It's about the geometry of spacetime.
(The what?)
In special relativity (SR), "boosting" to a frame of reference with a different velocity, mixes space and time coordinates, (x,t), in a similar (but not identical) way that rotating a geometrical figure in the xy-plane does with those two space coordinates.
When you start with an x-line-segment, and rotate it through an angle θ₁, you can see that it gets a slope of
s₁ = tanθ₁
If you rotate it further, by an angle θ₂, so that a second x-line-segment, acquires a slope of
s₂ = tanθ₂
the original segment doesn't get a slope of
s₁ + s₂
instead, it now slopes at an **angle** of
θ₁ + θ₂
In other words, we all realize that in successive rotations, slopes don't add; angles do.
In SR, when "boosting," there is a quantity that acts like the angle in a rotation; call it the "velocity parameter," α.
Now the analogy here is that velocity for boost transformations, is like slope for rotations.
So that when successive boosts are done, it isn't the velocities that add; it's the velocity parameters.
And the value of a velocity parameter is unbounded; it keeps growing, linearly, as you constantly accelerate. Only now, instead of the
"slope = tangent of angle" . . . . s = tanθ
rule for rotations, there is a
"β = velocity/c = hyperbolic tangent of velocity parameter" . . . . β = v/c = tanhα
rule for boosts.
From this, you can derive the "velocity-addition formula" for SR, from the formula for tanh of a sum of angles:
tanh(A+B) = (tanhA + tanhB)/(1 + tanhA tanhB):
β = (β₁ + β₂)/(1 + β₁β₂)
This is the analog of the tan of a sum of angles:
tan(A+B) = (tanA + tanB)/(1 - tanA tanB)
which can be used to find the slope of that original x-segment after two rotations:
s = (s₁ + s₂)/(1 - s₁s₂)
You can also see that, as α→∞, v→c⁻, not ∞, because of the behavior of tanh, as its argument grows arbitrarily large.
And that, for non-relativistic boosts, for which the α values are very small (|α|<<1), v ≈ α, and velocities *do* add (in the sense that the fractional error goes to 0 in the limit), for this Newtonian approximation.
Just as, for rotations by very small angles, slopes *do* add (fractional error again going to 0 in the limit).