The building blocks
Compute, network, storage, power and cooling on one page: what each is in an AI data centre, how they depend on each other, why each matters and what it does not solve.
Every AI data centre is built from the same five things. Compute does the work; the network and storage keep it busy; power and cooling keep it running. None of them is useful alone, and the weakest one sets the pace for the rest.
The GPUs and servers that do the actual work: training models and answering requests.
- A CPU does a few complicated tasks one after another. A GPU runs thousands of simple calculations at once, which is exactly what a neural network needs.
- It is built up from chip to server to rack to cluster, so a single job can span thousands of GPUs working as one.
- Training is one long job with every GPU in step for weeks; inference is many short requests, each answered in a fraction of a second.
Why it matters
The most expensive part of the building and the fastest to go out of date. Everything else on this page exists to keep it busy.
What it does not solve
More GPUs do not make a job faster on their own. Without the network, storage, power and cooling to match, they sit idle.
Key Takeaways
- Compute does the work; everything else exists to keep the GPUs busy.
- The network lets thousands of GPUs act as one machine, and is sized for bursts, not averages.
- Storage is tiered, and checkpoints arrive as a burst from every GPU at once.
- Power is the scarcest input and often the slowest to secure.
- Every watt becomes heat; at AI densities, liquid cooling has replaced air.