Manufacturing IoT
RISC-V Is Becoming More Important for IoT Hardware
Published by IOT San Jose Research & Editorial Team

- From Cheap Controller to AI Host
- Where Silicon Valley Fits In
- Where the Software Is Catching Up
- What an Open Instruction Set Changes in Practice
- When RISC-V Makes Sense, and When It Does Not
- A Worked Example: A Smart Meter Controller
- Business Value for South Bay Hardware Teams
- Implementation Advice
- Common Challenges
- What Comes Next
Ask two research firms how big the RISC-V market is and you will get two different answers. One values it at 2.3 billion dollars in 2025 and 3 billion in 2026, heading toward 34.7 billion by 2035. Another puts it at 1.35 billion in 2025 and 1.91 billion in 2026, reaching 10.7 billion by 2031. The disagreement runs to nearly a factor of two on the starting line. It is a useful reminder that RISC-V, an open instruction set architecture born at UC Berkeley, is growing quickly enough that analysts cannot agree on how to count it.
The cost argument for RISC-V in IoT is old news: no royalties, freedom to customize. We covered it in our piece on chiplets and RISC-V in IoT hardware. What has changed in 2026 is the argument that was missing before. RISC-V is becoming important to IoT because it is starting to answer the question that matters most this year, which is where the AI goes.
From Cheap Controller to AI Host
For most of its life in IoT, RISC-V was a way to save licensing fees on simple 32-bit controllers. That role has not gone away; 32-bit cores remain the default for cost-sensitive nodes. But edge AI is pulling the architecture upward. Mordor Intelligence lists rapid growth in AI and machine learning workloads at the edge as the first driver of RISC-V demand, and notes that 64-bit designs held the largest share of the market, 41.85 percent, in 2025. Renesas is among the vendors moving premium microcontrollers to 64-bit address spaces.
The mechanism is straightforward. RISC-V's ratified vector extension, RVV 1.0, gives chip designers a standard way to accelerate the math behind neural networks, and its open design lets them add custom instructions for operations a particular product needs. IoT Analytics expects RISC-V adoption to expand during 2026 across low-power IoT edge devices, edge AI processors, and automotive subsystems. ABI Research projects RISC-V processors addressing edge AI devices will reach 129 million shipments by 2030, driven largely by edge AI gateways that connect sensors in the home and run inference on what they hear and see.
That 129 million figure is worth putting in perspective. It is real, and it is small next to the total volume of IoT chips. RISC-V is entering edge AI from a modest base, which means the trend is about direction and share gain, not about replacing existing architectures overnight.
Where Silicon Valley Fits In
The region touches this story at several points. SiFive's X280 RISC-V cores sit beside Kinara's Ara-2 NPU on a board rated up to 40 TOPS, pairing an open instruction set with dedicated AI silicon, which we described in our article on edge AI chips and the Kinara acquisition. Cadence in San Jose supplies the design tools that make custom RISC-V chips practical for companies without a giant verification team. Startups such as Axelera AI and Tenstorrent are using RISC-V to go after heavier inference jobs like computer vision for security and automotive systems.
There is also a policy thread. Market analysts cite national chip-sovereignty programs in the United States, the European Union, and China as a tailwind, because an open architecture is easier to adopt without depending on a single foreign licensor. For South Bay companies selling into regulated or government-adjacent markets, that makes RISC-V a procurement question as well as an engineering one.
Where the Software Is Catching Up
Hardware openness means little without a working software stack, and this is the area that has moved most. Alignment around GCC, LLVM, and the RVA23 profile has given developers a more unified toolchain, and Java and Python runtimes have matured. Real-time operating systems and embedded AI toolchains are following. The practical effect is that a firmware team can now plan a RISC-V product without budgeting for a custom compiler port.
The gap that remains is matrix math. ABI Research notes the community is still waiting for matrix extensions, a key part of AI workloads, to be added to the open standard. Until they arrive, RISC-V designs that need serious inference throughput will pair the cores with a separate NPU, as the Kinara and SiFive board does, or rely on vendor-specific extensions.
What an Open Instruction Set Changes in Practice
An instruction set is the vocabulary a processor understands. With a licensed architecture, the vocabulary and the permitted ways to extend it belong to the licensor. With RISC-V, the base is open and the extension mechanism is built in, so a chip designer can add a custom instruction for a step that dominates a product's workload, such as a filter used in vibration analysis or a hash used in secure boot. If that step runs ten times faster in hardware, the whole device can use a slower clock and a smaller battery.
