That reach now extends across driver assistance, perception, cabin experience, energy management, predictive maintenance, and fleet learning – and it is arriving far faster than the traditional automotive development cycle can absorb. Technology market research company Informa valued the global automotive AI processor market at $5.2 billion in 2025 and expects it to reach $12.6 billion by 2030, a compound annual growth rate (CAGR) of 19.4%.

Each new AI-augmented workload places new demands on the silicon underneath. And those demands are changing far more quickly than the automotive sector has been used to moving. A vehicle program is specified years before start of production, and may stay on the road for decades. Yet the AI we know today is almost unrecognizable from the AI we knew 10 years ago.

When the “Attention Is All You Need” research paper was published by Google engineers in 2017, the AI we knew was fundamentally based on Convolutional Neural Networks (CNNs). Three years later, and in the wake of a global pandemic, everything had changed. The transformer model the paper described was so powerful that it rapidly replaced CNNs, and today almost everything we call ‘AI’ is based on that transformer model. Whatever comes next, the industry expects it to arrive just as quickly and unexpectedly.

This means that silicon decisions taken towards the software-defined vehicle (SDV) today must support at least two generations of AI that have not been invented yet.

Why RISC-V?

RISC-V Enables Pervasive AI in the Software-Defined Vehicle

Today, two broad categories are emerging in automotive AI. Rather than two ends of a compute spectrum, many chips designed for SDVs need to be capable of both – often at the same time.

Physical AI brings intelligence closer to the sensors, actuators, and real-time control loops that govern how the vehicle behaves on the road. Whether serving a zonal controller / hub or a deterministic microcontroller at the edge, physical AI perceives the physical world and acts on it in real time. Its latency budget is measured in milliseconds and it may need to be guaranteed with deterministic timing and certifiable behavior.

Agentic AI draws on human-centric central compute between cockpit and cloud. It acts as an intermediary between a vehicle and its occupants, personalizing the digital cockpit and enriching the in-car or in-cabin experience through workloads such as conversational voice assistants, intelligent navigation, comfort and orchestration layers. Expectations here are being set by consumer electronics rather than by automotive precedent, and they are rising quickly.

While many agentic AI workloads will be non-safety-critical, often referred to as Quality Management (QM), the boundary is porous. Any feature that can act on the vehicle’s physical state carries a functional-safety dimension, and any agentic AI that monitors a driver – be that their physical interactions with the vehicle or simply drooping eyelids – must work closely with physical AI to ensure safe operation of the vehicle.

The Industry Is Embracing the Middle

These capabilities often contribute to the same use case from different angles, increasingly coexisting on the same chip within a mixed-criticality architecture. The models themselves are merging; vision-language-action models integrate perception, natural-language reasoning, and vehicle control in a single policy. A model of this kind may simultaneously drive the most agentic and the most physical workload in an SDV.

Today, the first commercialized SoCs unifying digital cockpit and driver-assistance workloads on one die are shipping, and Tier 1 compute modules are now marketed as serving either domain from one scalable platform. The physical separation between “cockpit chip” and “ADAS chip” is being designed out.

It’s also important to recognize that, as explained above, the field of AI is moving incredibly quickly. As such, the categories of ‘physical AI’ and ‘agentic AI’ should be regarded as a useful simplification rather than industry standard taxonomy. Leading silicon vendors describe these categories as sequential waves rather than opposites, pair them as “embodied and agentic” intelligence delivered across cockpit and driver assistance, and in at least one case argue directly that physical AI is simply agentic AI at the edge.

However you describe it, the important takeaway is that pervasive AI does not mandate the same AI everywhere in the SDV, but a wide range of AI workloads to meet an ever-growing number of use cases. And it’s this that lies at the heart of the opportunity for OEMs and Tier 1s embracing RISC-V for automotive.

AI has already proven its value in central compute and the digital cockpit. Now, as workload-specific silicon makes heterogeneous, AI-native compute the norm throughout the vehicle, the next question is what happens when intelligence proliferates out from the center to touch every processor on board?

Chasing that question has produced a paradigm that the industry is beginning to refer to as the AI-defined vehicle (AIDV). This is less a replacement for the SDV than an acknowledgement of what pervasive AI adds to the original software-defined vision. In effect it is a declaration of intent: moving AI from a bolt-on feature to an architectural component threaded through the vehicle’s nervous system, from its traditional home in central compute out to the edge.

“SDV” took the better part of a decade to settle into common lexicon. “AIDV” seems to have progressed from a handful of silicon-vendor keynotes to trade-press explainers, analyst frameworks, and OEM product language within a year, and it dominated the automotive halls at CES 2026 and Auto China in Beijing.

But again, language is fluid. Look to those using the term ‘AIDV’ and you’ll see it applied to three distinct jobs: 1) a vehicle that adapts to its occupants, 2) an E/E architecture that allocates silicon and thermal budget to AI from the outset, and 3) a development process in which vehicle behavior is trained from fleet data rather than frozen at start of production.

Which definition prevails matters commercially, but matters far less to those designing silicon for pervasive AI. All three converge on the same requirement: more AI, at more locations within the vehicle, with headroom reserved for models that do not exist yet.

A number of RISC-V members are probing where physical AI belongs in safety-critical microcontrollers. If AI is probabilistic by nature, what can it contribute to workloads that demand deterministic outcomes? Answering that means working through safety islands, control and judgment planes, and the boundaries of safety certifications such as ISO 26262.

At the vehicle’s edge sit the microcontrollers and safety islands responsible for deterministic real-time control, monitoring, and fallback. RISC-V privilege levels, memory protection, and deterministic execution support anomaly detection and predictive maintenance within a safety island, while certified control retains final authority over braking, steering, and power delivery.

This is physical AI at its most constrained: inference at the edge crosses into physical AI at the moment it can act on a motor, a brake, or a steering column, and that is precisely the moment that functional safety obligations attach. The same model running as a diagnostic that never actuates anything sits outside that boundary.

The currently agreed design discipline is to let AI contribute, without ever letting it become the single point of authority. AI informs and monitors through anomaly detection, plausibility checks, and predictive maintenance, while a deterministic mechanism retains final say. The prevailing architectural pattern keeps probabilistic AI outside the highest safety boundary and has a deterministic safety island validate its output before any physical action is taken. Agentic AI stays adjacent to certified control loops, not in charge of them.

Standards Have Caught Up With AI

The safety conversation around automotive AI is often oversimplified to ISO 26262. This is no longer the whole picture, and treating it as such understates how far automotive standards work, particularly in AI, has moved in recent years.

ISO 26262 addresses malfunctioning behavior. Something in the electrical or electronic system fails, systematically or randomly, and the system must detect it and fail safely. ISO 21448, ‘safety of the intended functionality’, addresses a different failure: nothing is broken, the system does exactly what it was specified to do, yet the design implementation itself turns out to be insufficiently robust in its operation. Perception systems fail ISO 21448 certification far more often than they fail ISO 26262.

Meanwhile, the newer ISO/PAS 8800:2024 standard covers what the others do not: specific malfunctions and lifecycle factors for AI and machine-learned models used in safety-related road vehicle elements. It addresses functional insufficiencies arising from model performance limitations, training data quality, and the AI development lifecycle.

Together with ISO/SAE 21434 for cybersecurity and UN R155/R156 for type approval, these form the framework any AI-enabled SDV, or AIDV, must be capable of meeting.