tinygrad is a neural network framework that reduces all ops to three types — Elementwise, Reduce, and Movement — with CONVs and MATMULs apparently assembled from those primitives, a mystery the page says you solve by reading the code. It has full autograd, a PyTorch-like API, lazy tensors, and compiles a custom kernel per operation for shape specialization. It's in alpha until it reproduces a common set of papers 2x faster than PyTorch on one NVIDIA GPU; ETA is "Q2 next year," a date that has historically been the most stable part of the project. It's proven in production: it runs the openpilot driving model on a Snapdragon 845 GPU. Funding exists, hiring is contribution-gated, and the only accepted payment for their hardware is wire transfer. Which is either extreme operational discipline or a company that has seen what happens when the FAQ answers "no." The tinybox hardware line ranges from 778 TFLOPS FP16 (4x 9070XT, 64 GB) up to the exabox: ~1 EXAFLOP across 720 GPUs, 600 kW, 20,000 lbs, mounted on a concrete slab, shipping 2027. Jensen says the more you buy, the more you save; at 720 GPUs, that must be a lot of saving.