End-to-end driving models Market Outlook 2026–2034: Growth Drivers, Key Companies, Strategic Developments & Adoption Forecast — A Meticulous Next™ Foresight Brief
What This Brief Covers
This Meticulous Next™ brief examines end-to-end driving models, neural networks that turn camera and sensor input directly into driving actions and replace the separate, hand-written modules for perception, prediction, planning and control. It tests whether single neural stacks become the default driving software for assisted and automated vehicles by 2034, and what that means for automakers, robotaxi operators, chip suppliers and software developers.
Investors have backed the approach at scale. Wayve announced a USD 1.2 billion round on 24 February 2026 at a valuation of USD 8.6 billion, led by Eclipse, Balderton and SoftBank Vision Fund 2, with Microsoft, NVIDIA and Uber returning and Mercedes-Benz, Nissan and Stellantis participating, with a further USD 300 million possible from Uber if Wayve deploys robotaxis, starting in London. Wayve sells its driving software to automakers and does not build vehicles, and Nissan plans to upgrade its driver assistance with Wayve software from 2027. The conclusion is that a licensing model for a neural driving stack has the backing of both automakers and a ride-hailing network.
Chip suppliers are opening the technology. NVIDIA announced the Alpamayo family on 5 January 2026 with a 10 billion parameter reasoning vision-language-action model, an open simulation framework and more than 1,700 hours of driving data, and on 4 August 2026 cleared the 34 billion parameter Alpamayo 2 Super for commercial use, with about 400,000 downloads of the family reported since launch. The full model is meant to be distilled into smaller versions for in-car chips. The conclusion is that a capable driving model is available to any automaker, which lowers the barrier to building a neural stack and raises the pressure on firms that sell closed systems.
China is moving in the same direction. XPeng began rolling out VLA 2.0 in March 2026 on the P7, G7 and X9 in Ultra trim, running on its Turing chip at up to 2,250 TOPS, and Volkswagen was reported as the first external customer. XPeng describes a single vision-to-action model in place of separate perception, planning and control modules. The conclusion is that the neural stack is reaching production cars in the largest vehicle market and is being sold to a foreign automaker.
The brief treats end-to-end driving models as a high-impact signal with an eight year horizon to 2034. It identifies which automakers, software developers and chip suppliers capture value as single neural stacks replace rule-based autonomy. It is a focused 110-page decision brief that presents an indicative trajectory rather than a segmented market model.
| Parameter | Details |
|---|---|
| Forward horizon | 2026–2034 (8 years) |
| Emerging force | Vision-language-action driving |
| Technology readiness | In production in assisted driving at several automakers; open models available for developers; fully driverless use limited to defined areas |
| Indicative market size & forecast | The end-to-end driving models market, covering neural driving software that maps sensor input to driving actions, including licences, per-vehicle fees, subscriptions and the training data and simulation tools dedicated to these models, for assisted and automated vehicles, and excluding rule-based driver assistance software, vehicle hardware and robotaxi ride revenue, is valued at USD 1.20–2.20 billion in 2026 and is projected to reach USD 13.24–24.27 billion by 2034, at a CAGR of 35% |
| Mainstream inflection | ~2029, when most new assisted driving systems from major automakers use a single neural stack and licences are sold across several automaker brands |
| Signal strength | A USD 1.2 billion Wayve round, commercial release of a 34 billion parameter open driving model and production rollout of VLA 2.0 in China |
| Primary beneficiaries | Automakers; software developers; chip and compute suppliers; robotaxi operators |
| Brief length / format | 110 pages · PDF + executive summary deck · instant delivery |
Understanding the Technology
A conventional driving stack passes information through separate modules, each written and tuned by engineers. An end-to-end model learns the whole mapping from sensors to actions from driving data. XPeng describes its system as replacing separate perception, planning and control with one model. The conclusion is that improvement comes mainly from more data and compute instead of more hand-written rules, which changes the cost structure of development.
Vision-language-action models add reasoning. NVIDIA describes Alpamayo as a chain-of-thought reasoning model that outputs a reasoning trace, a trajectory and a plain-language explanation, using video from up to seven cameras. The conclusion is that the same model can explain its decision, which matters for safety assessment and regulators.
Size and compute are the main constraint in the car. The 34 billion parameter Alpamayo 2 Super needs tens of gigabytes of GPU memory and is meant to be distilled into smaller models for in-car chips. XPeng runs its model on a chip of up to 2,250 TOPS. The conclusion is that large models are used for training and teaching smaller ones, and chip performance in the vehicle sets what the car can run.
