This is called aliasing. I recommend reading the linked Wikipedia article. The sample rate (in this case the number of time samples in your t vector, per second), must be at least twice the frequency of the signal in order to even approximately represent the signal. A better rule of thumb is at least 10 samples per period of the signal.
So, for a signal at approximately 8 GHz you need to sample it at around 80 GHz to get a relatively clean representation of the signal. This means that, for an 16 second time span you would need this many samples:
num_samples = 16*80e9
Of course that is going to take about a terabyte of memory and, if you had that much memory, your plot would look like a blue rectangle and you couldn't see the wave shape at all.
Each period of an 8 GHz sine wave is only
period = 1 / 8e9
In other words, 0.125 nanoseconds.
Shorten the time span of your plot. For example:
f=7.9*10^9; % frequency of my cosine 7.9 GHz
w=2*pi*f;
t=linspace(-8e-10,8e-10,400); % time interval with 400 is the number of samples that I change
y=cos(w*t);
plot(t,y)
grid on
