Step-by-step guides on prompting, styles, and getting the most out of AI image generation.
Apple's cancelled self-driving car program — not its consumer device roadmap — is the reason the company's M-series chips deliver the on-device AI performance that creators and developers rely on today.
Early in the development of its self-driving platform, Apple concluded it needed serious on-device AI processing — the kind that could handle real-time sensor fusion and decision-making without leaning on a data center. That requirement pushed Apple's chip teams toward neural engine designs that were far more capable than anything needed for a smartphone at the time. The car processor itself was never finished, but as The Verge reports, the engineering groundwork it generated fed directly into what eventually became Apple Silicon's AI architecture.
The practical result: M-series chips — from the M1 through the M4 Ultra — carry neural engine hardware whose ambition was originally scoped for a vehicle, not a laptop. That over-engineering, by automotive standards, turns out to be exactly the right spec for running large AI models locally on a Mac.
Apple's M-series chips trace their AI processing roots to an abandoned self-driving car program.
Image: The Verge / The Verge AI
For AI-art creators running image generation locally — Stable Diffusion, FLUX, or similar pipelines via tools like Diffusers or ComfyUI — the practical upshot is that Apple Silicon's neural engine headroom is deeper than it needed to be for consumer devices alone. That surplus capacity is why M3 and M4 MacBooks can run quantized models at speeds that would have required a discrete GPU just two years ago.
The on-device advantage matters beyond raw speed. Local inference means no API costs, no usage caps, and no image data leaving the machine — a meaningful consideration for creators working on commercial projects or sensitive character designs. If you're exploring what's possible with local generation, the Charmloop guides cover model setup and optimization for exactly these workflows.
The lineage also hints at where Apple's hardware is heading. If the car program's chip ambitions were already baked into the M1 generation, the M4 Ultra and whatever follows it represent several more years of iteration on top of that foundation — not a fresh start. That compounding effect is what makes Apple Silicon increasingly competitive with Nvidia's consumer GPU line for inference workloads, even if training still belongs to CUDA.
Apple's car isn't the only example of a failed moonshot quietly funding useful technology. The pattern — a high-ambition project gets cancelled, but its engineering output survives in adjacent products — is common enough in large tech companies that it's almost a funding strategy in disguise. The difference here is scale: Apple reportedly spent over a decade and billions of dollars on the car before shelving it, which means the chip R&D budget was correspondingly large.
For anyone choosing hardware for an AI-generation setup, the takeaway is concrete: the neural engine in current Mac hardware was designed to handle a self-driving car's workload. Running a 12-billion-parameter image model on it is, by comparison, a lighter ask than the original spec. That gap between designed capacity and actual workload is where performance headroom lives — and why Apple Silicon continues to punch above its thermal envelope for on-device AI work.
Apple's next chip generation, expected to carry the M5 designation, will be the first designed entirely after the car program's formal cancellation in 2024. Whether that changes the architectural ambition — or simply refines what the car program started — is the question worth watching.
