Physical AI After the Compute Race: Deepx, Open Workflows and the Production Test for AIdataNewsGlobal General

AIdataNewsGlobal briefing. The central challenge for physical AI is no longer simply whether an accelerator can post competitive numbers; it is whether developers can carry a validated model through procurement, integration, manufacturing and long-term operations without repeated rewrites. In an interview with EE Times Asia, Lokwon Kim, founder and CEO of Deepx, framed that challenge as the boundary between early hardware wins and durable deployment across cameras, robots, vehicles, drones and industrial equipment.

Kim's argument does not treat silicon as irrelevant. Hardware remains the entry point. The more durable question is whether software compatibility, ecosystem access and production support convert performance into large-scale commercial use.

Deepx frames its plan around that separation. The company is moving from low-power AI accelerator development toward a broader platform strategy, connecting its NPUs to existing model frameworks, development platforms and embedded systems rather than building a wholly closed software stack.

Key Takeaways

• Physical AI spans heterogeneous devices with different power, thermal, sensor, operating-system and real-time requirements.
• Deepx describes the semiconductor business as developing through three stages, from initial customer selection to design wins and IP, then to developer ecosystem adoption.
• The firm identifies four major barriers between proof of concept and mass deployment, including power and thermal limits, model compatibility, total system cost and production readiness.
• Early developer engagement produced three lessons about simple entry points, prototype-stage influence and the need to move models into industrial validation.
• Deepx says its first commercial product shipment is supported by partnerships with 27 distributors, whose role now includes engineering and design-in support.
• Intelligent vision looks closest to large-scale commercialization because of installed camera bases and comparatively mature models; robotics may become larger but is harder.
• Kim argues that distributed fleets will need operations discipline, a model Deepx labels Federated AI, while data centers remain central to training and heavier workloads.

Why Fragmentation Changes the Deployment Problem

Cloud AI can often be executed in comparatively standardized compute environments. Physical AI cannot. A camera, a robot, a vehicle, a drone, a medical device and an industrial controller may differ in power envelope, thermal design, sensors, software stack and real-time expectations. That heterogeneity makes the path from demonstration to production more difficult than a single benchmark table suggests.

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.

Why TOPS Is Not the Whole Business Case

TOPS and performance per watt remain relevant, especially in power-constrained products. But they do not describe the full production process. For a customer, the decisive step is converting accelerator performance into a manufacturable system that can be developed, certified, supplied and supported across thousands or millions of units.

Nvidia is cited as an example of how hardware advantage can be amplified by ecosystem infrastructure. CUDA created a development environment around the hardware, giving researchers and engineers established tools and workflows. In Deepx's reading, physical AI still lacks an undisputed incumbent with comparable ecosystem dominance. That creates an opening, but also raises the stakes for whoever can build early developer trust.

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.

For that reason, Deepx follows an open route, connecting its NPUs to ecosystems developers already use. The company points to work involving Ultralytics YOLO, Baidu PaddlePaddle and Raspberry Pi, along with effort to support Nvidia Isaac applications and other software environments. The goal is simple: cut the number of times developers must rebuild software as an AI application shifts from experimentation to production.

From Demonstration to Production: The Four Barriers

For many physical AI programs, the first question is whether a model can run. The next set of questions is harder: whether the system can operate consistently across thousands or millions of units, meet a defined power budget, stay within thermal limits, support required models, hit cost targets and provide predictable supply and software support.

Deepx sees four major barriers between proof of concept and mass deployment: power and thermal constraints, model compatibility, total system cost and production readiness.

1. Power and thermal constraints

Physical AI products cannot usually accept data-center-class processors and cooling systems. The accelerator must fit inside the product's physical and thermal boundaries.

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,

2. Model compatibility

Benchmark performance can be lost during conversion, optimization or adaptation. If accuracy drops or a customer must redevelop software extensively, the hardware advantage may disappear.

