The brief

Analyst view

• Rebel100 claims 3.8x perf-per-watt, 1.9x lower capex, 2.1x lower opex versus GPU nodes.
• POC trials with Meta and xAI. Planning KOSPI IPO in 2027.

Technical approach

As reported

Chiplet design

Modular NPU scaling

Flexible scaling of performance and memory across configurations.

Software stack

Open standards

Kubernetes, PyTorch, and Hugging Face support.

Supply chain

Memory partnerships

Samsung and SK Hynix relationships provide memory allocation advantages.