ItInsightsWorld Startups: What Deepx’s Physical AI Roadmap Tells Early Teams About Ecosystem, Production and Growth
Founders working on AI devices often start by asking whether a faster chip will win customers. The more useful question is whether a startup can take a model from a bench demo to a manufacturable, supportable fleet. That is the lesson embedded in Deepx’s strategy. In an interview with EE Times Asia, founder and CEO Lokwon Kim described the edge AI race as a contest of developer workflow, production handoff and ecosystem reach, not merely a race in TOPS or performance per watt.
Case Snapshot: Deepx’s Path From Chip Design to Platform Work
Deepx started with low-power accelerators and is now presenting a broader platform thesis. Instead of forcing every customer into a closed proprietary stack, it is linking its NPUs to development routes the market already knows: Ultralytics YOLO for vision models, Baidu PaddlePaddle for another major model path, Raspberry Pi for accessible prototyping, and support work connected to Nvidia Isaac for robotics. The Nvidia example is a warning and a template: a hardware advantage becomes durable when the tooling around it becomes hard to leave, as CUDA did for Nvidia’s GPU world.
A three-stage lens for startup teams
Kim’s framework can be read as a roadmap for investor and founder decisions.
“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.
- In the first stage, silicon earns the customer’s attention.
- In the second stage, repeat design wins, production quality and IP build a commercial moat.
- In the third stage, developers can keep shipping without learning a new stack for every processor.
Why the Next Contest Is About Integration, Not Just Compute
Physical AI is not one product category; it is many. Cameras, robots, vehicles, drones, industrial machines and medical equipment have different power ceilings, thermal paths, sensors, operating systems and timing rules. A model that runs in a cloud or a reference environment may fail when it must meet those device-level limits.
Kim says the opportunity is open because no single platform has locked down this segment. “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.
Production Reality Check: From Pilot to Volume
The commercial gap for many hardware startups is not demo quality. It is repeatability. A product must behave predictably across thousands or millions of units while staying inside energy, heat, model, cost, supply and software-support boundaries.
Deepx groups the hard parts into four major barriers: power and thermal constraints, model compatibility, total system cost and production readiness.
| Barrier | What early teams should pressure-test | Startup consequence |
|---|---|---|
| Power and thermal constraints | Whether the accelerator can live inside the product’s actual energy and heat envelope, not just an evaluation kit’s. | “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 | Whether customer models can be converted and optimized without a damaging loss of accuracy or a large rewrite. | Benchmark strength disappears if the production software path is slow, fragile or expensive. |
| Total system cost | The complete bill of materials: memory, host processors, power management, cooling and other components around the NPU. | A device may miss its commercial price target even when the chip itself looks attractive. |
| Production readiness | Supply predictability, manufacturing yield, quality control, long-term software support and field evidence. | Procurement teams care less about a board demo than about repeatable deployment. |
The chip vendor cannot stop at delivery
Deepx’s conclusion is that semiconductor suppliers must act as system partners. “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, Channels and the First Beachhead
For founders, portability is not simply moving code between two boards. It means preserving the initial software investment while the application travels from a prototype camera and development kit into an industrial product, pilot line and long-lived fleet.
Deepx’s Raspberry Pi work is aimed at that transition. The idea is to give teams a familiar entry point, then help them carry validated vision software into industrial cameras, robots, gateways and production systems without rebuilding everything.
The company also distilled its early developer engagement into three lessons: simple entry paths matter, hardware choices can be shaped before production, and accessible tooling alone does not create real portability unless the software can survive industrial validation.
Distribution is another part of the platform strategy. As Deepx entered mass production, it built partnerships with 27 distributors, including Avnet, WPG, Macnica and DigiKey, along with regional partners. For industrial AI, these relationships are not only about inventory; they can carry technical support, customer discovery and design-in work across categories such as robotics, machine vision, factory automation, security and healthcare.
Where a Startup Should Begin: Vision, Manufacturing and Robotics
Intelligent vision looks closest to broad commercialization because the installed base is already there. Security, video surveillance, transportation, retail, factory monitoring and machine vision have many cameras in the field, and many of the required models are relatively mature. The economic case is also direct: continuous inspection can lower labor needs and surface incidents or defects earlier, often without replacing the whole optical system.
Smart manufacturing is another near-term path, especially in Asia-Pacific, where supply-chain density, labor shortages and quality demands create openings for machine-vision inspection, worker safety, predictive maintenance and logistics automation.
Robotics may become a larger physical AI market, but it asks for more. A robot must perceive, decide and act in real time. A mistake in perception can turn into a wrong movement or a safety problem, so model robustness, connectivity, cost and reliability all become stricter.