3. Total system cost

Buyers evaluate the complete bill of materials, not the NPU component alone. Memory, host processors, power management, cooling and other parts determine whether an AI-enabled product can meet a commercial target.

4. Production readiness

A working evaluation board does not prove manufacturability. Customers look for dependable supply, manufacturing yield, quality control, long-term software support and evidence from actual production and field deployment.

This changes what a semiconductor vendor is expected to provide.

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.

Portability as a Product Requirement

Software portability becomes central because physical AI projects rarely begin on the final production platform. Developers often start with a familiar model, a camera and an accessible development board. After validation, the workload must move to industrial hardware.

That transfer can create substantial engineering work when tools, drivers and execution environments change between two boards. It is not merely about moving an application from one development kit to another; it is about preserving the developer's initial software investment across the product lifecycle.

Deepx's collaboration with Raspberry Pi is designed around this sequence. The idea is to give developers an accessible starting point and then a path to industrial deployment, allowing vision models and applications to be validated on familiar hardware before being carried into industrial cameras, robots, gateways and production systems using the same Deepx execution environment.

The company says early developer engagement revealed three lessons.

1. The entry point must be simple: documentation, installation procedures, familiar workflows and technical support carry major weight.
2. Hardware decisions can be influenced during prototyping, well before production commitments are made.
3. An easy development platform is insufficient unless developers can carry their models and software into industrial validation and production.

Open Ecosystems, Not a Single Winner

Deepx does not expect physical AI to consolidate around one software platform in the way Android became a dominant smartphone operating system. A robot's requirements are fundamentally different from a camera, vehicle, drone or medical device. Kim expects convergence around multiple de facto standards at different layers.

PyTorch, Ultralytics and PaddlePaddle are positioned around model development. ExecuTorch can support lightweight edge deployment. llama.cpp is becoming relevant for generative AI. ROS 2 and Nvidia Isaac provide important assets for robotics development. The company points to work involving Ultralytics YOLO, Baidu PaddlePaddle and Raspberry Pi, along with effort to support Nvidia Isaac applications and other software environments.

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. For the company, the aim is not to own every layer of software. It is to make its NPU and execution environment a common hardware foundation across several ecosystems.

Distribution Moves Into Engineering

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

Kim argues that AI's movement into industrial applications has changed the role of distribution. Inventory, logistics and sales remain necessary, but partners increasingly provide technical support, identify customers, engage in engineering work 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.

Intelligent Vision and Robotics: Where Scale Arrives First

Among physical AI applications, intelligent vision appears closest to large-scale commercialization. Installed base is one reason: security, video surveillance, transportation, retail, factory monitoring and machine vision already have extensive camera deployments, and many relevant AI models are comparatively mature.

The economic case can also be direct. AI can automate continuous video monitoring or visual inspection, reduce labor requirements and identify incidents or defects faster. In many settings, companies can 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 become the largest physical AI market, but it is also considerably harder. 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 therefore become an incorrect movement or a safety event, raising 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 the major remaining barriers. Physical AI models must maintain performance despite changes in lighting, vibration, weather, camera angles and obstacles.

That architecture cannot depend entirely on cloud connectivity. A robot or industrial machine must continue 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.

Federated AI and Device Operations

In Kim's view, the next stage of physical AI 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 lists five operational 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 uses the label Federated AI, while distinguishing it from federated learning.

Under that framing, the data center remains responsible for training and centralized management, while edge devices handle inference and control where latency, security and energy efficiency matter; gateways connect the two layers when needed.

The company is collaborating with AWS IoT Greengrass to explore a setup in which data-center-trained models are securely shipped to devices, execute locally on Deepx NPUs and are administered as a fleet.

The larger shift is workload placement. The point is not to move all AI out of the cloud, but to divide work where computation makes the most sense.

Infrastructure Limits and the Edge Inference Case

That division matters more as AI infrastructure hits physical constraints. Kim points to a widening gap between the pace of AI model development and the 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's 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.

