Everyone remembers the autonomous vehicle hype. Fewer remember the crash that followed. Between 2020 and 2024, the industry torched tens of billions of dollars and most of its credibility: Uber sold its AV unit after a fatal crash, Argo AI collapsed despite $3.6 billion from Ford and VW, Apple killed a decade-long project, GM shut down Cruise after burning through $10 billion. The 2010s promised a revolution. The early 2020s delivered a reckoning.
But something shifted while the headlines moved on.
THE AUTONOMOUS MOMENT
This time, the technology finally delivers.
Let’s start with whatâs already happening.
Fully driverless robotaxi services, which only appeared in 2018, now deliver 700,000 rides per week: 450,000 in the US (led by Waymo) and 250,000 in China (led by Baidu Apollo Go). The global fleet is still small, a few thousand vehicles in geographically limited zones, but broader adoption is emerging in Europe, the UAE, and parts of Asia. Under reasonable regulatory and cost assumptions we can assume one million autonomous vehicles by 2035 and a $100 billion annual service market.
Consumer vehicles are a different story. Tesla has the lead on deployment, millions of cars running its ‘Full Self-Driving’ software. But it still requires hands on the wheel and eyes on the road. Legacy automakers have been more cautious: Mercedes briefly offered true hands-off highway driving in the US before pausing it this month. BMW and Honda have limited systems in Germany and Japan.
This isn’t hype. It’s a technical unlock.
What changed? Four things converged at once: (1) compute power that can actually process massive datasets, (2) new methods for generating training data through simulation, (3) learning-based systems that improve from experience rather than hand-coded rules, and (4) sensor costs that finally fell enough to make unit economics work.
The capital markets noticed. Waymo is in discussions to raise at a $100 billion valuation. Tesla is building purpose-built robotaxis without steering wheels. Pony.ai and WeRide completed US and Hong Kong listings, raising over $3.7 billion combined. Wayve closed a $1 billion Series C and is reportedly in talks for another major round. NVIDIA’s Jensen Huang keeps calling this “the decade of robotics and autonomous vehicles”, and he’s positioning the company like he means it.
The graveyard years are over. Let’s look at where the industry stands.
THE LEVELS THAT MATTER
And the liability shift that defines the market.
Before diving into players and technology, you need to understand how autonomy is classified, and where liability sits. SAE International defines six levels. The first three leave the driver responsible. The last three shift that burden to the machine. That line is where the business model breaks or works.
Level 0: No Automation. The human drives. Radar and cameras handle basic alerts, emergency braking, lane-departure warnings, but nothing more. The car watches; you act.
Level 1: Driver Assistance. The car helps with steering or speed, but not both. Think adaptive cruise control or lane centering. Hands stay on the wheel. By 2024, these features were standard on roughly 90% of new cars in Europe and the US.
Level 2: Partial Automation. The car controls steering and speed. You can take your hands off the wheel briefly, but your eyes stay on the road, you’re still responsible. Level 2+ extends hands-off time but doesn’t change the liability. This is where the consumer market lives: in 2025, Level 2 and 2+ accounted for about 39% and 5% of global new car sales respectively. McKinsey expects Level 2 and 2+ to account for nearly 60% of new car sales by 2035.
Level 3: Conditional Automation. The car drives itself under defined conditions, specific highways, speed limits, weather. You can look away. But here’s the shift: when the system is engaged, liability moves from the driver to the automaker. That’s why almost no one ships it. Mercedes was first in the US, then paused it this month. BMW offers it in Germany only, capped at 60km/h. Honda ran a 100-unit lease experiment in Japan and stopped. The technology exists. The key to unlocking broader adoption of L3 is the proof of safety. Many players are now well advanced in their ability to showcase this from the data and real-life test cases.Â
Level 4: High Automation. The car handles everything within its operating zone. No human needed. If it hits a situation it can’t manage, it stops safely on its own. This is robotaxi territory: Waymo and Zoox in the US, and Baidu Apollo Go, Pony.ai and WeRide in China, all running Level 4 fleets in geofenced areas with HD maps and remote support. The emerging challengers are taking different bets: Wayve is pursuing a map-light, AI-first approach with Nissan, targeting consumer vehicles around 2027. Similarly, Tesla is pushing vision-only for both robotaxis and private cars. Neither has achieved regulated Level 4 status yet, however, Wayve has been approved in the UK for L4 robotaxi trials in late 2026/early 2027, in partnership with Uber. Â
Level 5: Full Automation. The car drives anywhere, in any condition, with no human involvement. No one is close. Major technical, regulatory, and safety hurdles remain. Don’t underwrite to it.