Data Signals: What Deepx Claims About Efficiency
Deepx has published a set of internal reference points to argue that local inference can change infrastructure economics.
| Reference | Claim | Why a startup should care |
|---|---|---|
| Internal system comparison | The company says its comparison of data center GPU systems and on-device AI systems showed potential for about a 100-fold improvement in resource efficiency in some applications and system configurations. | It supports a workload-splitting story: keep heavy training in the data center, move suitable inference into products. |
| Simulation of a large model | In a 100-billion-parameter mixture-of-experts simulation, 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. | It gives founders a way to think about local generative AI within device power limits. |
| Power caveat | The 5-W figure applies to the chip itself, while total system power varies with memory capacity, configuration and operating conditions. | Do not quote chip-only power as system power in product roadmaps or investor materials. |
| Workload coverage | Models at or below the 100-billion-parameter class could potentially handle more than 80% of inference requests in targeted physical AI scenarios. | A startup may not need frontier-size models running on every endpoint to make edge products viable. |
In that architecture, billions of devices become computing nodes rather than passive endpoints waiting for cloud answers.
The Ecosystem Thesis: Many Standards, Less Friction
Deepx does not expect physical AI to converge on one winner the way Android became dominant in smartphones. A robot, camera, vehicle, drone and medical device do not share the same runtime assumptions. Kim instead expects several de facto standards at different layers.
PyTorch, Ultralytics and PaddlePaddle sit close to model development. ExecuTorch can help with lightweight edge deployment. llama.cpp is becoming relevant for generative AI on constrained systems. ROS 2 and Nvidia Isaac provide robotics building blocks.
Kim framed the competitive test around connection, not ownership. “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.
Federated AI and Operational Scale
The next startup problem is operational: once devices are deployed, how do they update, monitor and survive failures? Cloud AI has mature training-to-deployment workflows, while physical AI remains scattered across custom device and project stacks. Kim lists five requirements: secure cloud-device communication, reliable fleet-wide model software, deployment with rollback, simple management of 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. The data center handles training and central management; edge devices handle inference and control when latency, security and energy efficiency matter; gateways connect the two layers. The company is working with AWS IoT Greengrass toward a path where models trained centrally can be deployed securely, run locally on Deepx NPUs and be managed across fleets.
From inference to agency
Deepx uses the label “Free-Token Machines” for systems where local inference can avoid a per-interaction charge on centralized infrastructure and reduce constant network dependence. The deeper shift is agentic: if a camera, robot or industrial system can perceive, decide and act locally, it becomes part of the computing fabric rather than a simple terminal.
Common Misconceptions for Physical AI Startups
- “Higher TOPS wins the market.” Compute matters, but integration, tooling and production support decide whether a product reaches scale.
- “A good proof of concept is a product.” The same behavior must survive across thousands or millions of units, with fixed power, thermal, cost and supply constraints.
- “Portability means code can run on two boards.” The real requirement is surviving from prototype to industrial validation, production and long-term maintenance.
- “Distributors only ship inventory.” In industrial AI, channel work often includes engineering assistance and design-in; Deepx’s 27 distributors are part of its platform thesis.
- “Edge means abandoning the cloud.” The better frame is workload division: training and heavy tasks in the data center, time- and privacy-sensitive tasks on the device.
Action Suggestions for Founders, PMs and Corporate Innovation Teams
- Begin with the developer’s first hour: documentation, installation, familiar model formats, sample apps and responsive technical support.
- Choose a beachhead where existing hardware and mature models shorten the sales cycle, such as intelligent vision or factory inspection.
- Turn the four major barriers into a supplier checklist: power and thermal fit, model compatibility, total system cost and production readiness.
- Plan portability across the lifecycle, not just across evaluation kits.
- Build channel partnerships early as technical and discovery assets, not only logistics assets.
- Design fleet operations before scale: update safety, rollback, monitoring, error tracking and model performance visibility.
- Keep cloud training centralized, but push suitable inference to devices where latency, privacy, power or connectivity make cloud processing awkward.
FAQ
Why should a startup care about an open physical AI stack?
Because buyers usually do not want to rewrite software for each new processor. An open path can reduce integration risk, speed design-in and make the first product-to-production handoff less painful.
Is robotics the best first market?
Not necessarily. Vision systems may commercialize sooner because installed cameras, mature models and direct inspection economics make adoption easier. Robotics is potentially larger but carries stricter real-time, safety and reliability demands.
What should investors ask instead of “Which chip has more TOPS?”
Ask how customers move from benchmark to production: model conversion quality, total system cost, power envelope, channel support, design-win conversion, supply predictability and software maintenance.
What is Deepx’s main thesis?
Hardware opens the deal; ecosystem and production capability determine the durable commercial position. The company is trying to make that handoff smoother across cameras, robots, industrial devices and edge fleets.
For the Startups reader, the takeaway is practical: the most promising physical AI ventures may not be those with the flashiest accelerator, but those that make deployment repeatable. The future belongs to teams that can turn on-device intelligence into a process, not a one-off experiment.