The industry is changing, and AI adoption is becoming more prominent even in the embedded systems world. AI tools like Claude Code are becoming part of everyday development. Sure, it reshapes the code-writing and review process, but that’s not the only change happening. Spending on embedded systems is rising and is forecast to reach over $700 billion by the end of 2029. 

At the end of 2025, the number of Internet of Things devices stood at 21.1 billion. It is expected to surpass 50 billion by 2034, so the number of devices will more than double within a decade. Keeping that in mind gives great perspective on the future of embedded systems. But before looking that far ahead, let’s focus on five shifts redefining engineering in 2027.

They cover edge and physical AI in embedded systems, agentic workflows in embedded development, and the Zephyr vs FreeRTOS question. The list closes with embedded systems cybersecurity trends driven by the CRA and the challenge of adopting all of them.

Edge AI and physical AI in embedded systems

Probably every developer right now is aware of the memory price increase and the reason behind it: Artificial Intelligence. AI data centers’ demand for high-bandwidth memory (HBM) is taking wafer capacity away from the LPDDR used in embedded devices. Semiconductor manufacturers are launching new chips with the processing power necessary to run AI, and AI in embedded systems is no longer limited to high-end hardware.

Edge AI; WizzDev

Edge AI keeps time-critical decisions on the device

In this environment, edge AI in embedded systems should be on the table when designing the next generation of devices. It provides the ability to process data without sending it to the cloud. A round trip to the cloud can take tens or hundreds of milliseconds, and sometimes it fails because there is no connection at all. In many applications, the delay is acceptable.

In machine control, safety monitoring, or real-time inspection, it is not. Local processing also keeps sensitive data on the device and reduces bandwidth costs. Today, only about 1% of IoT connections include a dedicated edge AI component, and IoT Analytics expects this to reach around 25% by 2035.

Why physical AI needs a deterministic control layer

But it goes even further, and physical AI in embedded systems is emerging. It interacts directly with the physical world: it perceives the environment through sensors, makes decisions, and acts on them through motors and actuators in robotic arms, autonomous vehicles, or drones. AI no longer runs only as software or in the cloud. It now has a direct link to the physical world.

Physical AI; WizzDev

It’s not simply about putting an AI model into a robot, but designing a whole system around it. Motor controllers need predictable timing, and safety mechanisms must work no matter what the model decides. The problem is that AI models aren’t deterministic, so in order to achieve this, we must have an architecture where real-time control and safety stay in the deterministic layer, and the AI layer works on top of it.

Agentic workflows in embedded development

Do you remember long hours spent developing certain features in a device? Trying to make it work and getting the same outcome all over again. And after a couple of hours, you made it through. But it cost you a couple of meetings with clients, dozens of liters of coffee, and a headache. Right now the reality looks completely different. Of course, we didn’t eliminate coffee and meetings with clients because they are what give projects the right direction, but we speed up the development process by using AI tools to create code.

Most engineers still use AI as a chat

It is not surprising that developers are using AI to write code. Based on available data, nearly every engineer is using AI daily, but only about 2% go beyond basic prompting. Most open a chat, ask a question, take the generated code, and go back and forth until it works. That holds even for tools like Claude Code or Codex, which were built to work as agents.

We’re no exception. Here is a real prompt one of our developers used when he needed a quick frontend:

I need you to build me a static frontend (just the front end only) for a chat app in the style of Claude, ChatGPT, etc. First there has to be a login screen (consider one-page vs. multi-page), then the chat itself. It must support uploading files with the prompt, followed by a waiting/loading state.

Claude Code Prompt; WizzDev

The prompt does its job: it is specific, sets the scope, and leaves design decisions to the model. For a side task like a demo UI, this is perfectly fine. But it is still a conversation, not automation. There are no agents, RAG, or MCP servers, and no complex task handed over to AI end to end. The developer stays in the loop for every step: copying, running, checking, and asking again.

