Rebellions Inks Deal to Plug Its AI Chips Directly Into Nvidia's Infrastructure
As data center operators seek to lower the massive energy costs of generative AI, Korean chip startup Rebellions has secured a landmark deal to integrate its neural processing units directly into Nvidia's server racks.
The dominance of Nvidia in artificial intelligence hardware has forced most competing chip startups to position themselves as complete alternatives. However, Korean AI chip developer Rebellions is taking a hybrid approach. The company has partnered with Nvidia and SK Telecom to connect its upcoming Rebel R200 neural processing units (NPUs) directly to Nvidia's computing platform using NVLink Fusion. The rack-scale systems will be built on Nvidia's MGX architecture and deployed in SK Telecom's data centers starting in 2028.
The collaboration relies on a technique called disaggregated inference, which splits the workload of generating AI responses between two different types of processors. When a user submits a prompt, Nvidia's GPUs will handle the compute-heavy "prefill" stage to read and process the context. The task is then handed off to Rebellions' NPUs for the "decode" stage, generating the actual response token by token. By assigning each phase to the processor architecture best suited for it, the setup aims to boost throughput while cutting power consumption.
For Rebellions, the integration represents a major technical validation, allowing the startup to plug its proprietary silicon into Nvidia's high-speed scale-up network rather than building an entire networking ecosystem from scratch. For the broader AI industry, the deal signals a shift toward heterogeneous computing environments. As frontier models grow to trillions of parameters, data centers are increasingly looking to mix specialized hardware to manage the escalating costs and thermal limits of AI infrastructure.
INTELLIGENCE BRIEF
WHY IT MATTERS
The partnership demonstrates that AI chip startups do not have to replace Nvidia entirely to gain market share. By focusing on the highly specific task of token generation, specialized NPUs can coexist with GPUs in the same server rack to optimize energy efficiency.
WHO IS INVOLVED
Rebellions CEO Park Sung-hyun, Nvidia VP of Robotics and Edge AI Deepu Talla, and SK Telecom.
MARKET IMPACT
The move toward disaggregated inference could reshape data center economics, allowing operators to lower total cost of ownership by offloading the less compute-intensive parts of AI generation to cheaper, lower-power chips.
This story was drafted with AI assistance and reviewed by TurkSpark editors before publication. Facts, figures, and names may be inaccurate — verify important details independently.
Panthalassa to Build Floating AI Data Centers in Washington Following $365M Funding Run
Read
Mavria Emerges from Stealth with $100M Seed to Build AI Data Centers on an Assembly Line
Read