DeepX to spend $100 million on 5-watt AI chip for on-device use

A major bet on local AI

DeepX, a company specializing in ultra-low-power artificial intelligence chips, has announced a significant investment to develop its next-generation hardware. As reported by Seoul Economic Daily, the company plans to spend more than $100 million, or roughly 133 billion won, on this project. This funding will cover essential costs such as production engineering and performance verification.

The primary goal of this investment is to create a chip capable of running complex AI models directly on local devices. By doing this, the company hopes to move away from a heavy reliance on remote data centers for computing power. This shift toward on-device AI is intended to reduce the time, cost, and energy required to send data back and forth to the cloud.

The company has already completed the design for its new chip, which it has named the DX-M2. It has now handed over these designs to the foundry business of Samsung Electronics. This partnership ensures that the hardware will be built using some of the most advanced manufacturing techniques currently available in the industry.

Technical specs and manufacturing

The DX-M2 chip is being designed specifically to utilize Samsung Electronics' advanced 2-nanometer process. This manufacturing node is critical for achieving the high performance and low power consumption that DeepX is targeting. The company wants the chip to handle large language models and vision language models while consuming only about 5 watts of power.

Running large models at such a low wattage is a significant technical hurdle for the industry. It requires highly efficient architecture to ensure the chip doesn't overheat or drain batteries too quickly. By focusing on these specific metrics, DeepX aims to make AI more accessible for portable and integrated electronics that cannot be plugged into a constant power source.

The development timeline for the DX-M2 is clearly defined by the company. They expect to secure their first silicon by March 2027, followed by the distribution of evaluation samples to customers in the next quarter. Initial production volumes are scheduled for the fourth quarter of 2027, with full-scale mass production beginning in 2028.

Target markets and applications

DeepX is targeting several specific industries for the deployment of the DX-M2 chip. One of the primary areas of interest includes robotics, where real-time processing is necessary for movement and interaction. The chip's ability to run vision language models makes it particularly useful for machines that need to understand and navigate visual environments.

Beyond robotics, the company is looking at industrial equipment and personal computers. Modern PCs are increasingly being equipped with dedicated AI features that require significant local processing power. By providing a 5-watt solution, DeepX can offer a way for these devices to run sophisticated tasks without needing a constant internet connection or high-end server access.

The move toward on-device AI services is a strategic play to improve the user experience in these sectors. When AI runs locally, it can provide faster responses because there is no latency from cloud communication. It also offers a layer of privacy and cost-efficiency that is highly attractive to both industrial clients and individual consumers.

Broader ecosystem partnerships

While the Samsung partnership focuses on the physical manufacturing of the DX-M2, DeepX is also expanding its reach through other collaborations. The company announced it will work with Intel Korea and a South Korean server maker called KTNF. This trio aims to develop on-premise inference servers that allow corporations to run AI entirely within their own private networks.

In this specific arrangement, DeepX will provide the neural processing units (NPUs) and the technology needed to optimize models. Intel Korea will contribute its central processing unit (CPU) and extended processing unit (XPU) platforms, along with the necessary software. This multi-company approach ensures that the hardware and software are compatible for enterprise-level use.

KTNF is also expected to play a role by linking its capabilities in server design, manufacturing, and system integration to the project. By combining these resources, the partners hope to create a comprehensive solution for companies that want to keep their data secure. This allows businesses to utilize AI without the risks associated with sending sensitive information to public cloud providers.