Autonomous Driving’s AI Shift Forces Automakers to Rebuild

Autonomous driving programs are hitting a ceiling that has little to do with compute or sensors. Parthiv Shah, Senior Vice President of Automated Driving at Mercedes-Benz Research and Development India (MBRDI), argues the primary obstacle is organizational readiness: automotive engineering teams haven’t built the cross-functional, data-centric operating model that AI-driven vehicle development demands. He made the case in an interview with TechCircle, describing a structural transformation already underway at MBRDI, the largest R&D center Mercedes-Benz operates outside Germany, with offices in Bengaluru and Pune employing over 8,500 engineers.

The stakes are concrete. The new Mercedes-Benz CLA, unveiled at CES 2026 with NVIDIA’s complete DRIVE AV software stack, is the first production vehicle on Mercedes’ MB.OS software platform to carry this architecture. Its U.S. launch, expected before December, is the earliest production-scale test of whether that infrastructure performs under real conditions.

The Three Phases of Driving Intelligence

For a decade and a half, Advanced Driver Assistance Systems (ADAS) ran on manually authored logic. Radar data arrived as raw returns, engineers clustered those returns into object representations, and functions like Adaptive Cruise Control (ACC) fired according to predetermined conditions. The software knew what it knew because a human wrote that knowledge into it.

Machine vision changed the perception layer but not the architecture. Camera feeds trained convolutional models to recognize pedestrians, cyclists, lane markings, and traffic signs, a significant improvement but still modular: perception in one pipeline, planning in another, vehicle control in a third. Each component was auditable and correctable in isolation.

End-to-end AI erases those seams. “The shift is being driven by AI-powered driving models that ingest large-scale, multi-sensor datasets and learn driving behaviour in an integrated manner rather than through separate stages,” Shah said. These models absorb camera, radar, and other sensor data simultaneously and produce vehicle commands as a single output. Perceiving, planning, and controlling merge into one learned process.

NVIDIA’s Alpamayo 1, a 10 billion parameter vision-language-action model (VLA) announced alongside the Mercedes-NVIDIA automotive partnership at CES 2026, is a concrete instance of this architecture. The model uses chain-of-thought reasoning to explain each driving decision as it makes it, a property regulators have specifically flagged as necessary for post-accident accountability.

Why AI Models Need 10 to 50 Times More Infrastructure

End-to-end training has an immediate hardware cost. He put the scale plainly: autonomous driving models today are “often 10 to 50 times larger than earlier architectures.” Those models don’t train on a cluster of servers over a weekend. They need sustained access to GPUs, high-performance computing, and the data pipelines to feed them continuously.

• 10 to 50x: size of current AI driving models relative to earlier ADAS architectures, per MBRDI’s Shah
• $79.25 billion global ADAS market value in 2026, up from $68.26 billion in 2025, per Fortune Business Insights
• 42%: projected annual growth rate of automotive AI accelerator chip deployments through 2035, per SNS Insider

“Scale alone is not sufficient, and the data must be curated to ensure it remains context-aware,” he said. A model trained on highway footage won’t generalize to dense urban intersections, and a dataset collected in clear weather will produce unpredictable results in rain. Curation, the systematic work of selecting, labeling, and categorizing data by scenario type, becomes a primary bottleneck.

Scenario analytics, MLOps (machine learning operations, the practice of deploying and maintaining AI models in production), simulation engineering, logging systems, and data traceability are no longer support functions. “This shift also moves complexity from algorithm design to data and platform engineering,” he added. The hiring profile for an autonomous driving program increasingly resembles a cloud software company.

Simulation’s Useful Ceiling

Virtual testing offers a path around the data scarcity problem: generate the edge cases a real-world fleet can’t reliably encounter, such as a child stepping off a curb while looking at a phone or a truck partially blocking a stop sign at a congested intersection. “Virtual validation and simulation play a key role by enabling teams to deliberately generate and amplify edge cases using synthetic data, something that is difficult, time-consuming, and often unsafe to achieve through real-world testing alone,” he said.

Digital twin simulation platforms that form part of MBRDI’s AV development infrastructure let engineers replay a single recorded drive as millions of varied scenario permutations. A data-collection session on a test route becomes a broad training dataset through this multiplied simulation approach.

But synthetic data has a ceiling. Real roads produce textures, lighting conditions, and sensor noise that simulation never fully replicates: the way late-afternoon sun reflects off wet asphalt, the ambient radar interference from roadside construction, the half-open umbrella that confuses a camera in ways no simulation engineer anticipated. “On-road testing remains indispensable for capturing real-world nuances and validating system behaviour in live environments,” he said.