Simulation and data close the loop. NVIDIA released an open simulation framework and a driving dataset with its first model, and has open-sourced the auto-labelling pipeline for the second. The conclusion is that data and simulation tools become a market of their own, and open releases lower the entry cost for smaller developers.
The limits are the difficulty of proving safety for a model that learns from data, the compute needed in the vehicle, the cost of collecting rare-event data and regulatory acceptance of explanations from the model. The brief tests each limit against the evidence to 2034.
Market Outlook
The end-to-end driving models market, covering neural driving software that maps sensor input to driving actions, including licences, per-vehicle fees, subscriptions and the training data and simulation tools dedicated to these models, for assisted and automated vehicles, and excluding rule-based driver assistance software, vehicle hardware and robotaxi ride revenue, is valued at USD 1.20–2.20 billion in 2026 and is projected to reach USD 13.24–24.27 billion by 2034, at a CAGR of 35%.
The 2026 base rests on assisted driving subscriptions and licences that use neural stacks, software sold to automakers by specialist developers, and training data and simulation tools. Growth then follows the number of vehicles with neural stacks, the revenue per vehicle, the number of automakers that license rather than build and the spread of driverless operations.
The forecast is sensitive to the safety record of neural stacks, regulatory approval of driverless operation and the extent to which open models reduce prices for licences.
Scenarios
The base case assumes that major automakers launch neural stacks in assisted driving by 2028, that driverless operation reaches more than 20 large cities by 2031 and that licensing grows alongside in-house development, bringing the inflection to about 2029 and a 2034 value near the middle of the range. An accelerated case assumes strong safety records and early regulatory approval, moving the inflection to 2028 and lifting the 2034 value to the top of the range. A delayed case assumes a safety incident or tight regulation that slows deployment, moving the inflection to 2031 and holding growth at the bottom of the range.
Factors Behind Growth
Growth drivers
- Funding for developers of neural driving software.
- Open driving models and datasets from chip suppliers.
- Production rollout in China and in assisted driving.
- Automaker interest in licensing instead of building.
Enablers
- In-vehicle compute chips of high performance.
- Large driving datasets and simulation tools.
- Vision-language-action architectures.
- Over-the-air update capability in vehicles.
Restraints and barriers
- Difficulty of proving safety for learned models.
- Compute and cost needed in the vehicle.
- Cost of collecting rare-event data.
- Regulatory approval for driverless operation.
The Forces at Play
The direction of end-to-end driving models is set by safety evidence, regulation of driverless operation, the performance and cost of in-vehicle chips and the choice of automakers between licensing and building.
The brief assesses standards and policy, developer and investor funding, technology readiness, vehicle and chip readiness and customer readiness, and rates each force for direction, speed and confidence.
Adoption Outlook
How the shift is likely to unfold across three time horizons.
Single neural stacks reach more assisted driving products, and automakers sign licences with software developers.
Neural stacks are the default in new assisted driving systems, and driverless operation expands to more cities.
Driverless operation spreads across regions with models trained on fleet data and checked against explainable outputs.
Latest Strategic Developments
| Date | Development | Type | Significance |
|---|---|---|---|
| Jan 2026 | NVIDIA announces the Alpamayo family with a 10 billion parameter reasoning driving model, an open simulation framework and 1,700 hours of data | Product | Opens a reasoning model to developers |
| Feb 2026 | Wayve announces a USD 1.2 billion round at an USD 8.6 billion valuation with Mercedes-Benz, Nissan and Stellantis participating | Funding | Backs the licensing model |
| Mar 2026 | XPeng begins over-the-air rollout of VLA 2.0 on the P7, G7 and X9 in Ultra trim | Product | Brings a single neural stack to production cars |
| Apr 2026 | Reports name Volkswagen as the first external customer of XPeng VLA 2.0 for an electric SUV in China | Partnership | Shows a foreign automaker licensing a Chinese stack |
| Aug 2026 | NVIDIA clears the 34 billion parameter Alpamayo 2 Super for commercial use | Product | Makes a large open driving model available commercially |
| Aug 2026 | Waymo unveils a custom onboard compute chip for its robotaxi fleet | Product | Moves compute into the vehicle |
Key Players & Competitive Landscape
The key participants in end-to-end driving models include Tesla, Inc., Alphabet Inc., NVIDIA Corporation, XPeng Inc., Li Auto Inc., BYD Company Limited, Mobileye Global Inc., Qualcomm Incorporated, Baidu, Inc., Pony.ai Inc., WeRide Inc., Mercedes-Benz Group AG, Volkswagen AG, Stellantis N.V., Nissan Motor Co., Ltd., Uber Technologies, Inc., Microsoft Corporation, Toyota Motor Corporation, General Motors Company, Huawei Investment & Holding Co., Ltd., Wayve, Momenta, Horizon Robotics, Waabi and Nuro.
