Ohad Yaniv, chief executive of Variscite, says the company’s System on Modules have gone into products from more than 7,000 customers in more than 70 countries over 20-plus years, and he argues that the hardest part of turning an AI prototype into a manufacturable machine today is no longer the AI. It is getting the memory.
Yaniv describes a market where neural processing units, or NPUs, now ship inside mainstream processors such as NXP’s i.MX 8M Plus, i.MX 93 and i.MX 95, so customers ask about NPU performance and camera inputs in their first conversation. He also describes a memory shortage that began as AI-driven data-center demand accelerated through 2025 and has left medical, industrial, IoT and robotics buyers struggling to get DRAM and flash. Variscite, the embedded-module maker he leads, says it orders more than 4 million RAM and flash components a year and keeps six to 12 months of buffer stock on critical parts. Below, Yaniv answers questions on what customers build, why they run models on the device instead of in the cloud, and what separates AI machines that last in the field.
1. For readers meeting Variscite for the first time, what does the company make, and where does its technology show up in the real world?
We design and manufacture System on Modules, or SoMs. A SoM is a compact computer board that carries the processor, memory, storage, power management, and often certified Wi-Fi and Bluetooth, ready to plug into a customer’s own carrier board. Instead of spending time and resources designing that core computing layer from scratch, our customers’ product teams build around a SoM that is already tested, certified, and supported, and are free to put their engineering effort into what makes their product different.
In more than 20 years, we have worked with over 7,000 customers in more than 70 countries. Our modules are inside medical devices, industrial controllers, robots, vision systems, agricultural machinery, EV chargers, building management systems, point-of-sale terminals, and more. If a machine has a screen, a camera, or a network connection, or has to make decisions locally, there is a good chance a SoM like ours is inside.
2. What kinds of customers and applications does Variscite focus on today?
Our customers are companies that build professional equipment, from small production runs to tens of thousands of units a year. For them, a SoM offers a proven, production-ready design, lower development cost and risk, and a faster route to market.
Medical, industrial automation, robotics, IoT, energy, agriculture, and retail customers all use our modules, and their core needs are similar: reliable hardware, long product availability, secure software, and direct access to engineers.
Our production is certified to ISO 13485. While this is mandatory for medical device manufacturers, many of our non-medical customers value this for its quality control and full traceability.
The people we work with are mostly engineers and engineering managers. They start with the processor and features, then look closely at support and supply.
3. How has the rise of AI changed what customers ask Variscite to help them build?
One major change is in the processors themselves. Not long ago, adding AI to an embedded product meant adding an external accelerator. Today, processors such as NXP’s i.MX 8M Plus, i.MX 93, and i.MX 95 include a neural processing unit (NPU), so the ability to run a vision or classification model locally is part of the base design. Customers now ask about NPU performance, camera inputs, and image signal processing in the first conversation, alongside the usual questions about interfaces and temperature range.
Standalone accelerators haven’t gone away, though. Dedicated AI/ML chips are still very much in demand, and they integrate easily with our SoMs. They offload computation from the CPU and let customers scale AI performance by adding more accelerators as their needs grow.
These shifts reflect the growing edge-AI trend, as more processing moves onto the device itself. That said, most AI and ML workloads still run in the cloud, and the customers who run them there need reliable, secure connectivity. We deliver that too, with dual Gigabit Ethernet with TSN support, and pre-certified Wi-Fi 6 and Bluetooth.
4. Where are you seeing the strongest demand for AI on devices themselves: robotics, industrial equipment, medical devices, or somewhere less obvious?
We are seeing demand for AI capabilities across the board in robotics, industrial equipment, medical devices, and well beyond. Mobile robots use it to detect obstacles, inspection systems to find defects on a production line, and medical devices to analyze images and sensor data. Building management systems use it to respond to occupancy, EV chargers and battery systems to detect abnormal load patterns, and predictive maintenance systems to flag a motor or pump before it fails. Retail machines recognize products, and surveillance and defense systems identify activity that needs attention. What these devices share is the need to interpret data locally and act on it.
5. Can you share an example of what a Variscite-powered machine can see, understand, or do using AI?
Picture an inspection station on a fast-moving production line. It is built around a Variscite SoM that has a camera connected directly to the processor. The image signal processor receives the images, the NPU runs an object detection model that finds and classifies defects, and a real-time core triggers the reject mechanism. All of that happens on a module smaller than a credit card, and nothing leaves the machine unless an operator needs to see it.
Now picture an autonomous mobile robot moving pallets through a warehouse. The NPU detects people, forklifts, and obstacles while a separate real-time core handles the motor control so the robot can slow, stop, or change direction without waiting on the application cores.