Where the market is heading
For now, consumer vehicles have only been approved for Level 2. The hardware is mature, the regulations are clear, and the liability stays with the driver. Level 3 exists but remains boxed in by geography, speed limits, and legal uncertainty. Itâs clear that the roadmap is being set for L3 in the coming years.
The jump to true autonomy is happening in commercial fleets first. Private vehicles will follow, slowly.
THE STACK
Who builds what, and who captures value.
The autonomous vehicle industry isn’t one market. It’s three layers:
- OEMs build the vehicles, and either develop autonomy in-house or buy it.
- Software providers build the brains: perception, prediction, decision-making. Some license to OEMs; others run their own fleets.
- Hardware providers supply the compute. NVIDIA dominates, but the field is shifting.
Some players span all three. Others specialize. Where you sit in the stack determines your margins, your defensibility, and your path to scale.
1 |Â VERTICAL INTEGRATION
The moat is the stack.
Full-stack control: owning both autonomy software and compute hardware enables tighter co-optimization, faster iteration, and sole ownership of the data flywheel. It also requires massive capital, scarce AI and silicon talent, and years of sustained R&D burn.
Three OEMs have made that bet: Tesla, XPeng, and NIO. Each develops autonomy software and custom compute internally. Everyone else remains partnership-dependent across one or both layers, focusing on vehicle integration while outsourcing the capital-intensive work to suppliers.
On the supplier side, the market has converged toward bundled hardware + software offerings. Two paths lead there:
Hardware-first:Â NVIDIA, Qualcomm, and Horizon Robotics built their positions on compute silicon, then layered autonomous driving software on top to lock in the full stack.
Software-first:Â Mobileye and Huawei lead with perception and planning algorithms, but control proprietary hardware to run them.
The bundling isn’t accidental. Suppliers who own both layers capture more margin, reduce integration risk for OEM customers, and make switching costly. OEMs who depend on them will trade long-term leverage and IP for speed-to-market.
2 |Â THE SOFTWARE DIVIDE
 A convergence toward end-to-end intelligence.
The strategic question isn’t just who builds the system, it’s how the driving intelligence is architected.
The figure below classifies autonomy software providers by their core approach.
Three architectures have emerged: (1) rule-based systems (AV 1.0), (2) learning-driven end-to-end models (AV 2.0), and (3) hybrids that combine both. Each reflects a different trade-off between scalability, interpretability, and safety.
AV 1.0: Rules
This was the first approach used in the early 2010s, requiring significant capital resulting in constrained performance and scalability.
A modular pipeline where each task follows predefined instructions: sensors collect data â perception detects objects â planning applies rules â control executes. Each stage is separate. Humans write the logic.
This works in structured environments. It struggles everywhere else. Edge cases like a plastic bag, an ambiguous hand signal, or an unmarked construction zone require manual fixes, one rule at a time. This is the long-tail problem. Comprehensive coverage at scale is nearly impossible.
The hardware cost reflects the architecture’s rigidity: AV 1.0 systems compensate with sensor redundancy, multiple lidars, radars, cameras, pushing per-vehicle costs above $150,000 in early deployments. Most OEM in-house ADAS (Advanced Driver Assistance Systems) development still follows this approach, which explains why consumer autonomy remains stuck at Level 2. Early adopters were companies such as; Waymo, Cruise and OEMs.
AV 2.0: Learning
A single neural network learns the entire driving task, perception to control, from data. No hand-coded rules. The system improves through experience. Tesla, Wayve and a few others have been at the forefront of this approach, betting that unified neural networks trained on massive driving datasets can replace modular pipelines entirely.
The advantage is scalability. Less manual inputs of each edge case and exponentially more learned data means adaptability to new situations and better performance.
The problem is interpretability. Neural networks generate outputs, but unlike hand-coded rules, they struggle to explain why a particular decision was made. This black-box behavior creates material regulatory and safety hurdles.
That said, meaningful progress is underway. Wayve, for example, is pioneering Lingo, a language-based interface designed to surface and explain the systemâs reasoning in human-interpretable terms. Even so, purists hedge this issue: both Tesla and Wayve deploy safety wrappers, hard constraints that can override the learning system in edge cases or safety-critical scenarios.