The brief profiles representative participants in each archetype and assesses which are positioned to lead as single neural stacks become the default.
Tesla, Inc., XPeng Inc. and Alphabet Inc. lead among developers of in-house neural stacks, Wayve and Momenta lead among licensors to automakers and NVIDIA Corporation, Qualcomm Incorporated and Mobileye Global Inc. lead among chip suppliers. Competitive intensity is high and rising as open models lower the cost of entry.
| Archetype | Representative players | Position | Outlook |
|---|---|---|---|
| In-house stack developers | Tesla, Inc., Alphabet Inc., XPeng Inc., Li Auto Inc., BYD Company Limited, Huawei Investment & Holding Co., Ltd. | Assisted driving and robotaxi | Build and run own neural stacks |
| Licensors of driving software | Wayve, Momenta, Mobileye Global Inc. | Software for automakers | License stacks to several automakers |
| Chip and compute suppliers | NVIDIA Corporation, Qualcomm Incorporated, Microsoft Corporation, Horizon Robotics | In-vehicle and training compute | Supply the compute that runs the models |
| Robotaxi operators and networks | Baidu, Inc., Pony.ai Inc., WeRide Inc., Uber Technologies, Inc., Waabi, Nuro | Driverless services | Operate driverless fleets and ride networks |
| Global automakers | Mercedes-Benz Group AG, Volkswagen AG, Stellantis N.V., Nissan Motor Co., Ltd., Toyota Motor Corporation, General Motors Company | Vehicles | Choose between building and licensing |
| Data and simulation providers | NVIDIA Corporation, Microsoft Corporation | Simulation and datasets | Supply data and simulation tools |
Where value migrates.
In 2026 value sits with developers of neural stacks that already ship in vehicles and with chip suppliers whose compute runs them. By 2029 it moves toward licensors that serve several automakers and toward automakers that own the data from their fleets. By 2034 it settles with operators and automakers that run driverless fleets at scale and with chip suppliers whose platforms are standard. Margins hold for developers with large data sets and narrow for suppliers of models that are available openly.
Who Will Win — and Why
The archetypes best positioned to capture value as the shift matures.
Data from vehicles in use is the main input for improving neural stacks.
Automakers that do not build their own stack buy from specialists, and the licensor spreads cost over many brands.
Every neural stack needs compute in the vehicle and in training.
Regulatory Landscape
| Jurisdiction | Milestone | Indicative timing | Effect on adoption |
|---|---|---|---|
| Driverless operation permits | Cities and states grant permits for driverless services with safety conditions | Ongoing | Sets the pace of robotaxi growth |
| Assisted driving rules | Regulators set conditions for hands-off and eyes-off driving | Ongoing | Sets the features that can be sold |
| Over-the-air updates | Rules on software updates for vehicle safety systems | Ongoing | Affects how stacks are improved |
| Data rules | Rules on collection and transfer of driving data between countries | Ongoing | Affects training data and cross-border sales |
| Explainability | Interest in outputs that explain decisions of driving models | Emerging | Supports reasoning models in assessment |
Investment Signals
Funding, open models, production rollout in China and automaker licensing support growth, while the difficulty of proving safety, in-vehicle compute cost, the cost of rare-event data and regulation of driverless operation hold back the pace.
A safety incident, a regulatory decision or a new open model could shift the pace. The brief therefore tracks four indicators: vehicles shipped with neural stacks, automakers licensing from specialists, cities with driverless operation and safety statistics per million miles.
Neural stacks in assisted driving, then default architecture across automakers, then scaled driverless operation is the expected order of change.
Questions This Brief Answers
Strategic Implications
- Automakers: decide whether to build or license a neural stack, and secure rights to fleet data.
- Software developers: build licensing relationships with several automakers and publish safety evidence.
- Chip and compute suppliers: offer platforms that run distilled models in the vehicle at lower cost.
- Robotaxi operators: use neural stacks to enter more cities and compare safety records with human drivers.
- Investors: favour developers with fleet data and automaker contracts over those with models alone.
"The rule book is being replaced by a driver that has seen more roads than any human could."
Table of Contents
Access & Licensing
A focused foresight brief, priced to circulate. Every option is delivered instantly and backed by analyst support.
every brief