Rebellions Challenges Nvidia: A More Efficient and Affordable AI

⚡

Key Takeaways

1Rebellions, a start-up from Seoul, has raised $400 million to develop more energy-efficient AI inference systems.

2The RebelRack promises a sixfold reduction in consumption and an acquisition cost up to 75% lower than Nvidia GPUs.

3Rebellions' systems are compatible with existing data centers, without requiring liquid cooling.

💡Why it matters — Rebellions could disrupt the data center market by providing more accessible and energy-efficient AI solutions.

Rebellions Raises $400 Million to Revolutionize AI

The Korean start-up Rebellions, founded in 2020 in Seoul, has recently raised $400 million in a pre-IPO round. This funding round was led by Mirae Asset Financial Group and the Korea National Growth Fund, bringing the total funds raised to $850 million and valuing the company at $2.34 billion. Rebellions positions itself as an innovative player in the field of artificial intelligence, with the ambition to drastically reduce the costs and energy consumption of AI inference systems.

More Energy-Efficient Inference Systems

Rebellions has launched two inference systems, the RebelRack and the RebelPOD, which promise to reduce energy consumption by six times compared to current solutions, while offering an acquisition cost up to 75% lower than Nvidia GPUs. These systems are designed to easily integrate into existing data centers thanks to their standard 19-inch rack format and air cooling, eliminating the need for liquid cooling. According to Sunghyun Park, co-founder and CEO of Rebellions, "AI is now measured by its ability to operate in the real world, at scale, under energy constraints, and with a clear return on investment."

Compatibility and Performance

The RebelRack consists of four nodes, each integrating eight cards, totaling 32 Rebel100 accelerators. These accelerators offer a performance of 64 petaFLOPS in FP8 and an aggregated memory bandwidth of 153.6 terabytes per second. Each card has a TDP of 600 W in PCIe format. The RebelPOD, on the other hand, can be configured in clusters of 8 to 128 nodes, interconnected via an 800 Gbps Ethernet network. Marshall Choy, Chief Commercial Officer of Rebellions, stated that anyone who has worked with vLLM, PyTorch, or Red Hat OpenShift already knows how to use the platform, as the entire stack is open source, from vLLM to OpenShift. Rebellions is also a member of the PyTorch Foundation, a notable fact in the limited world of AI chip start-ups.

Support from Samsung and SK Hynix

The Rebel100 chips utilize a chiplet architecture with four compute dies, manufactured by Samsung using the SF4X process, and are equipped with 144 GB of HBM3e memory, offering a bandwidth of 4.8 TB/s per chip. Rebellions benefits from the support of Samsung and SK Hynix, two giants in HBM memory, ensuring supply chain flexibility. Currently, Rebellions sources HBM from Samsung but can switch to SK Hynix without any particular friction.

International Expansion and Outlook

With a total of $850 million raised and a valuation of $2.34 billion, Rebellions is expanding internationally, with offices in Japan, Saudi Arabia, Taiwan, and the United States. The start-up plans to file for an IPO by the end of 2026 or early 2027, with JPMorgan as its financial advisor. Marshall Choy also emphasized that developing the interconnection network is a priority, aiming to move from rack-level thinking to row-level, and then to the entire data center.