How an agentic workflow takes over the whole loop

That approach saves time on one-off tasks. But what about the repetitive, core embedded work, such as writing a driver from a 100-page datasheet and mapping dozens of register bitmasks by hand? This is where an agentic workflow can take over the whole loop. The agent reads the datasheet, generates the register map and driver code, runs the cross-compiler and static analysis, flashes the target or a virtual board, and runs the tests. When something fails, it fixes the code and tries again until the tests pass. The engineer no longer has to copy, compile, and report every error by hand.

Agentic Workflows; WizzDev

Turn proven workflows into reusable AI skills

That’s why agentic workflows are worth a closer look. They let you hand over repetitive, well-defined tasks and get them done faster. This isn’t about giving AI full control or replacing your own work. AI can’t be solely responsible for the final result: the engineer still defines the requirements, reviews the output, and owns the outcome.

AI is a tool, and like any tool, it’s worth using to its full potential. Instead of repeating the same prompts, turn proven workflows into reusable AI skills, such as a driver template, a test procedure or a code review checklist, and share them with your team. That way one developer’s experience becomes the whole team’s standard.

Zephyr vs FreeRTOS and why Zephyr is gaining ground

Zephyr RTOS, an open-source project started by the Linux Foundation 10 years ago, is steadily gaining ground. Continuous improvements and strategic partnerships have turned Zephyr from just another RTOS into a serious choice for a wide variety of use cases ranging from wearables, healthcare, and automotive to industrial IoT. It runs under 8 KB of Flash and 5 KB of RAM, and offers flexibility by supporting over 1,000 boards (such as the nRF9160, ST Nucleo F103RB, or ESP32) and 300+ sensors out of the box.

Zephyr vs FreeRTOS; WizzDev

Its unique contributor base is constantly growing. In 2026, Zephyr has around 440 active monthly contributors and more than 3,500 in total. For an open-source project, that is a solid measure of adoption: more developers learning Zephyr means more drivers, better tooling, and faster fixes. It provides an entire ecosystem: hardware abstraction, drivers, Devicetree configuration, the build system, and testing – all in one place.

Vendor-neutral governance and five-year LTS

The real advantage, though, isn’t only the ecosystem. It’s that Zephyr is truly open-source and vendor-neutral under the Linux Foundation: silicon vendors such as Nordic, NXP, and Intel contribute heavily, but no single company controls it, which reduces the risk of lock-in. Since 2025, LTS releases have been supported for five years – an important factor for products that must receive security updates under the EU Cyber Resilience Act.

Zephyr as the foundation for Matter on the nRF54

Matter is a good example of where Zephyr makes a real difference. Matter devices built on Nordic’s nRF54 chips need the full connectivity stack: Thread, Bluetooth LE for commissioning, and Matter itself.

In the nRF Connect SDK, that stack runs on Zephyr; Nordic’s bare-metal option covers only simple Bluetooth LE applications. In our Matter work on the nRF54, Zephyr allowed us to integrate Thread, BLE, and Matter and define the required Matter endpoints without building the protocol layers from scratch. That leaves time to focus on device logic and features.

Embedded systems cybersecurity trends across the product life cycle

Ten years ago, secure boot, SBOMs (Software Bills of Materials), security updates, and vulnerability tracking were considered good practice. In the European Union, they are now becoming law. The Cyber Resilience Act (CRA) makes security a core requirement for every product with digital elements sold in the EU, not an afterthought or an optional feature.

What the CRA requires and when

The CRA comes into force in stages, and the first stage is already here. Since 11 September 2026, the CRA reporting obligations require manufacturers to report actively exploited vulnerabilities and severe incidents to ENISA and the national CSIRT.

For an exploited vulnerability, that means an early warning within 24 hours, a full notification within 72 hours, and a final report within 14 days after a fix is available. For a severe incident, the final report is due within one month. This applies to products already on the market, not just new ones.

Cyber Resiliance Act; WizzDev

From 11 December 2027, the full set of requirements applies: security by design, an SBOM in the technical documentation, vulnerability handling processes and security updates for at least five years. Non-compliance can cost up to €15 million or 2.5% of global annual turnover.

From voluntary standards to mandatory security

For years, standards like ETSI EN 303 645 existed, but they were voluntary, and many manufacturers simply skipped them. That didn’t make every device insecure, but it left users with no common baseline to trust. The RED cybersecurity requirements, mandatory since 1 August 2025, closed part of that gap, but only for radio equipment.

