The AI Software Stack, Live: The Real Architecture, Every Tool, Data in Motion
The software of an AI factory drawn the way engineers draw it: a bare-metal hardware foundation across the whole width (chips, memory, fabrics, storage, racks, with power and cooling as supporting infrastructure), firmware, infrastructure software, runtime and a shared ML platform above it, six lifecycle stages side by side on top (data prep, pre-training, post-training, evaluation, serving, agents), a control plane on one side and observability and security on the other. 84 technologies appear as boxes, each carrying its own line of innovations, and the traffic between them is animated as streams. Watch a training step reach the GPUs and HBM, a user request, an agent task, a post-training loop, a hardware failure and recovery, a new rack bring-up, and power and cooling, all at the same time. Turn up the load to find the bottlenecks, or drag the year back to 2012 and watch the stack disappear.
How the architecture is organised
Only the bottom of the picture is a true stack: bare metal → firmware → infrastructure → runtime → shared frameworks and kernels, where each layer calls the one below. Above it, data prep, pre-training, post-training, evaluation, serving and agents are stages side by side, all running on that same platform and passing datasets and checkpoints left to right. Scheduling is a control plane beside the stack, and observability and security cut across everything. Simplified stack shows the older one-layer-per-row teaching view.
Related: Stack Innovation Map (when each technology arrived) · The AI Factory in 3D (where it physically runs) · Who Runs the AI Factory (the people) · ML Layer, Data Layer, Evaluation and Agent Layer hubs.