The text of the analysis of the prediction time is wrong. Currently it reads:

However, prediction of 100000 target values is more than tree times faster with SVR since it has learned a sparse model using only approx. 1/3 of the 100 training datapoints as support vectors.

But both on the main branch with 100k samples and in your PR with 70k samples, the KRR model predicts faster. So we could fix the text to be something like:

The speed of prediction of SVR could in theory be 3x faster than KRR because SVR uses approximately 1/3 of the 100 training datapoints as support vectors. However here we observe that this not the case, probably because of implementations details (the SVR prediction code does not seem to be as well optimized as the KRR prediction code).

and reduce the prediction set to 10k samples instead (large enough to measure a timing that is not too noisy but small enough to make this example run significantly faster).

Thank you for your comments.

If we reduce the prediction set to 10k samples we now have this outputs:

SVR complexity and bandwidth selected and model fitted in 0.801 s
KRR complexity and bandwidth selected and model fitted in 0.428 s
Support vector ratio: 0.290
SVR prediction for 10000 inputs in 0.029 s
KRR prediction for 10000 inputs in 0.069 s

plot_kernel_ridge_regression.py has been accelerated in #21791.
I'm closing this pull request.