The CRA changes that for every product with digital elements: security is now part of the design and can’t be skipped if you want to sell a product in the EU. And the scope is broader than it used to be. Products often stay in the field for ten years or more, so teams need to plan for secure OTA updates, cryptography that can be replaced during the product’s life, field monitoring and, increasingly, protection of the AI models running on the device.

How WizzDev helps with CRA compliance

In practice, most CRA gaps we find are not in documentation but in the device itself: no secure boot, no way to roll back a failed update, no list of third-party libraries in the firmware. Fixing these late can mean a PCB redesign, so it pays to check early.

CRA readiness; WizzDev

We don’t stop at the audit. In our clients’ projects, we implement CRA requirements as part of development, from secure boot and update mechanisms to SBOM generation built into CI. Compliance then becomes part of the product, not a last-minute checklist.

December 2027 is closer than it looks, especially if hardware changes are needed.

The real challenge with embedded systems trends 2027 brings is adoption

With more devices each year and new technologies emerging at the same time, adopting them all is becoming a tough task. Edge AI, agents, Zephyr, the CRA, new MCUs, protocols: everything takes time to learn, test, and integrate into projects. At this point, our own capacity is becoming the limiting factor, not the technology itself. How should we deal with it?

Start from your own bottlenecks

The answer is to start, but not chase every trend. Nowadays, it is impossible to be an expert in everything and know every single technology perfectly. While developing new embedded systems, you should definitely focus on solutions that are truly necessary and will deliver good value to your clients. The best way to choose the right technology is by looking at your own bottlenecks.

Embedded Systems Trends; WizzDev

If you spend days on repetitive tasks, consider agentic workflows. Do clients require near-zero latency response times? Think about Edge AI and its capability to process data locally. Are you planning a product launch in the European Union? Definitely look into the CRA, because it’s not optional, but a mandatory part of digital product development.

The embedded systems trends of 2027 should solve real engineering problems, not create new ones. With this in mind, to benefit most from new trends and tools, teams should adopt them in small steps while keeping ongoing work on track.

Conclusion

The embedded systems landscape is evolving, and IoT is evolving with it. New trends create opportunities to write faster, simpler, and better code. That doesn’t mean we should rush to vibe-code everything. Instead, AI should be a tool that helps us write meaningful code and deliver software faster.

Edge AI is essential wherever response time and, above all, privacy matter: data that never leaves the device can’t be intercepted in transit. Zephyr RTOS, with its growing adoption and vendor neutrality, is worth considering for your next project, especially with its built-in Wi-Fi, Bluetooth LE, and Thread stacks and support for over-the-air updates. On top of that, we have the CRA: its reporting obligations already apply, and from December 2027 the full set of requirements becomes mandatory for every product with digital elements sold in the EU.

The coffee and the client meetings aren’t going anywhere. But the hours spent fighting the same bug again and again can shrink if we choose the right tools and adopt them wisely. The trends are already here. The question is which of them will solve your next problem.

Key takeaways

  • The embedded systems market is projected to reach over $700 billion by 2029, with IoT device counts expected to more than double within the next decade.

  • Edge AI is shifting from a niche component to a standard requirement, with 25% of IoT connections expected to include edge AI by 2035 to solve latency and privacy issues.

  • Physical AI requires a hybrid architecture where non-deterministic AI models run on top of deterministic layers to ensure real-time safety and motor control.

  • Embedded development is moving toward agentic workflows that automate the entire cycle—from datasheet register mapping to cross-compilation and automated testing—freeing engineers from repetitive manual work.

  • Zephyr RTOS is outstripping vendor-specific options due to its neutrality, its support for over 1,000 boards, and long-term support (LTS) cycles that align with regulatory requirements.

  • The EU Cyber Resilience Act (CRA) is a mandatory design constraint; full compliance, including SBOMs and five-year security update windows, becomes enforceable on 11 December 2027.

  • Adopting 2027 trends shouldn’t be about chasing every new tool, but identifying specific bottlenecks like development speed, regulatory risks, or hardware-level latency.