Dr. Sunghyun Park earned his B.S. from KAIST and his Ph.D. from MIT’s CSAIL (Computer Science and Artificial Intelligence Laboratory). He built his career across Intel, Samsung Research, and SpaceX as a semiconductor design engineer, before serving as Vice President at Morgan Stanley in New York — a rare combination of deep technical and financial industry experience.
In 2020, he co-founded Rebellions, growing it into Korea’s first AI semiconductor unicorn. Rebellions has raised $850 million in total funding and is valued at approximately $2.34 billion, backed by the Korean government’s strategic AI investment program alongside leading private investors. With a public offering on the horizon, Rebellions is establishing itself as a credible global player in the AI inference infrastructure market.
Tell us about your company?
Rebellions is a full-stack AI inference chip company building token factories from South Korea. Five of us founded it in 2020 with one belief: AI would eventually be limited by power and cost, not by model capability. I came to this from Intel Labs, Samsung, SpaceX and Morgan Stanley, and my co-founders came from IBM and top research labs. Between us, we saw both the silicon side and the economics side of the problem. That’s also why we started in Korea. The ecosystem is here, with Korea’s strength in sovereign AI and the broader ecosystem across Asia.
As South Korea’s first AI chip unicorn, we’ve raised about $850M in total. That includes a $400M pre-IPO round in March 2026 at a roughly $2.3B valuation, in which we were the first recipient of Korea’s National Growth Fund. Our backers include Samsung, SK hynix, SK Telecom, KT, Arm and Aramco.
What problems are you solving?
The center of gravity in AI has moved from training to inference. Training happens once. Inference runs every time someone uses the model: millions of times a day, 24/7, for years. That changes what matters.
Most data centers aren’t short of chips. They’re short of power. Operators have fixed megawatt budgets, and GPUs designed mainly for training bring high cost and high power draw into a job that runs forever. So the question our customers ask is no longer “how fast?” but “how many tokens per watt, and per dollar?”
Efficient silicon doesn’t matter if it takes months of custom engineering before a model can run on it. Teams need to deploy quickly and keep iterating as models change. That’s why software is part of the product. The RBLN SDK connects the frameworks developers already use – PyTorch, TensorFlow, and Hugging Face – to our NPUs through a compiler, runtime, and optimized software stack, with no retraining required. Models can then be served through familiar frameworks like vLLM, while the same software stack scales from a single RebelServer to RebelRack and RebelPOD.
We design for that question end to end. Our core value is performance-per-watt for inference, delivered as a full system – from silicon and software to racks and pods – that runs today’s largest models in production.”
What application areas are your strongest?
Large-scale LLM inference is our core, including MoE models, coding agents and more. Our second-generation chip, REBEL100, is built on Samsung’s 4nm process. It is the industry’s first chiplet architecture with 144GB of HBM3E, and it delivers 2 PFLOPS at FP8. That combination fits the memory-bound, latency-sensitive nature of serving large models.
We’re deployed in production with Korea’s telcos and cloud providers. KT Cloud became the first Korean cloud to offer NPU-based AI services on our chips. We see strong pull from sovereign AI programs, from telcos building AI infrastructure, and from enterprises that want to run models on-premise at predictable cost.
What keeps your customers up at night?
Three things:
First, power. Many operators can’t get more grid capacity fast enough, so every watt spent on inference is a watt taken from something else.
Second, total cost of ownership. As AI moves from pilots to production, CFOs are asking what each token costs. Hardware that looked fine in a demo can break the business case at scale.
Third, supply and lock-in. Customers want a credible second source, and they want the lowest possible switching cost which includes open source based SW.
What does the competitive landscape look like and how do you differentiate?
NVIDIA is the benchmark and will remain very important. We don’t pretend otherwise. But inference is a different market from training, and it’s big enough to support architectures built only for it. There are also other specialized players: hyperscaler in-house chips and several well-funded startups.
We differentiate in three ways:
- Efficiency by design. We build dedicated inference silicon, not a general-purpose chip adapted to the workload. Performance-per-watt is the metric our architecture is optimized for.
- Open software. Our stack is built on the tools developers already use: PyTorch, vLLM, Triton, Hugging Face, Kubernetes and OpenShift. Customers shouldn’t need to learn a proprietary ecosystem to switch.
- Ecosystem depth. We work closely with Samsung Foundry, SK hynix for HBM, and Arm. Those relationships give us access to leading-edge process, memory and supply that most startups of our size don’t have.
What new features/technology are you working on?
Our focus is moving from chips to systems. In March, we launched RebelRack, a production-ready inference compute unit, and RebelPOD, which connects multiple racks into a scalable cluster. Customers today don’t want a card. They want capacity they can plug into a data center and start serving tokens with.
We recently announced a 100-rack deployment in Japan in partnership with AI&, supporting Japan’s sovereign AI initiatives.
On the silicon side, we keep investing in chiplet integration, memory bandwidth and scale-out networking. On the software side, we’re deepening support for the latest open models and serving frameworks so new models run on day one.
Our next-generation silicon, now in development, will integrate the newest technologies, from memory-centric architecture to hardware-software co-design using AI.
How do customers normally engage with your company?
It usually starts with a technical evaluation based on customer’s specific needs. We give access to hardware, either on-premise or through cloud partners, and our engineers work alongside their team to benchmark real workloads, not synthetic ones. From there, it moves to a pilot deployment and then to rack- or pod-scale rollout.
We work with hyperscalers, telcos, sovereign AI programs, enterprises and system integrators. Some buy systems directly, and others use our capacity through cloud partners. We have teams in Korea and are expanding in the U.S., Japan, Singapore and the Middle East.
Final thoughts?
AI is now measured by how well it works in the real world: at scale, under power limits, and with a clear economic return. The definition of sovereign AI is also changing. Data sovereignty and owning the stack matter, but heterogeneous compute matters more than simply diversifying HWs.
That’s the problem Rebellions was built for. We’d welcome conversations with anyone deploying inference at scale.
Also Read:
Executive Interview with Raghu Belur, Co-Founder of Enphase Energy
CEO Interview with Charlie Cheng of CEO of TeRAM
SemiWiki CEO Interview with Dr. Nissan Maskil of Qarakal Quantum
Share this post via:
Why ASML Is Racing to Build 110 EUV Machines