Electronicseng Semiconductors Deep Dive: How Deepx Turns NPUs Into Physical-AI Production Infrastructure

By Rachel Williams, Electronics & Engineering | For questions about this Semiconductors report: info@electronicseng.com

This report for Electronicseng separates three questions that often get merged in edge-AI purchasing: which chip shows the strongest silicon figures, which model stack can survive the move from bench to product, and which supply chain can support a fleet after the first prototype passes. In an EE Times Asia interview, Lokwon Kim, founder and CEO of Deepx, framed physical AI as a platform problem: NPUs are the entry point, but lasting commercial strength depends on workflow, compatibility, production support, and ecosystem reach.

Why Fragmentation Makes Physical AI Different

Cloud inference can often run in comparatively standardized compute environments. Physical AI cannot. It must operate across cameras, robots, vehicles, drones, industrial equipment, medical devices, and endpoints with different power envelopes, thermal limits, sensors, operating systems, and real-time requirements. This fragmentation is not a minor integration detail; it is one of the main reasons many programs stall before broad production.

Deepx describes its shift from low-power AI accelerator development toward a wider physical-AI platform. In that framing, NPUs are connected to existing model frameworks, development platforms, and embedded systems rather than being wrapped in a closed software stack.

The Three-Stage Business Sequence

Lokwon Kim told EE Times Asia: We see the physical AI semiconductor business developing in three stages. First, differentiated hardware performance must earn the customer’s initial selection. Second, a company must expand design wins and mass production, commercialize at scale, and build a defensible position through a strong portfolio of foundational IP. The third stage is to create an ecosystem in which developers can continuously build and deploy AI models and software using tools that are already familiar and broadly adopted, Kim said.

This sequence explains why hardware may secure the first customer choice but does not by itself guarantee lasting commercial strength. For Electronicseng readers, the semiconductor decision begins to look less like a single component selection and more like choosing a production path.

Performance Metrics versus Production Criteria

Evaluation layer	What a benchmark can show	What a production program must also prove
Silicon performance	TOPS and performance per watt remain useful in power-constrained systems.	The processor must fit inside the product physical and thermal boundaries and operate consistently across thousands or millions of units.
Model path	A device may look strong on a reference test.	Customer models must convert and optimize without accuracy loss or extensive software redevelopment.
System economics	The NPU price is easy to see.	Customers review the complete bill of materials, including memory, host processors, power management, cooling, and other components.
Lifecycle support	An evaluation board can demonstrate function.	Production readiness requires dependable supply, manufacturing yield, quality control, long-term software support, and evidence from actual production and field deployment.

Four Major Barriers Between Prototype and Mass Deployment

Deepx points to four major barriers between proof of concept and mass deployment: power and thermal constraints, model compatibility, total system cost, and production readiness. The power problem is especially unforgiving because a product cannot expand its thermal and electrical envelope the way a rack can.

The first is power and thermal constraints. Physical AI products such as cameras, robots, and drones cannot accommodate data-center-class power consumption or cooling systems. If power and heat exceed the physical limits of the product, no other advantage can compensate. This is a non-negotiable constraint, Kim said.

• Model compatibility can become a failure point when conversion or optimization reduces accuracy or forces extensive software redevelopment.
• Total system cost is decided by memory, host processors, power management, cooling, and other components, not only by the NPU.
• Production readiness requires dependable supply, manufacturing yield, quality control, long-term software support, and field evidence, not only a working evaluation board.

A semiconductor vendor therefore cannot stop at supplying a chip and demonstrating performance. It must help the customer run the required models, reduce system-level power and cost, and work alongside them from software integration and pilot production through quality validation and volume manufacturing, Kim said.

Why Openness Is a Competitive Advantage

Historically, edge AI hardware has often been judged through measures such as TOPS and performance per watt. Those figures remain meaningful in power-constrained systems, yet they do not capture every requirement of physical AI. For Deepx, the practical issue is converting hardware performance into products customers can develop and manufacture at scale.

Kim uses Nvidia and CUDA as a reference point. Nvidia differentiation has not rested only on one GPU generation performance; CUDA created a development environment around the hardware and supplied established tools and workflows. Physical AI semiconductors do not yet have an undisputed incumbent with that level of ecosystem dominance. That makes this period particularly important. Whoever succeeds in building an open developer ecosystem and establishing a de facto standard early can shape the market for years to come, Kim said.

Deepx therefore follows an open route by connecting its NPUs to ecosystems developers already use. Named collaborations include Ultralytics YOLO, Baidu PaddlePaddle, and Raspberry Pi, while the company is also working on support for Nvidia Isaac applications and other software environments. The goal is to cut the number of times developers must rebuild software as an AI application shifts from experimentation to production.

Portability Is Now a Design Requirement

For many physical AI programs, proving that a model can run is only the opening step. Production brings engineering constraints. A system must operate consistently across thousands or millions of units, meet a defined power budget, stay within thermal limits, support required models, meet cost targets, and provide predictable long-term supply and software support.

