Tesla Axes Dojo Team, Shifts to Inference Chips

On August 7, 2025, Bloomberg reported that Tesla CEO Elon Musk disbanded the Dojo supercomputer team, with leader Peter Bannon exiting the company. Dojo, unveiled in 2021, was central to training Tesla’s Full Self-Driving (FSD) neural networks using custom D1 and D2 chips. Once projected to add $500 billion to Tesla’s value, it faced hurdles, including issues in the June 2025 robotaxi pilot in Austin and the loss of 20 engineers to DensityAI.

Musk confirmed Tesla will now focus on AI5 and AI6 inference chips—optimized for real-time decision-making in vehicles and robots—produced under a $16.5 billion deal with Samsung. Training needs will be met through Cortex, a new Austin-based AI supercluster using over 100,000 Nvidia accelerators.

While inference chips can accelerate deployment, they risk compute bottlenecks, supply chain dependency, and reduced flexibility for multimodal AI. Here, a hypothetical alternative—FaceOff Lite with its Adaptive Computing Engine (ACE)—could address these challenges.

ACE integrates training and inference in one platform, supporting multimodal sensor processing, on-device learning, and dynamic compute allocation. It reduces reliance on cloud training, optimizes energy use, and scales across Tesla’s EVs, Optimus robots, and robotaxi fleets.

A hybrid model—Cortex for initial training, ACE for edge refinement—could boost FSD adaptability, cut costs by unifying chip functions, and reclaim Tesla’s AI differentiation from Nvidia-dependent rivals.

In the short term, Tesla’s pivot streamlines resources and speeds rollout. Long term, ACE could revive the original Dojo vision, enabling Tesla to lead in autonomous driving, robotics, and AI services—provided it retains top talent and rigorously tests in real-world conditions.

This shift underscores a strategic crossroads: pursue rapid market entry via proven inference chips or invest in integrated AI architectures that secure Tesla’s future autonomy and competitiveness.