Safety Standards Designed for Predictable Systems

Traditional automotive functional safety rests on a deterministic premise: identify all possible failure modes, design against each one, certify that no plausible malfunction leads to an unsafe outcome. ISO 26262, the primary functional safety standard for road vehicles, governs this process, covering hardware faults, software bugs, and electrical failures, the class of events that cause a system to deviate from its intended behavior.

AI perception systems break that assumption. A camera that correctly identifies 99.97% of pedestrians behaves exactly as designed, yet still misses pedestrians. That gap is purely statistical, inherent to any probabilistic model, and ISO 26262 has no framework to address it.

ISO 21448, known as SOTIF (Safety of the Intended Functionality), published in 2022, was created to fill this gap. SOTIF covers hazards that arise from intended behavior being insufficient for the full complexity of real-world driving, a category of risk that ISO 26262 explicitly excludes.

“Safety engineering for AI embedded in core driving functions is evolving from a requirements-driven, deterministic mindset toward system-level safety assurance,” he said. The practical consequence: safety certification stops being a milestone at the end of a development program and becomes a continuous operational discipline, monitoring deployed models for behavioral drift between training conditions and live ones.

MB.DRIVE ASSIST PRO Reaches Production

At CES 2026 in Las Vegas, MB.DRIVE ASSIST PRO debuted on the all-electric CLA, the first vehicle on MB.OS to carry the complete DRIVE AV stack. The hardware foundation is NVIDIA DRIVE AGX, a compute platform processing data from 10 cameras and five radar sensors in real time, handling perception, sensor fusion, and decision-making simultaneously.

Level 2++ is an unofficial designation for advanced Level 2 systems that go substantially beyond standard driver assistance. At the press of a button, the vehicle manages a complete urban route from parking lot to destination, handling stop signs, traffic lights, and U-turns without driver input while the driver remains ready to intervene. The system launched in China in late 2025, earned Euro NCAP’s highest overall vehicle safety score for 2025, and is set for U.S. production before year-end.

The architecture’s distinguishing feature is the closed loop between training infrastructure and deployed vehicle. Over-the-air updates push new model versions to production cars. DRIVE AV’s closed-loop training approach delivers rapid algorithm iteration through simulation-scale preparation before deployment and continuous updates after, meaning a vehicle on a customer’s driveway six months after purchase runs different software than when it shipped.

The Organizational Gap Most Automakers Haven’t Closed

What Engineering Teams Now Need to Build

Shah identifies organizational readiness as the “most critical and underestimated barrier” as autonomous driving programs move from prototype to production. The challenge is whether a company’s engineering culture can operate the kind of development program that end-to-end AI demands.

Traditional automotive development runs on the V-cycle: requirements flow down one side, verification flows back up the other. The model is sequential, stage-gated, and built for components that don’t change their behavior after deployment. A continuously learning AI system breaks every assumption that structure makes.

The most critical and underestimated barrier is organisational readiness. Without evolving into cross-functional, data-centric and security-aware organisations, even strong technology and infrastructure will not translate into safe, scalable and production-ready autonomous driving solutions.

Speaking to TechCircle, he said competitive advantage going forward comes from an organization’s combined capacity across data pipelines, compute, safety governance, and operational agility. The capabilities a traditional ADAS team didn’t need:

• Scenario analytics: systematic cataloguing of driving situations by frequency, edge-case status, and safety relevance
• MLOps: continuous integration and deployment pipelines for AI models, with production monitoring for model degradation over time
• Synthetic data engineering: simulation pipelines that generate labeled edge cases at industrial scale
• Data traceability: tracking which training examples drove which model behaviors, essential for regulatory and incident accountability
• Runtime safety monitoring: detecting distributional shifts between training conditions and real-world deployment

Where the Gap Gets Most Visible

The pressure is sharpest for programs targeting Level 3 and above, where the driver is legally permitted to stop monitoring the road. The automaker already holds Level 3 approval in Germany and in California and Nevada for its Drive Pilot system on the S-Class, a restricted use case involving controlled highway conditions below a speed threshold. The distance between that scenario and full urban autonomy is precisely where the organizational capabilities he describes, including the MLOps pipelines, scenario coverage databases, and runtime safety monitoring systems, become non-negotiable for any regulator granting approval.

Other automakers face the same constraint. General Motors deployed around 200 test vehicles in California and Michigan in March 2026 to evaluate eyes-off highway assistance, with commercial deployment targeted by 2028. Hyundai Motor Group has committed to building an AI supercomputer specifically for autonomous driving model training and validation. Each program carries the same hidden cost: the engineers, data pipelines, and governance processes described here are as expensive to build as the hardware and considerably harder to scale.

The CLA’s U.S. launch, expected before December 2026, will be the earliest production-scale evidence of whether that infrastructure holds up at scale.