Blog

DEEP RESEARCH · DEEPX

DEEPX: A Game Changer for the On-Device AI Era

A private AI fabless review centered on Samsung 2nm GAA, Gaonchips, DX-M1/DX-M2, and the IPO roadmap.

Date: 2026-01-09 · Edge AI semiconductor/foundry ecosystem lens · Naver Blog source and references

Investment decisions are your own responsibility. This material is research and is not a buy or sell recommendation.

0. Bottom line first

DEEPX's thesis is that as AI shifts from cloud training toward edge inference, low-power NPUs and the Samsung Foundry-Gaonchips collaboration model can be leveraged together. But as a private fabless company, cash flow, 2nm yield, and conversion of customer PoCs into mass production are the key tests.

DEEPX on-device AI semiconductor overview image

1. Paradigm: from training GPUs to edge inference NPUs

The source argues that NVIDIA GPUs led LLM training over the past five years, but the market is shifting toward running trained models on real-world devices. On-device AI addresses cloud AI bottlenecks around cost, power, latency, and security.

On-device AI shiftInference where data is created
Cloudtraining, large GPUs, high cost
Edgerobots, cameras, appliances, cars
NPUlow power, low latency, security
DEEPXDX-M1, DX-M2, IP
Performance per watt becomes the central competitive metric.

2. Samsung-Gaonchips-DEEPX triangle

Official fact: The source describes the 2024 DEEPX-Samsung Foundry-Gaonchips agreement as a milestone for Korea's system semiconductor industry. The core plan is to implement DX-M2, a next-generation generative AI chip, on Samsung's 2nm SF2 GAA process.

DEEPX

NPU architecture

Designs AI semiconductors for power, heat, and size constraints at the edge.

Gaonchips

Design solution

Turns logic into physical wafer implementation and links yield, packaging, and test.

Samsung

From 5nm to 2nm

Builds on DX-M1 5nm experience and pushes DX-M2 toward 2nm GAA.

3. Technology moat: 2nm GAA, DX-M2, IQ8

Official fact: The source states Samsung's 2nm process can improve power efficiency by about 25% or more, performance by 12%, and area by 5% versus 5nm.

DEEPX product portfolio and technology image

  • DX-V1/DX-V3: vision-focused line for smart cameras, CCTV, and appliances.
  • DX-M1: 5nm flagship product for robots, smart factories, and drones, presented as 25 TOPS at about 5W.
  • DX-H1: inference accelerator card for edge servers and data centers.
  • DX-M2: 2nm generative AI chip for on-device LLM and multimodal workloads, targeting 2027 mass production.

4. Financials and shareholders

Official fact: DEEPX raised KRW 110bn, about USD 80.5mn, in Series C funding in May 2024. The source cites a post-money valuation of about KRW 529bn and cumulative funding above KRW 150bn.

AreaSource viewInvestor checkpoint
CFOStill negativeTalent, EDA tools, MPW, and R&D costs
CFINegative due to 2nm and overseas expansionCAPEX-like spending for future competitiveness
CFFLarge positive in 2024 from Series CRunway for 2-3 years
IPOTargets KOSDAQ in 2027Samsung Securities selected; tech-special listing possible

5. Competition and growth strategy

Interpretation: If Rebellions and FuriosaAI lean more toward server inference and data centers, DEEPX is positioned around robots, appliances, smart cameras, and AIoT devices where power and price constraints are severe.

  • United States: collaborates with big tech and security companies through a Silicon Valley entity.
  • Greater China and Asia: targets smart-factory and AIoT demand in manufacturing hubs.
  • Customer validation: Hyundai-Kia Robotics Lab, POSCO DX, Jahwa Electronics, and PoCs with over 120 global companies are mentioned.
  • Policy: K-Cloud and the KRW 1tn AI semiconductor innovation/design center and fabless fund are presented as tailwinds.

DEEPX growth strategy and policy image

6. Conclusion and risks

DEEPX is a candidate to move from promising startup to global fabless company in 2026-2028. The bull case is the opening of on-device AI, stable Samsung 2nm yield, and conversion of 120-plus PoCs into mass-production contracts. Risks are post-IPO overhang, dilution from additional funding while losses continue, and Samsung 2nm process delay or yield failure.

Sources