The disruptive power of open-source ecosystems was the central message delivered by Park Sung-hyun, CEO of Rebellions, who is leading the drive to nurture a "K-Nvidia" in the artificial intelligence (AI) chip market, during his special lecture at "Seoul Forum 2026" held at the Shilla Hotel in Jung-gu, Seoul on the 28th. While U.S.-based Nvidia dominates the AI training market and hardware market through its proprietary CUDA software, Korean companies can turn the tables in the "inference" market—essential for mass-market services—by leveraging open ecosystems, he explained.
"In the past, the key weapon Intel used to topple IBM, the absolute powerhouse of the computer market, was not the performance of the semiconductor itself but its open ecosystem," Park said. "A similar phenomenon is occurring in the AI market today. As big tech companies unite around open source to avoid being tied to Nvidia's proprietary technology, opportunities are emerging for inference-oriented neural processing units (NPUs)."
Rebellions, led by Park, is a fabless company specializing in NPU design optimized for AI inference. His diagnosis is that leadership in AI semiconductors will inevitably shift from graphics processing units (GPUs), led by Nvidia, to NPUs.
"If AI training is the process of building massive foundation models with tens of billions of won invested each month—akin to athletes 'building their bodies' ahead of the Olympics—then inference is the domain where completed models are actually serviced to the public," Park said. "As AI, once used only by experts, is now being used in daily life even by our parents' generation, the largest semiconductor market in human history is opening up in the inference domain."
According to market research firms Gartner and Bloomberg, the global AI semiconductor inference market is projected to grow more than 12-fold over six years, from $39 billion (approximately 58.8 trillion won) in 2024 to over $475 billion (approximately 715.9 trillion won) by 2030. This means the inference chip market alone will surpass the size of the traditional "big three" semiconductor markets as of 2023: DRAM ($112 billion, approximately 169 trillion won), central processing units (CPUs, $97 billion, approximately 146 trillion won), and NAND flash ($72 billion, approximately 109 trillion won).
Behind the explosive growth of the inference chip market lies "power"—the core of AI infrastructure investment. Global data center power demand is expected to surge 30-fold, from 11 gigawatts (GW) in 2024 to 327 GW by 2030.
"The era has come when data centers are built based on total power capacity—whether they are 1 GW-class or 4.3 GW-class—rather than site area or computer specifications," Park said. "How much power is consumed to generate a single token has become the most critical indicator determining a company's profitability."
Rebellions is targeting the cost and power bottlenecks in the inference market. While Nvidia has dominated the AI chipset market through versatility, Rebellions is pursuing a "selection and concentration" strategy that maximizes energy efficiency through designs optimized solely for inference. In benchmark tests and real-world environments, Rebellions' NPUs deliver 5 to 7 times better price efficiency compared to competing systems.
"We respect the versatility and flexibility of Nvidia's ecosystem, but Rebellions has focused exclusively on inference efficiency," Park said, emphasizing the company's "comparative advantage."
Park also stressed that surviving on the global stage requires not only technological capability but also a "capital alliance." Indeed, Rebellions has brought world-class semiconductor companies on board as allies in its pursuit of global competitiveness beyond being Korea's Nvidia. Korea's leading semiconductor companies and the two pillars of memory—Samsung Electronics and SK hynix—have joined as strategic investors (SI), along with ARM and Marvell in the design field, and Amkor and Pegatron in the packaging and assembly field.
"In the global semiconductor market, friends and foes change frequently, but where money is mixed, true allies are made," Park said, emphasizing the meaning of building an ecosystem through investment.
Commercialization achievements are also taking shape. Rebellions is exclusively powering SK Telecom's "A. Call Recording Summary" service and has built NPU-as-a-Service (NPUaaS) through KT Cloud. Overseas, it is building sovereign AI infrastructure in partnership with Saudi Arabia's Aramco.
"Once the inference market enters a full-fledged monetization trajectory, operational efficiency rather than absolute performance will become the deciding factor," Park said in closing. "Rebellions' goal and journey is to confidently stand up to the giant called Nvidia and advance into the global market within the next two to three years."
