The Stack | Labs, X, and the chips underneath

This is the short sketch. The fuller map is the AI Systems Atlas. It dates every link to a company announcement, carries the 2025 compute announcements this sketch leaves off, adds labeled chip schematics with where the memory sits and a local LLM memory lab, and ends by setting each lab’s benchmark claims beside independent runs. The Stack stays here as the one-screen version.

Main path Also uses NVIDIA Dashed boxes are not there yet in the selected year.

What a person uses

Who builds the model

Where the machines sit

The chip doing the arithmetic

On a phone, swipe the diagram sideways. The lines are redrawn from the boxes, so they stay attached.

Same job: multiply huge grids of numbers. The blocks are roles, not silicon area, speed, or core counts. A school laptop’s neural chip is a different, much smaller machine.

What actually differs
Question	NVIDIA GPU	Google Cloud TPU	AWS Trainium
Original job	Draw graphics, then general parallel math.	Neural-network math in Google’s data centers.	Train models on AWS. Inferentia is the partner chip that serves answers.
Who designs it	NVIDIA. Labs buy it or rent it.	Google. Google uses it and rents it on Google Cloud.	Amazon. Customers rent it. They do not drop it into a laptop.
Software path	CUDA, since 2006, and the frameworks built above it.	Google’s own stack, including JAX and TensorFlow.	AWS Neuron.
In this map	The shared floor. OpenAI and xAI stand on it. Anthropic and Google also use it.	Google DeepMind’s own chip path, beside NVIDIA GPUs on Google Cloud.	The path Anthropic announced with AWS in 2023, in addition to chips it already used.

Tensor Cores are matrix units inside an NVIDIA GPU. A Cloud TPU is a different company’s accelerator. The Tensor chip in a Pixel phone is a third thing, a mobile system on a chip. The shared word does not mean a shared part.

Year 2026

Sources

Checked October 12, 2026. This is a map of public institutional history, not a product catalog. Model names and private valuations change faster than the drawing.