Tesla’s Full Self-Driving solution continues to look less and less like a must-license platform and more and more like a strategic fork in the road. When Elon Musk says that the traditional car manufacturers are “crazy” not to follow Tesla’s lead, what the industry has been up to lately suggests something more calm and calculated: control.
Autonomy is one of the few remaining aspects of a modern car that can have a distinct brand feel. The feel of the steering and suspension is still important, but now code influences everything from how a car checks a driver’s attention to how it selects a gap in traffic. Car companies that cede control of this aspect are essentially surrendering a unique driving experience to a settings menu.
This kind of thinking appears in the tech direction, not just the marketing speak. Rivian took the opportunity presented by its Autonomy and AI Day to explain how it is developing the “brain” of its autonomous driving solution around a custom compute stack, which includes a chip tailored for AI compute and an autonomy computer designed for scaling capabilities over time. One of the most interesting specs that Rivian revealed is 205 gigabytes per second of memory bandwidth for its custom chip, a reminder that autonomy is not just about the algorithm; it is also about getting sensor data through silicon quickly enough to make decisions.
But the legacy players are going down a different road, with the same endpoint: to bring more of the stack inside. Ford is going to introduce an eyes-off highway driver-assistance system in 2028 and develop it in-house to reduce costs and accelerate updates. “To integrate, I can’t do this with all these suppliers,” Ford engineering executive Paul Costa said. “We need to bring this stuff in-house, and it allows for this ability to do the trifecta at once: smaller, cheaper, and higher performance.”
The engineering rationale is direct. Autonomous systems are a tightly integrated system of software, sensors, and actuators that physically control steering, braking, and acceleration. When these components are sourced from multiple companies, integration is a never-ending project with shifting goals. Control of more of the stack means less friction, smaller interfaces, and easier iteration without having to re-negotiate each change in a supply chain.
One sentence summarizes why licensing Tesla’s system has become an increasingly tough sell: the hurdle to create autonomy is falling. At CES, Nvidia announced Alpamayo, a line that includes open models, simulation infrastructure, and data sets that target training systems to deal with the “long tail” of edge cases. Nvidia said that Alpamayo 1 is a 10 billion parameter vision-language-action model, combined with an open simulation framework and 1,700+ hours of driving data. In short, more of the building blocks are being made available to groups that want to control their own driving.
Mercedes is cited as an example of how this manifests. The company already has a limited Level 3 solution in place in certain circumstances, but it also intends to introduce the Nvidia solution in the CLA through an AI-based driver assistance system that aims for Level 2 point-to-point support. This “toolkit plus OEM integration” business model ensures that the car makers retain control of how the functions feel and where they are applied.
Industry expectations over a longer term also lean towards incremental, brand-owned systems over a single external stack. McKinsey’s research on autonomous vehicles points out that the mass market for privately owned vehicles is moving towards L2+ capabilities by 2035, with increasing expectations of hybrid systems that combine end-to-end AI with traditional verification. This approach favors those who can tune, validate, and defend their own systems, which becomes increasingly difficult when the underlying logic is licensed. Therefore, the industry’s hesitancy to license Tesla FSD is more a function of product strategy than denial. As autonomy becomes the interface, “companies want to own the interface.
