400 million and an ambitious goal: Rebellions is the Korean startup challenging NVIDIA
400 million and an ambitious goal: Rebellions is the Korean startup challenging NVIDIA
The South Korean startup Rebellions, founded in 2020 and specializing in AI chip design, has announced a funding round of $400 million, bringing its overall valuation to approximately $2.34 billion. The operation was led by Mirae Asset Financial Group and the Korea National Growth Fund, a vehicle supported by the Seoul government as part of the "K-Nvidia" initiative aimed at strengthening the national ecosystem of advanced semiconductors. With this funding, Rebellions has raised about $850 million in total, with $650 million raised in the last six months alone.
The new resources will primarily be allocated to international expansion, with a particular focus on the United States. The company has already started establishing operational locations in strategic markets such as Japan, Saudi Arabia, Taiwan, and the USA, aiming to build an ecosystem of technology and commercial partners. Their targets include not only large hyperscalers but also advanced research labs and new players in AI, such as Meta and xAI. Simultaneously, the company is conducting proof-of-concept projects with US clients while preparing for an IPO, expected between late 2026 and early 2027, although timelines and location remain unofficial.
Unlike many competitors focused on training, Rebellions aims at inference, that is, the execution of AI models in operational phases. With the proliferation of large language models (LLMs), this phase has become central to ensuring performance, energy efficiency, and economic sustainability. The company claims that its accelerators offer a significant advantage in terms of energy efficiency and performance compared to traditional GPU-based solutions. This positioning puts it in direct competition not only with NVIDIA but also with Cerebras, as well as proprietary projects developed by big tech companies like AWS, Google, and Meta.
The core of the offering is represented by the Rebel100 chips, accelerators designed for high-efficiency inference workloads. Architecturally, this is a chiplet-based solution, with four computing dies produced and assembled by Samsung. Each accelerator integrates HBM3e memory up to 144 GB, with a total bandwidth of 4.8 TB/s, achieving performance of up to 1 petaFLOP in FP16 or double in FP8. The declared TDP is around 600 watts per PCIe card.
On this hardware foundation, Rebellions has developed new infrastructural platforms:
- RebelRack: rack-scale configuration with 32 accelerators, 64 petaFLOPS FP8, 4.6 TB of HBM3e memory, and 153.6 TB/s of aggregated bandwidth.
- RebelPod: scalable solution from 8 to 128 nodes, designed for large-scale deployments in data centers and sovereign clouds.
A distinctive element is the adoption of standard form factors (19-inch racks and air cooling), allowing integration into existing data centers without the need for dedicated infrastructure, unlike the latest high-end liquid-cooled solutions.
On the software side, the platform is based on established open-source stacks like PyTorch, vLLM, and Triton, complemented by tools for disaggregated inference such as llm-d. This approach allows for separating the prefill and decode phases of models, optimizing the use of computational resources and memory.