South Korean chip startups say the country must build large-scale commercial references at home to expand exports of domestically developed neural processing units (NPUs). With the technical performance of Korean NPUs now reaching a globally competitive level, they argue the government should widen real-world deployment opportunities to help persuade overseas customers and win contracts.
Chief executives of leading domestic NPU startups including Rebellions, FuriosaAI and DeepX made the case at a National Assembly policy forum on expanding exports and diversifying sales channels for Korean NPUs, held the previous day at the National Assembly Members' Office Building in Yeongdeungpo District, Seoul, according to industry sources on the 11th. The session was chaired by Rep. Ko Dong-jin of the People Power Party, a member of a lawmakers' study group called the Forum for Structural Reform for Sustainable Growth.
Industry executives agreed that Korea needs a virtuous cycle in which larger domestic pilot projects and commercial deployments become references that lead to global orders.
Park Sung-hyun, chief executive of Rebellions, said government-provided opportunities to test chips on public infrastructure have helped Korean NPUs move toward commercialization abroad, but stressed that the scale of those projects must grow. "I always feel the volume is lacking," the CEO said, explaining that securing large-scale commercialization experience that generates actual service revenue — rather than stopping at small proof-of-concept projects — is what creates meaningful overseas business opportunities.
"This is no longer an era of competing on a single chip," the CEO said. "Software matters more than hardware, and the ability to actually deploy and optimize at rack scale matters more than software." The point is that what decides the outcome is not only the performance of individual chips but experience in deploying NPUs and running them reliably in large-scale environments combining multiple servers and accelerators.
The Rebellions chief added that the company must be able to say, "Use this product because it is competitive on the global stage," rather than simply, "Use it because it is a Korean product," and that large-scale commercialization experience at home provides the basis for that argument. "Our next-generation Rebel100, which is evaluated as delivering performance comparable to Nvidia's H200, will also be deployed at companies around the end of the year so we can build up actual operating experience," the CEO said.
Cho Jae-hong, an executive director at DeepX, said overseas demonstration support programs run by the Ministry of Science and ICT have been "a tremendous help" in widening sales channels for Korean NPUs abroad, but added that competing with Chinese and other foreign vendors requires price competitiveness and after-sales support systems as well as performance. "Voucher support from the government during the export process would be a big help," the executive said. "Price competitiveness is needed to overcome the burden that large overseas companies in countries such as Japan and Germany feel in buying chips from a Korean startup for the first time."
The DeepX executive also pointed to the need for budget support to build technical support systems capable of responding immediately at overseas sites. "Customers demand engineering support staff after purchasing an NPU and immediate responses regardless of time differences," the executive said. "If we could receive budget support for operating local offices, making regular business trips and building technical support systems, we would be able to secure competitiveness in winning orders."
Baek June-ho, chief executive of FuriosaAI, who joined via video, said it is time for Korean NPUs to move beyond technology development into mass production and commercialization. "Policy support is needed to boldly create markets, including carrying out large-scale demonstrations and commercial deployments of Korean AI chips at AI data centers being built in Korea," the CEO said.
The FuriosaAI chief also suggested that large-scale deployment experience gained at home should be channeled into exports, stressing the need to build and validate "full-stack" solutions in Korea that combine not only domestic NPUs but also AI models, software, servers and data center infrastructure. "We need to build comprehensive infrastructure solutions capable of running both domestic foundation models and global AI models," the CEO said. "The large-scale deployment capabilities accumulated this way can later expand into global AI infrastructure projects."
