Tesla is playing a different game: chips > vehicles.

TESLA JUST KILLED THE CAR
The vehicle is now a peripheral. The chip is the product.
AI5 tape-out confirmed. What nobody understands yet:
Tesla deleted the Image Signal Processor. Deleted the GPU. What remains is a half-reticle inference engine that ingests raw photons.
Not images. Photons.
The numbers:
40x faster in bottleneck operations.
8x raw compute.
9x memory.
144GB onboard.
250 watts.
SoftMax, the mathematical spine of every transformer model, ran in 40 emulation steps on AI4. On AI5, it executes natively in one.
This is not iteration. This is architectural extinction.
Samsung signed $16.5 billion through 2033 for AI6. Both fabs, Texas and Arizona, produce US silicon. Tesla is building chip independence while Nvidia sells to everyone.
The strategic trap nobody sees:
Cybercab launches 2026 on AI4. Volume AI5 arrives mid-2027. Eighteen months of exposure where the robotaxi runs on yesterday’s brain.
If the software team delivers unsupervised autonomy on constrained hardware, AI5 becomes a turbocharger.
If they fail, AI5 becomes a rescue mission that arrived late.
The 100GW thesis:
100 million vehicles. Each carrying 1kW of inference. Parked and plugged in, they form a distributed supercomputer larger than every data center on Earth combined.
Tesla is not building cars with chips.
Tesla is building chips that happen to move.
The age of software-defined mobility ends here. What follows is silicon-defined existence.
Waymo geofences cities. Tesla geofences physics.
Watch Q1 2026 silicon yields. Watch Cybercab disengagement rates. Watch the moment the car disappears and only the network remains.
The platform chip is real. The delay is real. The consequences are civilization-scale.