung Advanced Foundry Ecosystem Forum 2026 on 'Bertha,' an LPU-based AI inference accelerator successfully designed using Samsung Foundry's 4nm process with a Big Die exceeding 500mm².
- UpstagecoreHyperAccel joined the Upstage consortium for South Korea's government-led sovereign AI foundation model project ('독자 AI 파운데이션 모델 / 독파모'), providing LPU-based inference infrastructure, vLLM/PyTorch full-stack software, and efficiency expertise.
- AdvantechcoreMOU signed on February 2, 2026 at HyperAccel's Seoul HQ for AI infrastructure technology collaboration. Advantech brings ~30-40% global share in industrial IoT/embedded boards, expanding HyperAccel's reach into industrial computing, robotics, and medical segments.
- Amazon Web Services (AWS)flagshipHyperAccel publishes a joint AWS Technical Blog case study and lists its LPU AFI plus a HyperAccel AMI on AWS Marketplace, deployable on Amazon EC2 F2 instances. Provides on-demand cloud access to LPU inference for customers globally.
- TSMCcoreManufacturing partner for the smaller, on-device Bertha 100 chip (edge LPU) on TSMC's 6-nanometer process, enabling a separate supply chain path from Bertha 500.
- Motif TechnologiesminorHyperAccel joined Motif Technologies' consortium (alongside Elice for AI infrastructure, RLWRLD for physical AI, N.Light/Crowdworks/XYZ Robotics for 3D data and robotics) for South Korea's national AI foundation model project.
- MangoBoostminorMOU signed for next-generation AI infrastructure advancement, reported by Korean media.
- LG ElectronicsflagshipJoint development agreement (Bertha Edge Project Partnership Commitment signed Nov 2025) to build the Bertha 100 edge SoC for home appliances and robots. HyperAccel provides LPU IP/design services; LG provides in-house IP (PHY/memory controller) and backend services.
- DinotisiacoreJoint collaboration to develop a Retrieval-Augmented Generation (RAG) optimized AI inference system combining Dinotisia's Vector Data Processing Unit (VDPU) with HyperAccel's LPU into a single unified system, claimed to be the world's first RAG-specialized AI system.
- AMDcoreForte 55X uses AMD Xilinx Alveo U55C FPGA; HyperAccel and AMD co-market after functional and performance validation (case study published). HX-F55X system integrates the Forte 55X with ASUS servers for on-premise inference.
- RebellionscoreCo-lead of the government-led K-Cloud development overall project (HyperAccel–Rebellions consortium, KRW 4.030 trillion total program). Both companies are participants in the K-Perf joint benchmark at AISFC 2025.
- TrustlabminorCollaborates with AMD and Trustlab to expand real-world deployment of HyperAccel LPU semiconductor technology.
- Samsung Electronics FoundryflagshipManufacturing partner for HyperAccel's Bertha 500 LPU on Samsung's 4nm process (die area exceeding 500mm² per SEMIFIVE description). HyperAccel collaborates with Samsung Foundry and Korean design houses for end-to-end chip production.
- Naver CloudflagshipCore joint development agreement and ongoing PoC for LLM inference; Bertha 500 is co-developed with Naver Cloud and will be primarily supplied to Naver Cloud at launch. Joint demo of NAVER HyperCLOVA X running on HX-F55X.
Scale indicators13 records
Recent moves7 records
Expansion highlights9 records
HyperAccel competitors and assessment
Company assessmentDirect peers
- Graphcore: UK-based developer of the IPU (Intelligence Processing Unit) for AI workloads, recently pivoted toward inference; comparable specialized non-GPU architecture targeting transformer model acceleration.
- FuriosaAI: Korean AI semiconductor startup developing the RNGD NPU for data-center LLM inference; fellow Korean peer competing for the same domestic hyperscaler and sovereign AI workloads as HyperAccel.
- Cerebras Systems: Developer of the Wafer-Scale Engine (WSE) for AI training and inference; competes with HyperAccel for high-throughput, low-latency LLM serving customers, particularly hyperscalers and enterprises seeking NVIDIA alternatives.
- SambaNova Systems: Builds RDU (Reconfigurable Dataflow Unit) chips and SN40L accelerators purpose-built for enterprise generative AI inference; directly comparable product positioning and enterprise go-to-market to HyperAccel's Bertha 500.