Hybrid: Where most commercial players live
Hybrid models split the stack: neural networks handle perception and trajectory prediction; rule-based systems manage planning, localization, and safety validation, with hard overrides for critical moments.
Waymo, Baidu Apollo Go, Pony.ai, WeRide, all moved to hybrid architectures. But it’s not clean. Integrating two paradigms is expensive and still requires engineering-intensive work.
The question is whether hybrids are a permanent architecture, or a bridge to full end-to-end learning.
Three go-to-market strategies.
Software architecture is not the only differentiator among autonomous software providers: go-to-market strategies vary as well and ultimately shape how these companies generate revenue. There are three main approaches: OEM partnerships, partnerships with ride-hail platforms (Uber, Lyft, Grab, Bolt, DiDi), or operating proprietary robotaxi fleets.
3 |Â THE SILICON BOTTLENECK
Compute determines whatâs possible
Autonomous driving is a real-time inference problem with zero margin for error. Perception, prediction, and planning must run on massive neural networks, with millisecond latency and deterministic reliability. The chip makes or breaks the system.
Modern vehicles contain 1,000 to 3,000 chips. Only a handful run autonomy. These central SoCs (Systems on a Chip) integrate CPUs, GPUs, AI accelerators, memory, and safety systems on a single platform. They process sensors, run the models, and execute decisions. The architecture of this chip caps the complexity of what can be deployed, and therefore the level of autonomy achievable.
The hardware market is concentrated and getting more so.
NVIDIA, Mobileye, Qualcomm, Horizon Robotics, and Huawei controlled roughly 69% of the automotive AI SoC in 2025. By 2035, that’s expected to reach 78%. Tesla is the notable exception: building custom silicon in-house to tightly couple hardware, software, and data.
The spec that matters: TOPS.
SoC capability is measured in TOPS, trillions of operations per second. The rough thresholds:
- Level 2:Â 10â50 TOPS
- Level 3:100â300 TOPS
- Level 4:Â 500â1,000+ TOPS
- Level 5:Multi-thousand TOPS (no one is there yet)
The chart below ranks the most powerful single-chip SoCs from leading suppliers. Note: automakers often stack multiple chips to reach higher compute levels.
Where we are today:Â Most cars on the road run nowhere near these thresholds. A Tesla sold today has roughly 100 TOPS, enough for Level 2, arguably adequate for Level 3, but pushing limits for Level 4. Waymo’s robotaxis run custom server-grade hardware with significantly more compute, part of why they cost more per vehicle. NVIDIA’s next-generation chip (DRIVE Thor) delivers 1,000+ TOPS.
The gap is real: most consumer vehicles shipping today don’t have the silicon to run true autonomy, even if the software existed. Hardware remains a binding constraint.
THE REGULATORY MAP
Three models, one race
Regulation is the hidden variable. Technology may converge globally, but the rules won’t, and the rules determine who can deploy, where, and at what cost.
Three regulatory models have emerged: the US (fragmented, market-driven), Europe (harmonized, safety-first), and China (coordinated, state-led). Each creates different constraints and opportunities.
Who’s ahead? China leads on cost and scale. The US leads on revenue-generating deployment. Europe leads on legal frameworks for Level 3, but lags on Level 4 commercialization.
United States
Fast deployment, fragmented rules
The US has no federal framework for autonomous vehicles. Rules are set state by state: Arizona and Texas permit fully driverless operations; California is stricter; most states have no rules at all.
For robotaxi operators, this fragmentation is navigable. Pick permissive states, deploy in friendly cities, expand incrementally. Waymo operates ~2,500 vehicles across five cities (San Francisco, Los Angeles, Phoenix, Austin, Atlanta), completing 450,000+ paid rides weekly. In 2025 alone, Waymo completed 14 million trips, tripling the prior year. But the model is still brutally capital-intensive: each vehicle costs roughly $175,000, unit economics are not yet positive.
