Compiling source code also takes computing resources, only the scale is different. It's a very close analogy: source code with training data, and compiled binaries with model weights.

The weights are literally a binary blob.

The weights are the modifiable representation. You modify them with gradient updates.