DeepX, a Seoul-based neural processing unit (NPU) maker, is preparing to go public at a valuation near $700 million, marking the first major test of Korea's emerging edge AI chip sector. The company, founded in 2018, has spent eight years quietly building a portfolio of specialized chips designed to run artificial intelligence models directly on devices, from security cameras to industrial robots, without relying on expensive cloud computing resources.

The timing reflects a fundamental shift in how companies are thinking about AI infrastructure. With Nvidia's data-center graphics processing units (GPUs) sold out through 2027 and hyperscalers spending over $650 billion annually on cloud AI buildout, manufacturers are increasingly looking for ways to process AI locally on edge devices, where latency matters, power consumption is critical, and privacy concerns run high.

Why Are Edge AI Chips Suddenly Becoming Valuable?

DeepX's IPO timing coincides with three major structural shifts in the AI hardware market. First, the economics of on-device inference have become impossible to ignore; every watt of computing power saved at the edge represents one fewer expensive cloud GPU purchase. Second, the South Korean government has rolled an "AI Transformation" strategy into its 2026 budget, giving domestic chip startups procurement preference and a public-market tailwind. Third, with Nasdaq-listed Cerebras filing at a $23 billion valuation and major consortiums burning headlines, Korean investors are demanding a domestic AI silicon story they can buy.

The shift reflects a broader industry recognition that cloud-based AI, while powerful, is not always the right solution. Google's aggressive push to integrate its Gemini Nano model into laptops and smartphones signals that on-device AI is becoming a central competitive battleground. Gemini Nano is specifically engineered for efficiency, designed to run directly on device with minimal computational overhead, making it ideal for the constraints of consumer hardware.

What Makes DeepX's Chips Different from Competitors?

DeepX's flagship product, the DX-M1 chip, is rated at 25 trillion operations per second (TOPS) of integer 8-bit throughput with a typical power consumption in the sub-5-watt range, positioning it directly against Nvidia's Jetson Orin Nano and Jetson Orin NX modules. The company claims the DX-M1 achieves up to 95 percent utilization on transformer workloads through a proprietary quantization stack called "DXNN," which compresses full-precision models to lower-precision versions with less than 1 percent accuracy loss on common vision and language benchmarks.

DeepX's broader portfolio spans four distinct power-and-performance bands. The DX-V1 targets sub-1-watt vision applications for battery-powered devices and IP cameras. The DX-V3 is a multi-camera vision processor consuming 1 to 3 watts. The DX-M1 serves as the headline 25 TOPS edge inference engine. The DX-H1 is a server-class part designed for AI gateways and industrial workstations. Across the line, DeepX emphasizes TOPS-per-watt efficiency, with internal benchmarks claiming 5 to 10 times efficiency advantages over GPU-based edge modules at equivalent precision.

How Is DeepX Building Its Customer Base?

DeepX has steadily disclosed a roster of Korean industrial heavyweights as customers or pilot partners since 2024. Hyundai Motor Group confirmed a joint development effort with DeepX on robots powered by generative AI on the same day the IPO news broke, which Korean press read as a deliberate disclosure ahead of the prospectus. LG and POSCO have also been linked to DeepX in industrial-automation pilots, though neither has publicly confirmed production deployment volumes.

The supply-chain story may be even more important than the marquee Korean logos. DeepX has been reported to be working with both Foxconn and Pegatron on edge-AI module integration, relationships that, if they translate into Taiwanese original design manufacturer (ODM) design wins, would give DeepX a path into global brand-name consumer and industrial devices without having to win each original equipment manufacturer (OEM) individually.

Steps to Understanding the Edge AI Chip Market Landscape

• Cloud vs. Edge Economics: Cloud AI dominance is locked in by Nvidia, AMD, and hyperscaler in-house silicon like Google TPU and AWS Trainium, but the edge AI market is fragmented, latency-sensitive, and power-constrained, creating opportunities for specialized chip makers like DeepX.
• Hardware-Software Synergy: A powerful NPU is only as good as the models it can efficiently run; Google's Gemini Nano and similar optimized models extract maximum performance from these NPUs, delivering fluid AI experiences even on devices with constrained power budgets.
• Market Fragmentation Risk: While Google pushes Gemini Nano, other players like Microsoft are developing their own on-device models and frameworks, potentially leading to compatibility issues that require developers to optimize for multiple NPU architectures and software stacks.
• Real-World Use Cases: On-device processing excels in privacy-sensitive tasks, scenarios with intermittent internet access, and applications requiring extremely low latency, such as doctors summarizing patient notes or journalists transcribing in the field with unreliable connectivity.

DeepX's IPO filing reveals a company transitioning from pre-revenue narrative to one with audited revenue, customer logos, and ambitions to scale globally. The company completed what Korean financial press describes as a Series C round in 2025, with cumulative private funding consistently reported in the tens of millions of dollars across equity, strategic investment, and Korean government grants.

"Lokwon Kim has been building chip silicon for nearly a decade with a remarkable discipline about not over-marking. Now that DX-M1 has commercial design wins, the valuation conversation has finally caught up," a Seoul-based venture capital investor told KED Global in March.

Seoul-based VC investor, quoted in KED Global

DeepX is not alone in this race. Rebellions and FuriosaAI, two other Korean edge AI chip startups, are also preparing to test public markets within a 12-month window. Rebellions raised 340 billion Korean won in a pre-IPO round in September 2025 at a 1.9 trillion won post-money valuation, while FuriosaAI commands a similar valuation band to DeepX.

The broader AI chip market is experiencing a fundamental rebalancing. According to TrendForce analysis, custom application-specific integrated circuit (ASIC) shipments from cloud providers are projected to grow 44.6 percent in 2026, while GPU shipments are expected to grow only 16.1 percent, signaling a shift in the AI hardware landscape as hyperscalers increasingly invest in their own silicon.

For investors and industry observers, DeepX's IPO represents a bet that the edge AI market will mature rapidly enough to support a multi-billion-dollar ecosystem of specialized chip makers. The company's success or failure will signal whether Korean startups can compete globally in AI semiconductor design, or whether the market will consolidate around a handful of dominant players like Nvidia and AMD.