I have been experiencing the following problems with Mathematica 9 and also Mathematica 10:

ClearAll["Global`*"];
(* obtain constants *)
h = QuantityMagnitude[
   UnitConvert[Quantity["PlanckConstant"], "SIBase"]];
kB = QuantityMagnitude[
   UnitConvert[Quantity["BoltzmannConstant"], "SIBase"]];
c = QuantityMagnitude[UnitConvert[Quantity["SpeedOfLight"], "SIBase"]];
rIn = QuantityMagnitude[
   UnitConvert[Quantity[0.1, "AstronomicalUnit"], "SIBase"]];
rOut = QuantityMagnitude[
   UnitConvert[Quantity[10, "AstronomicalUnit"], "SIBase"]];
r0 = QuantityMagnitude[
   UnitConvert[Quantity[0.1, "AstronomicalUnit"], "SIBase"]];
\[Lambda]Low = 1*10^-6;(* Meters *)
\[Lambda]High = 10*10^-6;(* Meters *)
T0 = 1000;
\[Alpha] = -3/4;
a = T0/r0^\[Alpha];
F\[Lambda][\[Lambda]_] := 
  NIntegrate[
   4 \[Pi]*(h*c^2)/\[Lambda]^5*r/(
    Exp[(h*c)/(\[Lambda]*kB*a*r^\[Alpha])] - 1), {r, rIn, rOut}];
LogLogPlot[
 F\[Lambda][\[Lambda]], {\[Lambda], \[Lambda]Low, \[Lambda]High}]

Now this returns:

NIntegrate::inumr: The integrand (7.483543*10^-16 r)/((-1+E^((3.36356*10^-13 r^(3<<1>><<1>>))/\[Lambda])) \[Lambda]^5) has evaluated to non-numerical values for all sampling points in the region with boundaries {{1.49598*10^10,1495978707000}}. >>

NIntegrate::inumr: The integrand (7.483543*10^-16 r)/((-1+E^((3.36356*10^-13 r^(3<<1>><<1>>))/\[Lambda])) \[Lambda]^5) has evaluated to non-numerical values for all sampling points in the region with boundaries {{1.49598*10^10,1495978707000}}. >>

NIntegrate::inumr: The integrand (7.483543*10^-16 r)/((-1+E^((3.36356*10^-13 r^(3<<1>><<1>>))/\[Lambda])) \[Lambda]^5) has evaluated to non-numerical values for all sampling points in the region with boundaries {{1.49598*10^10,1495978707000}}. >>

General::stop: Further output of NIntegrate::inumr will be suppressed during this calculation. >>

And then the plot. But if I use Plot instead of LogLogPlot, there are no errors.

I could redefine the function:

F\[Lambda][\[Lambda]_?NumericQ] := 
  NIntegrate[
   4 \[Pi]*(h*c^2)/\[Lambda]^5*r/(
    Exp[(h*c)/(\[Lambda]*kB*a*r^\[Alpha])] - 1), {r, rIn, rOut}];

which really hides the errors but I am more interested why this problem happens. Because NumericQ is not always the good solution. What if I want to pass infinity which does not evaluate true as NumericQ or if I am using units!