Domestic NPUs Shift from Slogan to Reality... AI Semiconductor Landscape Transforms

Consensus Grows for AI Inference Semiconductors
Furiosa and Rebellion Secure Mass Production and Server Commercialization Funding via National Growth Fund
DeepX Targets Physical AI, HyperAccel Focuses on LLM Inference Market
K-Perf, Overseas Demonstrations, Cerebras IPO, and Groq-NVIDIA Collaboration

The key theme for the domestic artificial intelligence (AI) semiconductor industry in the first half of 2026 can be summarized as "tangible results." Until recently, domestic neural processing unit (NPU) companies were recognized for their technology, but there were ongoing concerns about insufficient market validation. With NVIDIA's graphic processing units (GPUs) dominating the AI training and data center markets, there were significant doubts about whether Korean AI semiconductors could actually secure real customers and penetrate overseas markets.

However, the atmosphere changed in the first half of this year. Domestic NPU companies have moved beyond prototypes and research and development achievements, advancing to the stage where they are producing actual chips and supplying them to customers as cards, servers, and rack-level infrastructure. The standard for competition has shifted from "Is it domestic or not?" to "Is it actually usable?" The most important factors now are whether AI models can be stably deployed, compatibility with existing software, the ability to reduce power costs, and immediate applicability in data centers and industrial sites.

Paek Joungho, CEO of Furiosa AI, is giving a presentation at the "Renegade 2026 Summit" held on April 2nd at SJ Kunsthalle in Gangnam-gu, Seoul. At this event, CEO Paek expressed confidence in the success of the Renegade chip. Yonhap News

◆ From development success to server delivery: National Growth Fund shares mass production risks = The business of AI semiconductors becomes more challenging after the design phase. Tape-out and prototype production are just the starting points. Actual mass production requires securing foundry production volume, high-priced memory such as HBM, advanced packaging, board design, PCIe card manufacturing, server integration, cooling and power validation, software optimization, and stabilization at the customer site. The "full-stack" approach emphasized by domestic NPU companies this year is essentially a declaration that they intend to handle this entire process themselves.

This is why the Financial Services Commission described the role of the National Growth Fund as capital that shares the risks of advanced industries. AI semiconductor companies must secure wafers, memory, packaging, and server components even before a product is launched. The period for customer validation is also long. With only research and development funds, it is difficult to clear the hurdles of mass production and delivery. This is why Furiosa AI faced a financial situation just a year ago where it had to worry about paying employee wages.

In the first half of this year, the National Growth Fund became a critical turning point for the domestic AI semiconductor industry. The Financial Services Commission approved a direct equity investment of around 800 billion won in Furiosa AI. Of this, the portion directly invested from the Advanced Strategic Industry Fund amounts to 370 billion won. The purpose of the investment is to expand production of the second-generation NPU, Renegade, and to develop third-generation products based on 2-nanometer process and HBM4/4E.

Rebellion was also selected as the first company to receive a direct investment from the National Growth Fund. The Financial Services Commission decided to directly invest 250 billion won from the Advanced Strategic Industry Fund in Rebellion's mass production and next-generation AI semiconductor development projects. The total capital increase, according to the Financial Services Commission, amounts to 600 billion won, structured as a combination of policy finance and private capital.

This is not a simple subsidy. It is long-term capital needed at the stage where domestic NPU companies are actually producing chips and delivering card, server, and rack-level products to customers.

At the "Renegade 2026 Summit" event held on April 2 at SJ Kunsthalle in Gangnam-gu, Seoul, the finished NPU product created using the Renegade chip and the server configuration were unveiled. Photo by Paek Jongmin, Tech Specialist

◆ Furiosa AI emerges at the forefront with mass production and Broadcom partnership = Furiosa AI most clearly demonstrated the trend toward tangible results in the domestic AI semiconductor industry in the first half of this year. The company began full-scale mass production of its second-generation AI inference accelerator, Renegade, and started supplying it to enterprise customers in the form of PCIe cards and turnkey servers. In April, it held the unusual "Renegade 2026 Summit" to present its product commercialization achievements and ecosystem expansion plans to both domestic and international key partners.