6. Why would a company run AI on a device instead of sending its data to the cloud?
Privacy and security are the main drivers. Take an office access system with a camera that recognizes employees as they come and go. Faces are biometric data, among the most sensitive personal information there is. Running the recognition model and storing the face data on the device itself keeps that data private and out of reach of anyone who intercepts traffic or breaches a cloud service. The same logic applies to medical and industrial customers, who often can’t, or won’t, send images and sensor data off-site for regulatory or confidentiality reasons.
Security matters beyond privacy, too. Every connection is an attack surface, so the less data a device sends out, the less there is to expose.
There are practical reasons as well. A robot or a production-line machine can’t wait on a remote server or stop working when the network drops. And streaming video to the cloud around the clock costs bandwidth and money for the product’s whole life, when usually only the result matters: a defect was found, someone entered a restricted area, or a motor is starting to fail.
Still, it’s rarely all or nothing. Many products are hybrid. Inference runs locally, and the device connects periodically to report results, receive updated models, and install security patches. The real design question is which decisions have to be made on the device.
7. What tends to surprise customers when they try to turn an AI prototype into a product they can manufacture at scale?
Many customers discover that the hardest part of moving to production these days has little to do with AI. A prototype can be built from a handful of boards, but a product needs the same components in volume, for years. That has become harder during the ongoing global component crisis. As AI-driven demand accelerated through 2025, memory manufacturers shifted capacity toward large-volume buyers in the data-center market, leaving sectors such as medical, industrial, IoT, and robotics struggling to get DRAM and flash. The shortage is still with us, and it has spread to the mid-range memory types that many embedded designs rely on, as well as to other components. A product built on components you can’t buy in volume is not really a product.
That makes the SoM supplier’s supply chain critical to the customer’s own product. Variscite orders more than 4 million RAM and flash components a year, which gives us allocation priority with suppliers we have worked with for over 20 years. We qualify multiple sources for critical parts and keep six to 12 months of buffer stock on critical components. We buy only through official distributors, never brokers, which guarantees the quality of the parts. And because we manufacture in-house, we control production scheduling and quality ourselves.
8. Variscite has introduced modules designed for AI-enabled devices. What do those give product teams that they could not easily achieve before?
The biggest benefit is that the NPU works as an integrated part of the SoM right out of the box. On our AI-capable SoMs, the NPU, CPU, image signal processor, and all other components are supported together in our board support package, so they operate as one system from the start. Customers don’t have to spend their own time and resources getting the NPU to work with the rest of the hardware and software before they can begin developing their application. NXP’s eIQ machine learning tools come integrated into our BSPs, our application examples give teams a working starting point, and if they get stuck, they talk directly to our engineers. That gives them a head start on their own application.
Beyond the AI-specific advantages, customers also benefit from all the broader advantages of using a SoM, and a Variscite SoM in particular. Compared with a chip-down design, development is faster because teams start with a pre-tested module, while upfront engineering costs and design risks are reduced. With Variscite, they also gain long-term software support and maintenance, free schematic reviews, and the flexibility to scale performance up or down with pin-compatible SoMs while minimizing carrier board redesign.
9. Looking ahead, what will separate AI-powered machines that succeed in everyday use from others?
As AI becomes commonplace, what will separate successful machines is how well they can be updated and secured over time. AI is heading into almost everything: as NPUs become a standard part of embedded processors, there is room for AI in almost any embedded system, and it will become an ordinary, everyday part of how those systems work.
That changes what a successful product looks like. A device with AI is a device whose software keeps evolving: models are retrained and redeployed, features are added, and vulnerabilities are found and fixed. The machines that succeed will be those designed from the start for long-term operation in the field, with secure over-the-air (OTA) updates, hardware-based security, and a software stack that is continuously maintained. In the EU, the Cyber Resilience Act makes much of this a legal obligation: manufacturers must handle vulnerabilities and provide security updates throughout a product’s support period.
Hardware needs room to evolve as well. As models grow more demanding, a product may need a stronger NPU a few years after launch. Our SoMs come in Pin2Pin families, in which every module is pin-compatible with the others and fits the same carrier board, so a team can move to a more powerful processor or a more cost-efficient one without redesigning its hardware. We support that with product lifecycles of up to 15 years, free software updates and security patches, TPM 2.0 security options across most of our portfolio, and a range of CRA-ready SoMs that help manufacturers prepare their own products for compliance. Together, these let customers keep their AI-powered products secure, up to date, and in production for as long as they are in the field.
About Variscite
Readers looking to learn more about Variscite and its System on Modules can visit www.variscite.com.
Leave a Reply