Reinventing autonomous driving in the age of generative AI

As AI-native, end-to-end systems take hold, autonomous driving is becoming a race for compute, software, data, and semiconductors. — McKinsey, June 2026

Executive summary

Autonomous driving is now as much an AI-infrastructure challenge as an automotive-engineering one

Generative AI is accelerating the shift to AI-native, end-to-end (E2E) architectures that learn driving behavior directly from data. Competitive advantage is moving from building a better vehicle to mastering the AI ecosystem around it — across five forces:

Market momentum: demand and investment rise as the ADAS/AD market compounds

Architecture: rule-based systems give way to end-to-end AI (AV 1.0 → AV 2.0)

Computing: in-car compute explodes, and memory bandwidth becomes the new bottleneck

Hyperscale: training and validation make AD economics resemble hyperscale AI

Value chain: hardware and software co-design today, but decouple and consolidate tomorrow

01 · Market momentum

Why the prize is real

Market momentum

Demand is real and the market compounds at roughly 16 percent a year

Consumers are ready

Most Chinese — and about 1 in 4 Western — consumers expect fully driverless cars by 2035

More than 60% would consider using robo-taxis

Roughly half expect ride fares to decline

Experts see Level 2+ dominating the mass market through 2035

The market scales

ADAS software and electronics: around $160 billion by 2035

Roughly 16% annual growth

Software and domain control units take the largest share

Market momentum

Three obstacles still make scaling autonomous driving hard

In a November 2025 survey of more than 40 industry leaders, the top ADAS challenges were:

Safety assurance — 23%

High computational demand for in-vehicle inference — 14%

Regulatory and legal uncertainty — 14%

Software development, safety validation, and large-scale data collection remain the major cost drivers.

02 · From rules to end-to-end

The strategic architectural shift

Architecture

The winning architecture is shifting from hand-coded rules (AV 1.0) to end-to-end AI (AV 2.0)

First generation — rule-based (AV 1.0)

Traditional ADAS: modular pipeline, hand-coded if-then rules, limited AI

Hybrid: end-to-end learning with rule-based safety guardrails, including vision-language-action models

Second generation — AI-native (AV 2.0)

End-to-end: transformer models trained on internet-scale and vehicle data

Learns behavior directly from data and generalizes to unfamiliar situations

Modular and monolithic design philosophies are still competing

Architecture

End-to-end AI generalizes better — but its "black box" nature challenges safety validation

E2E systems adapt to situations engineers never explicitly anticipated, which matters most in dynamic urban driving. The cost is explainability: teams can see what the system does without fully understanding why.

Level 2+ scales fastest — the human still supervises, easing the burden of proof

Level 3/4 with pure E2E needs data-efficient models, large-scale simulation, and regulatory acceptance

Hybrid supervisory layers could accelerate E2E adoption at Level 3 and Level 4

03 · In-car computing

A new era of in-car requirements

Computing

Memory bandwidth — not raw TOPS — is becoming the real bottleneck for end-to-end autonomy

Compute is exploding

ADAS/AD processing chips: ~$5.6B in 2025 to $46B+ by 2035, about 24% a year

Share of automotive semiconductor value: under 6% to 22%

NPUs overtake GPUs as the dominant AI-inference engine

Compute centralizes across ADAS, infotainment, and body control

But the constraint moves to data movement

E2E systems become memory-bound, not compute-bound

High-bandwidth LPDDR DRAM and larger on-chip SRAM become essential

Deterministic, bounded latency favors highly integrated designs

Packaging — monolithic SoC, system-in-package, chiplets — becomes a differentiator

AI demand strains supply: DRAM sales grew ~70% CAGR from 2023 to 2025

04 · The hyperscale shift

Why autonomous driving is becoming an AI-infrastructure challenge

Hyperscale

A strategic divide is forming between "compute-light" and "compute-heavy" players

Compute-light

Many traditional OEMs

Integrate off-the-shelf tech stacks from suppliers

Rely on supplier foundation models; only light vehicle-specific fine-tuning

Keep internal compute modest

Compute-heavy

Robo-taxi operators and vertically integrated OEMs

Build proprietary end-to-end models from scratch

Require enormous dedicated infrastructure

Leaders scaling toward ~90,000 H100-equivalent GPUs

Hyperscale

A hybrid cloud / on-premises model is emerging as the dominant infrastructure architecture

Players weigh cost, data sovereignty, elasticity, and compliance when deciding where compute lives:

Cloud — fastest scaling and flexibility, but costly and lock-in-prone at scale

On-premises — full data sovereignty and predictable economics, but heavy capex and no elasticity

Hybrid — sensitive data and steady-state on private infrastructure, burst training in the cloud (the preferred long-term model)

Cost discipline matters: GPUs can be up to 70% of AI data-center server cost — attacked through algorithmic efficiency, infrastructure optimization, and smart sourcing.

05 · Value-chain reconfiguration

How roles and relationships are changing

Value chain

Hardware and software are co-designed today, but the value chain is set to decouple — and consolidate to a few winners per layer

Extreme latency and safety needs push leaders toward tight hardware–software codesign now; most experts expect eventual decoupling toward independent software layers and open interfaces.

Semiconductor selection shifts downstream — automakers choose software first, chips follow

OEMs vertically integrate, co-developing custom AI silicon and their own AD software

Semiconductor firms move up into software, becoming end-to-end platform providers

New entrants — AI-native start-ups, robotics, cloud players — join as model and infrastructure providers

High costs mean only a few players per stack layer will reach market

"The race for autonomous-driving leadership will increasingly depend on who can build the industry's most powerful AI ecosystem."

McKinsey · Reinventing autonomous driving in the age of generative AI, June 2026