A chip 2 to 3x better than NVIDIA at 10% of the cost sounds like fantasy until you understand what NVIDIA actually sells. An H100 has to run every model architecture, for every customer, on every framework, with 15 years of CUDA compatibility behind it. That generality costs silicon. Die area for legacy blocks. Interconnect for data flows that might never happen. Memory bandwidth sized for workloads NVIDIA can't predict. Tesla's AI5 runs one company's neural nets. Engineers know exactly what data moves between which blocks, so they deleted the highways nothing travels on. Legacy GPU support, image signal processors, anything that doesn't serve Tesla's own models: gone. Half the die area, same job. Then add the margin math. NVIDIA's gross margin runs around 73%. Roughly a quarter of what you pay covers the actual chip. Tesla pays cost, because Tesla is the only customer. Before a single engineering advantage, vertical integration alone closes most of that 10x cost gap. Google figured this out in 2015 with the TPU. Amazon followed with Inferentia. Microsoft built Maia. Bitcoin miners learned it fastest: ASICs killed GPU mining in roughly two years once the workload stopped changing. That's the pattern. While a workload is still evolving, general-purpose chips win. The moment it stabilizes, specialized silicon beats them on every metric that matters. And the limitation is exactly what makes the claim credible. AI5 can't train frontier models. Nobody else can buy it without a decade of software work. It does one thing: run Tesla's inference at the lowest cost per watt, across millions of cars and robots. NVIDIA charges a premium for optionality. Tesla just stopped needing any.
