Rebellions RebelServer with Rebel100 Chiplet AI Accelerators

Rebellions is a South Korean AI semiconductor company with its address in Seongnam-si, Gyeonggi-do, and a Seoul presence. It builds inference accelerators and full-stack systems so that enterprises, telecom operators and governments can run large AI models on-premises with lower energy use than GPU-based infrastructure. Its first-generation ATOM and ATOM-Max chips are in mass production and deployed in Japan, Saudi Arabia and the United States.

RebelServer is a 5U air-cooled inference server with eight RebelCards, each powered by the Rebel100 chip, which uses a chiplet architecture with UCIe-Advanced interconnect and HBM3E memory. A server delivers up to 2 PFLOPS FP8, runs on two AMD EPYC CPUs with 1.5TB DDR5 and 400G networking, and integrates with vLLM and PyTorch for distributed inference of LLMs, mixture-of-experts and multimodal models. It scales into RebelRack and RebelPOD for larger deployments.

• Rebel100 chip: chiplet-based accelerator with UCIe-Advanced interconnect and HBM3E memory.
• Server performance: eight RebelCards delivering up to 2 PFLOPS FP8 in a 5U chassis.
• Host platform: two AMD EPYC 9355 CPUs, 1.5TB DDR5 and NVMe storage for KV cache.
• High-speed networking: 4x 400G QSFP112-DD ports for distributed inference.
• Open software: integration with vLLM and PyTorch, plus SDK documentation and a Model Zoo.
• Scalable systems: expands to RebelRack and RebelPOD turnkey inference infrastructure.

A Vietnamese telecom operator can host its own Vietnamese-language LLM on RebelServer clusters for customer service and enterprise APIs. A government data centre can deploy sovereign AI services on-premises. Cloud providers can offer inference instances based on vLLM with lower power per rack.

Sovereign AI workloads can run on-premises with full control over data and operations. Air-cooled design operating at typical 4-6kW lets data centres add inference capacity without liquid cooling. Support for vLLM and PyTorch eases migration of existing models. A three-year hardware and software support package is included.

Governments and telecom operators are building sovereign AI capacity and want alternatives to GPU supply chains. Chiplet designs with HBM3E are becoming the standard route to scale inference accelerators. Vietnam's push for national AI infrastructure and domestic LLMs creates interest in energy-efficient on-premises inference systems.

Address in Seongnam-si, Gyeonggi-do, South Korea. Raised a USD 250 million Series C in September 2025 at a USD 1.4 billion valuation, with Arm as strategic partner, Samsung Ventures, Pegatron VC, Korea Development Bank, Korelya Capital and Lion X Ventures. Earlier products ATOM, ATOM-Max and REBEL-Quad. Partners include SK Telecom and ai& in Japan.