Every system in this database — every chip, data center and power plant — exists to do one thing at absurd scale: multiply matrices. Here is why.
Text becomes numbers
A language model never sees words. Text is chopped into tokens — fragments like "infra" or "structure" — and each token becomes a long list of numbers. From that point on, everything is arithmetic.
The model is a stack of layers
A modern model is a transformer: dozens to hundreds of identical layers, each containing an attention step and a feed-forward step. Attention lets every token look at every other token and decide what matters — in "the chip that TSMC makes", attention is what connects "makes" back to "TSMC". The feed-forward step then transforms each token through learned weights.
Those weights — the parameters — are just numbers, billions to trillions of them. They are the model. When you download an open-weights model, you download these numbers.
Training: learn by predicting
Pre-training shows the model trillions of tokens of text and asks, over and over: what comes next? Every wrong guess nudges the weights slightly. Repeat quadrillions of times across months on tens of thousands of accelerators, and the weights come to encode grammar, facts, code style, reasoning patterns — everything needed to predict text written by humans.
Then post-training shapes the raw predictor into an assistant: examples of good behavior, human preference feedback (RLHF), and — since 2024 — reinforcement learning on problems with checkable answers, which produced the reasoning models.
Inference: run it
Inference is using the trained model. It generates one token at a time: compute a probability for every possible next token, pick one, append it, repeat. A 500-word answer is roughly 700 sequential passes through the entire network — which is why memory bandwidth (how fast weights stream from HBM into the compute units) usually limits speed, not raw compute.
Why this needs a database of infrastructure
Nothing above is exotic — multiply, add, compare. The exotic part is scale. Frontier training runs use 1025–1026 floating-point operations; serving hundreds of millions of users means data centers drawing more power than cities. The rest of this site maps the machine built to do that arithmetic: chips → systems → data centers → power.
