DeepRoute.ai VLA model, a 40-billion-parameter vision-language-action foundation model, took center stage at NVIDIA GTC 2026 in San Jose, California on March 18, 2026. The company presented the architecture as a unified system that integrates perception, reasoning, and action for autonomous driving at scale.
The model addresses a persistent bottleneck in traditional closed-loop data workflows. Conventional systems require manual data collection, review, labeling, and model retraining — a cycle that typically exceeds five days per iteration. DeepRoute.ai said its approach compresses that cycle to approximately 12 hours through intelligent automation.
Three Roles in One DeepRoute.ai VLA Model
The 40-billion-parameter foundation model performs three simultaneous functions. It acts as a driver, executing real-time actions from visual inputs. It also serves as an analyst, identifying critical driving events and explaining decisions through causal reasoning. Furthermore, it functions as a critic, evaluating trajectories for safety, comfort, and human-like behavior.
“Our solution to the industry bottleneck is a unified 40-billion-parameter vision-language-action foundation model. This model goes beyond basic vehicle control — it can analyze data and evaluate driving behavior. Simply put, it does not just play the role of the driver, but simultaneously acts as the analyst and the critic.”
Tongyi Cao, Chief Technology Officer, DeepRoute.ai
Self-Evolving Data Flywheel
The unified architecture enables a self-reinforcing development cycle. Improvements in driving performance feed directly into the system’s ability to process and curate its own training data. The model independently handles data mining, cause diagnosis, and behavior evaluation without manual intervention.
“Traditional closed-loop data pipelines rely heavily on manual processes, severely limiting iteration speed. By leveraging our foundation model, we have restructured the entire workflow. Every iteration cycle translates directly into measurable improvement in our artificial intelligence capabilities.”
Tongyi Cao, Chief Technology Officer, DeepRoute.ai
Commercial Deployment: 250,000 Vehicles and Counting
DeepRoute.ai reported equipping more than 250,000 production vehicles with its autonomous driving systems by the end of 2025. In October 2025, the company captured approximately 40% market share among third-party suppliers in the high-level autonomous driving segment within a single month. The company now targets deployment in one million vehicles by the end of 2026.
DeepRoute.ai has raised more than $700 million in funding from major investors. The company said the foundation model forms the cornerstone of its next-generation advanced driver assistance systems and serves as a physical-world AI framework.
Outlook for Scalable Autonomous Driving
Through its GTC 2026 presentation, DeepRoute.ai demonstrated how the VLA foundation model architecture accelerates the path toward safe, scalable autonomous driving. The system achieves this through continuous data-driven learning and rapid iteration cycles, reducing dependence on manual labeling significantly.
“Autonomous driving is, at its core, a scaling problem. Despite significant progress across the industry, large-scale deployment remains out of reach because traditional implementation pipelines have fundamental flaws. The bottleneck is no longer about acquiring data — it is about how efficiently the system can filter noise and convert massive volumes of raw data into high-value training samples.”
Tongyi Cao, Chief Technology Officer, DeepRoute.ai
The company said the unified architecture enables autonomous driving systems to understand complex traffic environments, interpret the logic behind decisions, and evaluate driving behaviors — capabilities it described as broader cognitive and decision-making functions. DeepRoute.ai also stated its vision targets artificial general intelligence in the physical world, with commercial robotaxi operations as a long-term objective.