For consumer vehicles, the problem is liability. At Level 2, the driver is always responsible, hands on wheel, eyes on road, legally at fault if something goes wrong. At Level 3, the manufacturer assumes liability when the system is engaged. That’s a fundamental shift, and the US offers no framework for it. No federal safety standard, no certification process, no liability protection. Automakers are asked to accept legal responsibility for crashes with no regulatory safe harbor. State-by-state approval doesn’t help, get certified in California, you’re still exposed everywhere else. Europe-based Mercedes tried anyway, won approval in two states, and quit after 14 months: low demand, high costs, unlimited risk. Tesla’s Full Self-Driving remains Level 2. So does every other consumer system sold in America today. No automaker is currently shipping a Level 3 vehicle to US consumers.
Europe
Legal clarity, slow deployment
Europe took the opposite approach: regulation first. UN Regulation 157 defines what a Level 3 system must do, sensors, fail-safes, cybersecurity, data recording. Meet the standard, get type approval. Crucially, when the system is engaged, liability transfers to the manufacturer under a defined legal framework. This is what the US lacks.
Mercedes was first to certify, winning German approval for DRIVE PILOT in late 2021, later expanding to California and Nevada. BMW followed with Personal Pilot L3 in Germany in 2024. Both work on motorways up to 60 km/h. But even with a clear legal pathway, adoption was minimal, three markets for Mercedes, one for BMW. In January 2026, Mercedes paused DRIVE PILOT globally: low demand, high costs, narrow operating domain. The framework exists. The market doesn’t.
Robotaxis lag further behind. No European city has a driverless service at Waymo or Baidu scale, validation thresholds are high, and the economics don’t work.
China
State coordination, cost leadership
China took a third path: state-coordinated deployment. National guidelines set direction; local governments provide testing zones, permits, and subsidies. The result is the fastest scaling anywhere.
Baidu’s Apollo Go operates in 22 cities, completing 250,000+ fully driverless rides per week, half Waymo’s current volume, but growing at 212% year-over-year. Cumulative rides exceed 17 million; the fleet has logged 240 million kilometers. Safety record: one airbag deployment per 10 million km, no major injuries or deaths reported.
The cost structure is what separates China. Baidu’s RT6 robotaxi costs under $30,000, a sixth of Waymo’s vehicle cost. The next generation targets $20,000. In Wuhan, Apollo Go has achieved unit-level profitability, even with fares 30% below Beijing and Shanghai. A few other cities are now unit-positive. Since 2022, China has captured roughly 60% of global AV venture investment.
Now Baidu is going international. Apollo Go has permits in Hong Kong, Dubai, Abu Dhabi, and Switzerland! In 2026, it will launch in London and Germany, partnering with both Uber and Lyft. Chinese autonomy is no longer a domestic story.
Caution on Geopolitics
It is important to note that there may be constraints on who is ultimately approved for robotaxi licenses under geopolitical pressures. Similar to how taxi plates are currently controlled and sold in most countries.
INVESTOR IMPLICATIONS
The market bifurcated while most investors weren’t watching.
The thesis is infrastructure, not vehicles.
The highest-margin positions aren’t in the cars, they’re in the layers every car needs. Compute silicon, simulation, mapping. NVIDIA sells shovels to both sides of the AV divide, and the moat is widening.
Robotaxis are not if, but when.
Waymo proved the technological feasibility with incredible safety numbers and very content consumers.
The next question is unit economics and profitability. Waymo, is hardware intensive, quoted as costing almost $150k per vehicle. Apollo Go run by Baidu in China, is producing L4 capabilities at a fraction of the cost with end to end and cheaper hardware, at just $30k per car.
While still unproven at the L4 level, âhardware lightâ end-to-end models are offering lower cost approach to autonomy in the West. OEMs partnering with end-to-end model providers are hoping to maintain reasonable costs to consumers while providing the potential to run L4 robotaxi services at higher margins.
Consumer autonomy is a liability trade, not a technology trade.
Level 3 is now more of a regulatory issue than a performance one. Europe has the framework; the US doesn’t. Mercedes paused deployment after 14 months. Until legislation comes in or an appropriate insurance model emerges, consumer vehicles stay at Level 2++.
What to watch:
- Wayve à Nissan production (2027): proves licensing model at scale
- Waymo fundraise (H1 2026): validates ~$100B+ valuation
- Baidu London/Germany launch (2026), pending approval
- Tesla robotaxi approval (TBD): Level 4 certification or not
- Wayve x Uber: trial robotaxi rolling out in UK in 2026
The graveyard years killed the tourists. What remains is a smaller, better-capitalized field. Bet on the players gathering the data and building the software layer everyone else will need.