Physical AI projects rarely start on the final production platform. Developers typically begin with a familiar model, a camera, and an accessible development board. Once validated, the system must move to industrial hardware. That move can create substantial engineering work when software environments change from platform to platform.

Deepx and Raspberry Pi aim to reduce that friction by giving developers an accessible starting point and a path to industrial deployment. The intent is to let developers validate vision models and applications on familiar hardware before moving to industrial cameras, robots, gateways, and production systems using the same Deepx execution environment. The distinction is not simply moving an application between two boards. It is about keeping the developer's initial software investment intact throughout the product lifecycle.

From early developer engagement, Deepx identifies three lessons:

1. The entry point must be simple, with documentation, installation procedures, familiar workflows, and technical support carrying major weight.
2. Hardware decisions can be influenced during prototyping, well before production.
3. An easy development platform is not enough unless developers can carry their models and software into industrial validation and production.

Open Stack, Not One Dominant Platform

Deepx does not expect physical AI to settle on one software platform in the way Android became dominant on smartphones. A robot hardware and operating requirements differ from those of a camera, vehicle, drone, or medical device. Kim instead expects convergence around multiple de facto standards at different layers of the stack.

For model development, PyTorch, Ultralytics, and PaddlePaddle are positioned around model creation. For lightweight edge deployment, ExecuTorch can matter. For generative AI, llama.cpp is becoming relevant. For robotics development, ROS 2 and Nvidia Isaac provide important assets.

The central question is not who owns every layer of the stack. It is who can connect different models, frameworks, and hardware platforms to actual silicon and production products with the least friction. Developers will not accept having to rewrite their software for every new processor, Kim said.

The company objective is not to own every software layer. It wants its NPU and execution environment to become a common hardware foundation across several ecosystems.

Distributors as Engineering Support Nodes

As Deepx began mass production of its first commercial product, the company established partnerships with 27 distributors, including Avnet, WPG, Macnica, and DigiKey, as well as regional partners across several markets.

Kim argues that the role of a semiconductor distributor has changed as AI moves into industrial applications. Distribution is no longer limited to inventory, logistics, and sales. Partners increasingly provide technical support, identify customers, engage in engineering, and support design-in activity.

Different applications also require different channels. Robotics, machine vision, factory automation, security, and healthcare have distinct procurement structures and technical-support requirements.

First Markets: Intelligent Vision, Smart Manufacturing, and Robots

Among physical AI applications, intelligent vision appears closest to large-scale commercialization. Security, video surveillance, transportation, retail, factory monitoring, and machine vision already have extensive camera deployments, while AI models used in many applications are comparatively mature.

There is also a relatively direct economic case. AI can automate continuous video monitoring or visual inspection, reduce labor requirements, and identify incidents or defects faster. Companies can often add an edge AI processor without replacing the entire camera or infrastructure.

Smart manufacturing is another near-term opportunity. Asia-Pacific's position within global manufacturing and electronics supply chains, combined with labor shortages and pressure to improve quality, is driving adoption in machine-vision inspection, worker safety, predictive maintenance, and logistics automation.

Robotics may eventually represent the largest physical AI market, but it is also considerably more difficult. A vision system mainly needs to perceive and analyze. A robot must perceive its environment, decide, and execute physical action in real time. An incorrect perception can become an incorrect movement or a safety event. That raises requirements for reliability, model robustness, and system integration.

Kim identifies real-world model reliability, power and thermal constraints, system cost, software compatibility, connectivity, and safety as major remaining barriers. Physical AI models must maintain performance despite changes in lighting, vibration, weather, camera angles, and obstacles.

The resulting architecture cannot depend entirely on cloud connectivity. A robot or industrial machine must keep making immediate decisions when network connectivity is lost, or cloud responses are delayed. Data privacy can also keep sensitive industrial or personal information on the device. For these reasons, on-device inference is an architectural requirement, not merely an alternative to cloud computing.

From Edge Inference to Federated AI Operations

In Kim's view, next stage of physical AI development will require something resembling MLOps for distributed devices. Cloud AI benefited from mature workflows connecting model training, deployment, monitoring, and updates. Physical AI remains much more fragmented, with deployment and operational management often handled independently for individual devices and projects.

That model will not scale when AI is deployed across millions of cameras, robots, and industrial machines. Kim identifies five requirements: secure cloud-device communication; a reliable fleet-wide model; software deployment with rollback mechanisms; simple tools for managing field equipment; and unified monitoring of device status, performance, power, errors, and model accuracy.

Deepx describes the resulting architecture as Federated AI, while distinguishing it from federated learning. Under this model, the data center handles training and centralized management, edge devices perform inference and control where latency, security, and energy efficiency matter, and gateways connect the two where necessary.

