Chief product officer Sterling Anderson argues the speed does not come at the expense of safety. Instead of building a handful of physical prototypes for crash testing over six months, GM now runs thousands of virtual tests that return feedback much faster, which he says also improves coverage and understanding.

The approach leans on transformer-based foundation models, the same family of techniques behind modern chatbots, trained on decades of GM design files and analyses and co-trained with data from outside the company. Anderson says that external data tends to make the models better, so GM applies it across the board.

Where the models run depends on the job. Anderson says safety-critical functions must compute onboard the vehicle to avoid latency or connection dropouts, while features such as OnStar AI interactions can round-trip to the cloud, though GM is considering distilling those into onboard models too.

The pivot is a gamble. GM faces fast-expanding Chinese EV makers, softening US EV demand and software-first rivals such as Tesla, so delays or faltering AI systems would be costly while the company balances multibillion-dollar battery plant investments with ongoing software development.