Aside from simply looking at the Lorentz transformation, seeing a divergence and concluding "meh, it doesn't work", another way to gain insight into the divergence is through the statement:
no finite sequence of finite boosts will get you to a speed $c$ relative to your beginning inertial frame.
Imagine yourself in a spaceship with orientation controls and a booster such that you can accelerate yourself to any speed in some finite interval (let's say $[0,\,\Delta v]$ with $\Delta v\ll c$) in any direction relative to your present momentarily co-moving inertial reference frame in a unit time as measured by your on-board spaceship clock.
Group theoretically, this is equivalent to the assertion that after unit time, there is some neighborhood $\mathcal{N}_\mathrm{id}$ of the identity in $SO(1,\,3)$ such that I can impart any Lorentz transformation in that neighborhood to present my reference frame. As time (as measured by my trusty on board clock) goes by, I can impose any sequence of these neighborhood members; the overall transformation relative to my beginning frame is their product and so my overall transformation follows some continuous path through $SO(1,\,3)$. Our initial statement is equivalent to:
No member of the identity-connected component of $SO(1,\,3)$ corresponds to a relative speed of $c$
(indeed this is of course true of any member of $SO(1,\,3)$, but the identity component is the transformations we can physically reach with out controls, given enough fuel).
In particular, imagine heading off in a steady direction; and each unit time you are going to impose the same boost. As in WillO's answer, we concentrate on one spatial dimension, so our unit boost is:
$$\Delta\Lambda = \exp\left(\delta \eta\left(\begin{array}{cc}0&+1\\+1&0\end{array}\right)\right)=\left(\begin{array}{cc}\cosh\delta\eta&\sinh\delta\eta\\\sinh\delta\eta&\cosh\delta\eta\end{array}\right);\quad \delta\eta = \operatorname{artanh}\frac{\Delta v}{c}\approx \frac{\Delta v}{c}\tag{1}$$
The same, finite, $\Delta v$ relative to our present frame imparted $n$ times over is $\exp\left(n\,\delta \eta\left(\begin{array}{cc}0&+1\\+1&0\end{array}\right)\right)$. So, when we accelerate at a uniform velocity $\Delta v$ per unit time by to our clock, so we feel a constant accelerating force from our seat, an observer in our initial frame sees as accelerate so that our rapidity $\eta = n\,\delta\eta$ changes by an amount $\Delta v/c$ in a time interval $\cosh\eta$ relative to their frame. So the change in overall velocity between the two frames is
$$c\,\left(\tanh\left(\eta+\operatorname{artanh}\frac{\Delta v}{c}\right) -\tanh\eta\right) \approx \Delta v\,\operatorname{sech}^2\eta\tag{2}$$
and our apparent acceleration from our beginning frame is $\Delta v \,\operatorname{sech}^3\eta$; We seem to be accelerating more and more slowly, both because it is our rapidity, not speed, that is changing uniformly with boost interval number and also because these boost intervals are getting longer and longer relative to the beginning frame.
In neither frame will the overall speed difference ever reach $c$.
Note that the above arguments apply even if $\Delta v$ is a large fraction of $c$. When $\Delta v\ll c$ the approximation in (2) holds.