If the data is not transformed using a non-linear kernel function (as is this case, because we are still working with the data x and not the feature expanded version phi), then SVM gives linear decision boundaries in the original space. You can check the SVM slides in week 7 (slide 24, 25) for a short reminder.
In the end SVM actually always finds linear decision boundaries in whatever space you transform your data to. Let's say you have a 2 dimensonal x. You transform this using a non-linear kernel to 8 dimensions. Now, the decision boundary that SVM finds will be linear in this new 8 dimensional space (not the original space!). But if you look at this decision boundary in the original low dimensional space, it will be curvy.
I hope this clarifies things! Best,
Sena