The trade is responsibility. A custom instruction needs compiler support, test coverage, and documentation, and it creates a dependency the next engineering team must maintain. Companies that benefit most are those with a stable, high-volume product where the payoff repeats millions of times. Companies shipping a few thousand units a year usually gain more from a standard core and a good toolchain than from clever extensions.
When RISC-V Makes Sense, and When It Does Not
| Situation | Better fit | Why |
|---|---|---|
| New cost-sensitive sensor with a simple control loop | RISC-V | No royalty on high-volume units, and mature 32-bit tooling |
| Product needing a custom instruction for a signal-processing step | RISC-V | The open architecture allows extensions that a licensed core would restrict |
| Established product line with a large Arm codebase | Arm | Migration cost can exceed the licensing savings |
| Safety-certified application with pre-validated software libraries | Usually Arm for now | Certified RTOS and library choices are still broader on Arm |
| Edge AI gateway built around a separate NPU | Either | The NPU sets the performance; the host core is a cost and supply decision |
A Worked Example: A Smart Meter Controller
Consider a utility meter maker evaluating a new controller for a product that ships two million units a year. This is an illustrative scenario. If a licensed core carries even a small per-unit royalty, the annual cost reaches six figures and compounds over a ten-year product life. A RISC-V core removes that line, and the team can add a custom instruction for the metrology calculations that run constantly. The catch is that metering products face strict certification, so the team must confirm its operating system, libraries, and test tools are accepted by the certifying body before committing. In this case the savings are real, but only if the certification path is mapped first.
Business Value for South Bay Hardware Teams
The business case has three parts. Cost is the obvious one: for a product shipping millions of units, royalty savings accumulate quickly. Differentiation is the second, since custom extensions let a company build a feature competitors cannot copy by buying the same licensed core. Supply flexibility is the third, because an open instruction set widens the pool of foundries and design partners a company can work with. That last point links to the supply pressures described in our article on mature-node chip capacity.
It is worth being honest about the costs. Choosing RISC-V shifts some work from the licensor to the product team. Someone has to decide which extensions to include, verify them, and keep the software stack consistent across product generations. Mordor Intelligence lists fragmented ISA extensions and weaker backward-compatibility guarantees than Arm as genuine restraints, and notes that some enterprise technology leaders keep Arm roadmaps running in parallel for exactly that reason.
Implementation Advice
- Adopt on new designs first. Starting a new product on RISC-V avoids porting costs. Migrating a shipping product rarely pays back quickly.
- Stay on standard profiles. Targeting RVA23 and ratified extensions keeps your software portable, and custom instructions should be a deliberate exception.
- Pair cores with a proven NPU for heavy inference. Until matrix extensions land, a dedicated accelerator is the safer route.
- Budget for verification. Open does not mean free of validation. Include verification tools and time in the plan.
- Check your toolchain end to end. Confirm your RTOS, your embedded AI tools such as those described in our look at the deployment gap, and your debug stack all support the exact core you pick.
Common Challenges
| Challenge | Description |
|---|---|
| Extension fragmentation | Custom and optional extensions can split the software ecosystem if teams stray from standard profiles |
| Missing matrix extensions | Standardized AI matrix operations are still pending, so inference-heavy designs need extra silicon |
| Certification maturity | Safety-critical software libraries and certified tools are less abundant than on Arm |
| Uneven market data | Analyst forecasts vary widely, so plans should rest on your own unit economics, not a market-size headline |
What Comes Next
Three developments will decide how important RISC-V becomes in IoT. The first is whether matrix extensions are ratified and adopted, which would let RISC-V cores handle more inference without an external accelerator. The second is how many mainstream microcontroller vendors follow Renesas upward into 64-bit designs. The third is whether certified software catches up for safety-critical products, which would open the industrial and automotive segments more fully.
For now, the sensible position is practical rather than ideological. RISC-V has earned a place on the shortlist for every new IoT design, particularly where cost, customization, or supply flexibility drive the decision. It has not yet earned automatic selection. Teams that evaluate it on their own workload, with honest accounting of the verification and software costs, will make better choices than those following either the hype or the skepticism. For more silicon coverage, browse our manufacturing IoT and IoT news sections.