It 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's 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 the architectural leverage that becomes possible as more inference moves onto devices.

Data Points Worth Watching

Reference point from Deepx	How it should be read
Partnerships with 27 distributors for the first commercial product	Channel support is treated as part of deployment, not merely fulfillment.
Potential for about a 100-fold improvement in resource efficiency in some applications	A company analysis comparing data center GPU systems with on-device AI systems across power, servers, cooling and network resources.
Internal simulation involving a 100-billion-parameter mixture-of-experts model	Examines whether large models can be compressed through sparse activation into low-power devices.
Simulation assumption with approximately 10% of parameters active at any one time	The scenario is internal and architecture-level, not a claim that every model behaves identically.
Potential for a single 5-W-class AI accelerator to generate more than 10 tokens per second	An architecture-level potential identified by Deepx, not a universal benchmark.
The note that the 5-W figure applies to the chip itself	Total system power varies with memory capacity, configuration and operating conditions.
Models at or below the 100-billion-parameter class could handle more than 80% of inference requests in targeted physical AI scenarios	Suggests a workload split where data centers continue to handle remaining, more complex tasks.
The company frames billions of devices as distributed computing nodes	Reflects the broader vision: cameras, robots, vehicles, drones, industrial equipment and personal devices perform more local inference.

The Next Computing Layer

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

The company describes this emerging class of devices as Free-Token Machines. The label does not mean token generation is literally free. It means local inference can occur without a data-center charge for every interaction and without continuous dependence on network connectivity.

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 the 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.

FAQ

Why does the article focus on ecosystem tools rather than chip specifications?

Chip performance is necessary but not sufficient. The interview's central claim is that hardware may secure the first customer choice, while software compatibility, developer tools, portability, production support and long-term operations determine whether a product can scale.

What does Deepx mean by physical AI?

AI running in devices and machines that operate in the real world, such as cameras, robots, vehicles, drones, industrial equipment and other systems. These environments differ from cloud AI because they impose power, thermal, real-time, safety and cost constraints at the device or system level.

Are the efficiency and token-generation figures universal benchmarks?

No. They are presented as internal analysis and architecture-level simulations, including the potential for about a 100-fold improvement, the 100-billion-parameter mixture-of-experts scenario with approximately 10% active parameters, the 5-W-class accelerator estimate of more than 10 tokens per second, and the note that the 5-W figure applies to the chip itself. They illustrate architectural possibilities rather than guaranteed product results.

Which ecosystems does Deepx say it connects to?

The company references collaborations or compatibility work involving Ultralytics YOLO, Baidu PaddlePaddle, Raspberry Pi, Nvidia Isaac applications and other software environments. It also discusses PyTorch, Ultralytics, PaddlePaddle, ExecuTorch, llama.cpp and ROS 2 as important parts of the broader developer stack.

Why is intelligent vision highlighted before robotics?

Intelligent vision benefits from large installed camera bases, comparatively mature models and relatively direct economics. Robotics may eventually represent a larger physical AI market, but it adds perception-to-action safety requirements and more demanding integration.

What to Watch Next

• Whether open NPU paths can reduce the number of software rewrites between experimentation and production.
• Whether distributor networks can deliver technical support, not merely logistics, across robotics, machine vision, factory automation, security and healthcare.
• Whether intelligent vision converts installed camera bases into repeatable edge AI revenue faster than robotics.
• Whether distributed device operations, including rollback, monitoring and secure updates, become a competitive requirement for physical AI vendors.
• Whether edge inference can absorb a majority of targeted requests while data centers continue to train frontier models and handle more complex workloads.

Bottom Line

For physical AI, the decisive competition is moving from isolated performance claims to full production journeys. The question is not only which NPU delivers the most TOPS per watt, but whether an AI semiconductor platform can make model development, prototyping, deployment, fleet management and manufacturing simple enough for physical AI to become repeatable at scale.

By David Nakamura, AI & Data News. Contact: editorial@aidatanews.com.