The company is working with AWS IoT Greengrass toward an environment where models trained in the data center can be securely deployed to devices, run locally on Deepx NPUs, and managed centrally as a fleet. The point is less about moving all AI out of the cloud than about dividing workloads where computation makes the most sense.

Data-Center Limits and Resource-Efficiency Estimates

Deepx argues that physical AI is also an infrastructure response. Kim points to a widening gap between pace of AI model development and time required to build data centers and energy infrastructure. AI models can advance in months, while new power capacity and large data-center projects take years to build. He argues this mismatch will push more inference to the edge.

Deepx internal analysis compares data center GPU systems with on-device AI systems across power, servers, cooling, and network resources. The company says it found potential for about a 100-fold improvement in resource efficiency in some applications and system configurations.

The company is also examining whether increasingly capable models can operate within low-power devices. In an internal architecture-level simulation involving a 100-billion-parameter mixture-of-experts model, with approximately 10% of parameters active at any one time, Deepx identified the potential for a single 5-W-class AI accelerator to generate more than 10 tokens per second. The company notes that the 5-W figure applies to the chip itself, while total system power varies with memory capacity, configuration, and operating conditions.

Deepx workload research further suggests that models at or below the 100-billion-parameter class could potentially handle more than 80% of inference requests in the physical AI scenarios it is targeting. Under that scenario, the data center would handle the remaining, more complex workloads, potentially allowing existing central infrastructure to serve more users or devices. These figures are not presented as universal benchmarks. They illustrate architectural leverage that becomes possible as more inference moves onto devices.

Free-Token Machines and the Next Computing Layer

Deepx describes this emerging class of devices as Free-Token Machines not because token generation is literally free, but because local inference can occur without a data-center charge for every interaction and without continuous dependence on network connectivity.

Kim sees the transition as analogous to the move from desktop computing to mobile computing. Low-power CPUs moved computation into handheld devices and created a mobility-focused computing model. Deepx expects ultra-low-power AI semiconductors to enable a similar transition for AI, moving inference from centralized data centers into machines and objects operating in physical world.

The concept extends beyond inference. As models become capable enough and power-efficient enough to run locally, cameras, robots, and industrial systems can become agentic machines that perceive their surroundings, make decisions, and act. That does not eliminate data centers. It changes what they are responsible for.

The cloud remains center of frontier-model training and large-scale computation, while local devices absorb workloads where latency, privacy, power efficiency, and connectivity make centralized processing less practical. In this architecture, billions of devices stop being simple endpoints waiting for instructions from centralized infrastructure. They become distributed computing nodes.

Common Misconceptions in NPU Selection

• Thinking that highest TOPS wins is incomplete. TOPS and performance per watt still matter in power-constrained systems, but model conversion, thermal limits, bill-of-materials cost, production readiness, and support decide scale.
• Thinking that edge AI means all AI leaves data center is wrong. The architecture divides workloads where computation makes most sense; training and complex tasks still belong in data centers.
• Thinking that board demo proves production compatibility is risky. A board can validate an idea, but industrial transition requires portability through model deployment, manufacturing, quality, and fleet management.

FAQ for Electronicseng Readers

Why do power and thermal limits dominate many physical-AI decisions?

Because cameras, robots, and drones have fixed physical envelopes. If power and heat exceed product limits, other advantages cannot compensate. The constraint is non-negotiable at the product level.

Does Deepx want to control the full software stack?

No. The stated aim is to connect models, frameworks, hardware platforms, and silicon with least friction, while making the NPU and execution environment a common foundation across several ecosystems.

What is the practical meaning of the 5-W-class estimate?

It refers to the accelerator chip itself, not the complete system. Total system power depends on memory capacity, configuration, and operating conditions.

Why should procurement teams care about distributors?

Because production support increasingly includes technical help, customer identification, engineering engagement, design-in activity, and vertical-specific procurement knowledge.

Action Recommendations for Engineering and Procurement Teams

1. Score NPUs against the four major barriers: power and thermal constraints, model compatibility, total system cost, and production readiness.
2. Map migration path from familiar development hardware to industrial hardware before selection, because redesign effort can erase an early performance lead.
3. Treat distributors as engineering resources and ask for field evidence, quality-validation data, and support models for robotics, machine vision, factory automation, security, and healthcare.
4. Specify fleet operations early: secure cloud-device communication, rollback mechanisms, field-management tools, unified monitoring, power, errors, and model-accuracy visibility.
5. Evaluate vendor support across software integration, pilot production, quality validation, and volume manufacturing, not only initial chip performance.

Final Perspective

Electronicseng reads this shift as a lifecycle-proof problem. The question is no longer only TOPS per watt; it is whether development, deployment, fleet control, and manufacturing can be made repeatable enough that physical AI moves beyond isolated pilots.

Prepared by Rachel Williams for Electronics & Engineering. Contact: info@electronicseng.com.