Where each part of the AI pipeline stands now
A short account of the nine parts of the AI pipeline, in order. Energy and chips have their own pieces. The other seven are marked coming soon, and a last piece will look at bottlenecks across the chain.
The map of this series names nine parts, grouped as power and chips, putting a model to work, and making a model. That map is The nine parts of the AI pipeline. This page is the short version: what each part is, in the same order, and where the longer piece is if it exists.
Energy and chips have their pieces. The other seven say Coming soon, with no link, because I would rather show the gap than send you to a page that is not there. After those nine, a last piece will look across the whole pipeline for where it is stuck.
The nine parts
Electricity and chips come first, because every other part runs on them. The next four cover putting a model to work: connecting it to your work, checking its output, running it, and how people use it. The last three cover how a model gets made, from the data to the training run to the model itself.
Area 1 Energy
AI runs in data centers that use electricity to run, move power, and take heat away. They also run the rest of the internet, so the total is bigger than AI alone. The longer piece is Energy, where it stands now.
Area 2 Chips
Most of the arithmetic runs on graphics processing units (GPUs) and on newer chips designed for AI. They do an enormous number of simple multiplications at the same time. The longer piece is Chips and hardware, where it stands now.
Area 3 Connecting the model to your work
This part is software wrapped around the model so it can search or use a tool. A gateway records each call, and an agent works through a task in many steps. Coming soon.
Area 4 Verifying the output
Checking is the decision that an output is good enough to use. In regulated work, such as making medicines, that check is a formal written step. Coming soon.
Area 5 Putting the model to work
Using a trained model is called inference, one run for each token it writes. That cost is paid on every use, in a data center or on a laptop or phone. Coming soon.
Area 6 How people actually use it
This part is people putting AI to work. In my view, a capable model often produces nothing useful here, because the work around it never changed. Coming soon.
Area 7 Where the data comes from
A model learns from examples. For a large language model (LLM), that mostly means text, counted in tokens (a word or part of a word). Coming soon.
Area 8 Who’s training models for pharma
Training is where data and chips meet. In pretraining, the model practices guessing the next token, and a second stage tunes it to follow instructions and turn some requests down. Coming soon.
Area 9 What the model actually is
What comes out of training is a file of weights, numbers that turn text in into text out. The design is the transformer, described in a 2017 research paper. Coming soon.
Forces outside the pipeline
Four forces press on every part. They are not extra stations on the chain. They change how fast each station can move, and whether it moves at all.
After the nine
I will write the other seven parts one at a time, each one on where that part stands, with a public source for every number. When a piece exists, this page will link it in place of Coming soon. The series then ends with a piece on bottlenecks across the whole pipeline: the places where one part holds the others up.
My own work sits near the answer. I built a small gateway from scratch to show what the layer between apps and model hosts does, and I keep the Pharma AI Tracker, a record of what pharma companies have said in public about where they use AI. Both sit downstream of the parts that come before the answer. That is why I wanted those early parts in the same plain words as the work I do every week.
My working view: a limit rarely stays in its own part. A shortage of electricity can show up later, as an answer that costs more or arrives slower. The last piece in the series is where I line those limits up.
The picture of all nine parts, and the four forces under them, is The nine parts of the AI pipeline.