- Rebellions: Korean AI chip startup building data-center NPUs (ATOM, ION) for LLM inference; co-leads the K-Cloud project with HyperAccel and is the most direct head-to-head competitor for Korean sovereign AI inference workloads.
- Hailo: Israeli fabless company building edge AI accelerators (Hailo-8, Hailo-10H); most directly comparable to HyperAccel's Bertha 100 edge SoC positioning for IoT, smart appliances, and on-device inference.
- Groq: Pioneer of LPU (Language Processing Unit) inference accelerators with deterministic LPU architecture and GroqCloud LPU-as-a-Service; closest architectural analog to HyperAccel's LPU strategy and competes for the same inference workloads.
- Tenstorrent: AI accelerator company founded by Jim Keller selling Grayskull/Wormhole chips and licensing RISC-V-based IP; comparable hybrid hardware-plus-IP model to HyperAccel's accelerator-plus-LPU-IP strategy.
Emerging players
- Lightmatter: Photonics-based AI compute and interconnect company targeting data-center inference; an emerging alternative silicon approach competing for the same inference-efficiency narrative as HyperAccel's LPU.
- Positron: US-based startup building inference-optimized accelerators (Positron Atlas) emphasizing memory-bandwidth-efficient transformer inference; emerging challenger with overlapping technical thesis to HyperAccel.
Market position
Strengths5 records
Weaknesses5 records
Competitive moat6 records
Key risks6 records
Key highlights7 records
Customer concentration
HyperAccel social profiles
Digital presenceHyperAccel financial estimates
Financial estimateRevenue estimate
Valuation estimate
HyperAccel leadership team
Management profileNumber of profiles
Profiles4 records
HyperAccel funding detail
Funding detailFunding overview
Funding rounds1 record
Investors11 records
Funding detail is available on the Subscription and Enterprise plan.Contact sales →
HyperAccel M&A and investment
M&A and investmentM&A
Investments
M&A and investment is available on the Subscription and Enterprise plan.Contact sales →
Frequently asked questions about HyperAccel
What does HyperAccel do?
HyperAccel designs and sells LPU (LLM Processing Unit) AI inference accelerators for cloud data centers and edge devices. Its product line spans the 4nm Bertha 500 cloud ASIC, the 6nm Bertha 100 edge SoC co-developed with LG Electronics, the AMD Alveo-based Forte 55X FPGA accelerator, the HX-F55X turnkey inference server, licensable LPU processor IP, and the HyperDex software stack supporting HuggingFace, vLLM, PyTorch, ONNX, and OpenAI-compatible APIs.
Is HyperAccel a public or private company?
HyperAccel is a private company. It is classified as venture growth investor backed and is currently operating.
When was HyperAccel founded?
HyperAccel was founded in 2023. It employs 51 to 100 people.
Where is HyperAccel based?
HyperAccel is headquartered in Seoul, South Korea, in the Asia region.
How does HyperAccel make money?
Five revenue lines are on record. Accelerator hardware sales are the primary driver. The others are server sales (Orion / HX-F55X), LPU IP licensing, design services and joint development and cloud-based LPU-as-a-Service.
Who are HyperAccel's main competitors?
Direct peers on record are Graphcore, FuriosaAI, Cerebras Systems, SambaNova Systems, Rebellions, Hailo, Groq and Tenstorrent. Emerging players are Lightmatter and Positron.
Does HyperAccel have an API?
Yes. HyperAccel offers a developer-facing SDK and inference API via the HyperDex Toolchain that supports HuggingFace-style generate() calls and vLLM-style API patterns. The platform also exposes OpenAI-compatible completions() and chat.completions() interfaces, allowing existing inference code to be reused with minimal modification on LPU hardware. An Amazon FPGA Image (AFI) and AMI are available on AWS Marketplace for cloud-based LPU access (F2 instances), with a bundled Chat UI, vLLM plugin, Docker runtime, and HyperDex Toolchain pre-installed. Developer documentation is at docs.hyperaccel.ai/1.5.2.
What industry is HyperAccel in?
HyperAccel's product category is AI Inference Semiconductors. Its primary akta.pro industry code is HDAHABAC, Compute & Acceleration ICs (CPU/GPU/AI/FPGA/ASIC), with a secondary code of HDAHAJAF, AI/Neural Processing Units (NPUs) & ML Accelerators. Its NAICS code is 54151 and its SIC code is 7373.