A Korean semiconductor startup is betting that the real money in AI chips isn't in training massive models, but in running them efficiently at the edge. Mobilint, founded in 2019 by CEO Shin Dong-ju, is ramping up production of neural processing units (NPUs), specialized chips designed to handle AI inference, the process of running already-trained models to make predictions or decisions. The company began mass-producing its first chip, called Aries, in the second half of 2025, and plans to launch a second chip, Regulus, in the second half of 2026.

The distinction matters because it reveals a fundamental shift in how the AI hardware market is dividing labor. While graphics processing units (GPUs) dominate the expensive, power-hungry work of training AI models, NPUs are emerging as the leaner alternative for inference, the task of actually using those trained models in real-world applications.

"Training is being handled by graphics processing units and inference by NPUs, so demand for NPUs will continue to grow," said Shin Dong-ju, CEO of Mobilint.

Shin Dong-ju, CEO of Mobilint

The economics are compelling. Shin explained that hyperscale GPUs cost tens of millions of won (roughly $20,000 to $30,000 USD equivalent), while edge GPUs cost millions of won, and both consume hundreds of watts of power. NPUs, by contrast, require significantly less investment and energy, making them practical for devices that can't afford the footprint of traditional GPU infrastructure.

Why Is Mobilint Focusing on Inference Instead of Training?

The answer lies in where the actual commercial opportunity sits. Data center applications, which handle training and large-scale inference, require massive infrastructure investments to link hundreds of thousands of chips together. That's territory where established players with deep pockets have already staked claims. But edge devices, the smaller computers embedded in cameras, drones, robots, and surveillance systems, represent a largely untapped market where startups can compete.

"For NPUs used in data centers, it is important to make a single chip well, but infrastructure investment, such as linking hundreds of thousands of chips, must accompany it. It is not easy for startups to work on data center applications," Shin explained.

Shin Dong-ju, CEO of Mobilint

Mobilint's flagship product, Aries, targets AI-powered surveillance facilities. Rather than installing AI chips in every single CCTV camera, which would require thousands of devices, the chip can be deployed in a central server room to monitor feeds from multiple cameras. Aries consumes just 25 watts of power, a fraction of what traditional GPU-based systems demand.

The second chip, Regulus, takes a different approach. Designed as an on-device processor, it will be embedded directly into drones, robots, and CCTV cameras themselves, enabling local AI processing without relying on central servers. Samsung Electronics manufactured Aries, while Taiwan Semiconductor Manufacturing Company (TSMC) will produce Regulus.

What Are the Real Challenges in Building NPUs?

Developing NPUs is far more complex than simply shrinking GPU designs. Semiconductor development is often called a "total art" because it requires balancing hardware components with software ecosystems. Even a technically superior chip can fail in the market if developers find it difficult to learn and use the software required to operate it.

Mobilint faces several interconnected challenges:

• Cost Efficiency: Developing an NPU requires investment ranging from tens of billions to hundreds of billions of won, with no guarantee the market will reward the effort.
• Power Efficiency: The chip must deliver strong performance relative to power consumption, a metric critical for edge devices with limited battery life or thermal budgets.
• Compatibility: NPUs must work seamlessly alongside existing chips in real products, supporting operating systems like Windows and Linux without integration headaches.
• Market Timing Risk: Unlike traditional chip manufacturing, where companies receive purchase orders before mass production, NPU makers must produce and sell simultaneously, betting that their chip will remain relevant as the AI market evolves rapidly.

Shin emphasized that stability and compatibility are non-negotiable. "When NPUs are used alongside other existing chips, they must work well in compatibility with other chips," he noted.

Shin

How Is Mobilint Positioning Itself for Growth?

The startup is pursuing a two-pronged strategy: proving the technology works through partnerships and securing funding to scale production. Mobilint has conducted proofs of concept with LG Electronics and Shinsegae I&C, major South Korean companies with real-world deployment needs. In February 2026, the company received 3 billion won in investment through POSCO DX's Corporate Venture Capital New Technology Investment Association, signaling confidence from a major industrial conglomerate.

"We will focus on enhancing AI capabilities while reducing power and cost. Our goal is to improve the strengths of our products," Shin emphasized.

Shin Dong-ju, CEO of Mobilint

The timing aligns with a broader industry trend. As physical AI becomes commercialized, demand for chips that can handle diverse tasks is accelerating. Humanoid robots, autonomous vehicles, smart surveillance systems, and edge devices all need processors that can run AI models locally without constant cloud connectivity. Mobilint's bet is that NPUs, not GPUs, will power this next wave of AI deployment.

The startup's three-year development cycle from chip design to mass production underscores the long lead times in semiconductor manufacturing. Regulus, which entered development years ago, is now approaching the production phase, suggesting Mobilint's earlier bets on edge AI demand are paying off. Whether the company can sustain its advantage as larger chipmakers inevitably enter the NPU market remains an open question, but for now, Mobilint represents a rare example of a Korean startup competing in the high-stakes world of AI semiconductor design.