x = [35.668 34.448 33.317 30.357 27.598];
y = [180 504 1260 10300 80900];
First, recognize you have only 5 data points. Any expectation of a good estimate of the parameters is wildly optimistic.
Next, recognize that your y variable varies by a factor of 400 over the curve. That means the point at x = 27.598 will have a HUGE impact on the estimates. Typically, this means you might need to consider other error structures, for example a proportional error might make more sense.
ft = fittype('a-b*exp(-c*x)','indep','x')
mdl = fit(x',y',ft,'start',[50 -1e13 1])
b = -1e+13 (-7.311e+13, 5.311e+13)
c = 0.6749 (0.4492, 0.9006)
The first thing you should recognize is the coefficient b is NEGATIVE. I don't know why you think your model should be of the form
In fact, that model, where you SUBTRACT the exponential term results in a NEGATIVE value of b.
So the resulting model looks like this:
y = -752 + 1e13*exp(-0.6749*x)
You should also see the model put confidence bounds on a that are huge, [-4674,3169]. See that a, which is the asymptote as x -- inf, is NEGATIVE. So this model predicts your function will go below zero for large x.
You need to recognize your data is literally worthless to estimate the parameters. You don't have enough data. And for some reason, you don't understand the shape that negative sign on b implies.