Furiosa AI is targeting the AI inference market. While large-scale AI training still revolves around GPUs, the costs and power burden of repetitive inference operations increase as actual AI services proliferate. As chatbots, AI agents, and enterprise generative AI services multiply, the operational costs of data centers rise rapidly. Furiosa AI is focused on performance per watt and total cost of ownership reduction at this crucial juncture.

Collaboration with Broadcom, a leading U.S. semiconductor company, was also a major achievement in the first half. Furiosa AI formed a strategic partnership with Broadcom for the development of a third-generation AI accelerator. The two companies plan to combine Furiosa AI's architecture with Broadcom's XPU technology, Ethernet scale-up, and fabric switch capabilities to develop a next-generation inference platform. This represents an attempt by a domestic NPU company to move beyond chip-level competition to compete at the rack-scale infrastructure level, focusing on communication efficiency across servers and racks, memory utilization, and power efficiency.

On March 17, at the Press Center in Jung-gu, Seoul, representatives of five AI semiconductor companies applauded while listening to the greeting speech of Deputy Prime Minister and Minister of Science and ICT Byung-hoon Bae at the joint public-private meeting of the National Growth Fund 'K-NVIDIA Project'. From left to right: Sungkyu Shin, CFO of Rebellion; Jooyoung Kim, CEO of HyperAccel; Junho Paek, CEO of FuriosaAI; Nokwon Kim, CEO of DeepX; Dongju Shin, CEO of Mobilint. March 17, 2026 Photo by Yongjun Cho

◆ Rebellion attracts first National Growth Fund investment, DeepX and HyperAccel target market segmentation = Rebellion has established another pillar with large-scale investment and a rack-level infrastructure strategy. The company attracted pre-IPO investment of 400 million dollars and unveiled RebelRack and RebelPOD. This strategy is to provide AI inference infrastructure that can be deployed directly in data centers, rather than just individual chips or cards.

In the AI semiconductor market, customers now want more than just the chip itself. They require servers, racks, software that can run AI models, management tools, and compatibility with cloud environments. Rebellion is strengthening its full-stack strategy at this point. Although being the first direct investee of the National Growth Fund is symbolic, ultimately, the key is whether the investment attraction and company valuation can be converted into securing actual customers and recurring revenue.

The progress of DeepX and HyperAccel is also noteworthy. DeepX, which was very active at CES 2026, is targeting the physical AI market, including robots, smart factories, video security, and smart cities.

At a press conference, the company unveiled its strategy as a "physical AI infrastructure company" and announced achievements in securing overseas purchase orders. Robots, cameras, and factory equipment are subject to strict constraints on power, heat, size, and cost, making low-power NPUs advantageous.

HyperAccel is differentiating itself with an LPU dedicated to large language model (LLM) inference. As generative AI services spread, companies feel the burden not only of model development costs but also of service operation costs. This is because as the number of user queries increases, token processing costs and electricity bills accumulate. HyperAccel is targeting the market for reducing LLM inference costs by combining its LPU with full-stack software based on vLLM and PyTorch.

Kim Nokwon, CEO of DeepX (second from the left), is seen discussing and taking a commemorative photo at CES 2026 last January.

◆ Demonstration and overseas markets: building trust is key = The Ministry of Science and ICT supported the market entry of domestic AI semiconductors in the first half of this year through the K-AI Semiconductor Growth Forum, K-Perf verification, and overseas demonstration projects. The Ministry announced that domestic AI semiconductors have entered commercialization this year and have begun to achieve export contracts and demonstration results. Hooneun Bae, Deputy Prime Minister and Minister of Science and ICT, evaluated domestic AI semiconductors as a core foundation for Korea to become one of the world's top three AI powerhouses and to achieve independent AI capabilities.

K-Perf also holds significant meaning. For domestic NPUs to be used in the market, claims by the suppliers alone are not enough. Verification must be conducted under the conditions required by buyer companies, including latency, throughput, number of concurrent users, power efficiency, and stability. K-Perf can serve as a trust-building infrastructure to help domestic AI semiconductors cross the threshold of customer adoption.

The global market is moving in the same direction. Cerebras has demonstrated the enthusiasm for investment in dedicated AI semiconductors by going public with its wafer-scale AI chip.

The inference technology collaboration between Groq and NVIDIA shows that the next battleground for AI semiconductor competition is expanding beyond training GPUs to large-scale inference infrastructure.

This content was produced with the assistance of AI translation services.