Korea's AI Semiconductor (NPU) startup FuriosaAI is accelerating full-scale commercialization, setting a goal of achieving 100 billion won in annual revenue this year. FuriosaAI is a fabless (semiconductor design) company that develops AI accelerators (NPUs) specialized for inference operations of Generative AI services, and its strategy is to expand mass production and overseas supply of its second-generation AI inference accelerator RNGD to turn its technology into actual business results through consolidation.
However, the industry is offering a cautious assessment that FuriosaAI and other domestic AI Semiconductor corporations have commonly entered a "death valley." Death valley refers to a period when a startup has succeeded in technology development and product mass production but faces difficulties in customer acquisition, repeat orders, and additional funding. Observers noted that proving chip performance and generating sustained sales in the actual market are entirely different challenges.
◇ FuriosaAI: "Mass-producing 20,000 units this year… targeting 100 billion won in revenue"
At a roundtable with the press on the 5th, FuriosaAI set a goal of producing 20,000 RNGD units this year to achieve 80 billion to 100 billion won in annual revenue. The purchase orders (PO) secured so far total about 20 billion won. The company said it plans to supply 7,000 to 8,000 units from this year's mass production to customers and use the remainder to build its own ecosystem and secure references.
A FuriosaAI representative said, "We secured POs worth 20 billion won just five months after starting mass production," and added, "In the second half, we expect to be able to share more meaningful commercialization results."
It is also expanding its customer base. RNGD has been applied to Samsung SDS Cloud, LG CNS, LG Uplus, and the Daum AI summarization service, and the company recently announced participation in a 15MW AI data center project in Sweden.
The company said, "As AI infrastructure expands faster than semiconductor supply capacity, demand for high-performance, high-efficiency inference solutions is rising rapidly," and explained, "RNGD maintains GPU-level performance while improving power efficiency to reduce total cost of ownership (TCO), which is its competitive edge."
Next-generation product development is also underway. FuriosaAI is working with Broadcom to develop a third-generation AI accelerator using the TSMC 2nm Process and HBM4, aiming to supply samples in the first half of 2028.
A FuriosaAI representative said, "If production capacity (CAPA) had not been secured, it would have been difficult to start development at all," explaining that it has established a collaboration base with key partners such as TSMC and SK hynix. Regarding cooperation with Broadcom, the company said, "It is not a simple IP supply but a form of jointly developing a system-level solution."
◇ "Commercial viability matters more than technology"… repeat orders are the test
Still, the market's view is somewhat cautious. In the AI Semiconductor industry, it is more difficult to get actual customers to make sustained purchases than to develop and mass-produce chips. The PO volume disclosed so far is about 20 billion won, leading to analysis that additional order expansion is needed to meet the annual revenue target the company presented. Given the structure in which it takes about 10 months from production to packaging and delivery to customers, actual revenue recognition can only proceed in stages.
Multiple industry representatives said, "Few people doubt RNGD's performance itself," but added, "From the perspective of corporations, Nvidia's CUDA-based software ecosystem and development environment are already in place, so the incentive to replace them is not yet sufficient."
Another representative said, "Supply through government projects or early pilot programs may be possible, but we need to watch longer to see whether private corporations will voluntarily move to large-scale adoption," adding, "The competitiveness of AI Semiconductor is revealed more in repeat orders than in first deliveries. Only when reorders continue can we say the commercial viability has been proven."
This concern is not unique to FuriosaAI. Domestic Neural Processing Unit (NPU) corporations such as Rebellions, DeepX, and Mobilint have also successively entered the mass production stage but are commonly facing the challenge of commercialization. In the AI Semiconductor market, comprehensive competitiveness is required, ranging from software ecosystems, development tools, customer support, server optimization, and long-term supply capabilities, in addition to hardware performance.
◇ Inference market expansion is an opportunity… ultimately, commercialization results matter
On the other hand, there are also opportunities. With the spread of Generative AI, the inference market is growing faster than AI training, and as data centers face heavier power consumption and operating expense burdens, power efficiency is emerging as a key competitive factor. The industry also offers the outlook that "while it will not be easy to replace Nvidia in the short term, the larger the inference market grows, the greater the role specialized AI accelerators could play for specific workloads."
Meanwhile, FuriosaAI is seeking to raise a total of 800 billion won through the Public Growth Fund. The company said it will soon wrap up a pre-IPO round and invest a significant portion of the secured funds in developing its third-generation product. However, during the investment process, an equity controversy arose as policy funds were approved first while private investment commitments fell short of the target. While the industry agrees on the need for policy support to foster the AI Semiconductor sector, opinions hold that the success or failure of the investment will ultimately be judged by actual commercialization results.
An industry representative said, "It will be difficult to replace Nvidia in the short term, but as the inference market grows, opportunities for high-efficiency NPUs will expand as well," adding, "In the end, what the market judges is not the technology but repeat orders and revenue."