As best I can tell this reduces to the computation below.
NIntegrate[(671.1784345137916*E^(-2.*(0.00183562 + Log[0.00128038*x2])*
(-162.12*(-0.00198044 + Log[0.000588107*x1]) +
82.7225*(0.00183562 + Log[0.00128038*x2]) -
37.947*(0.000608031 + Log[0.000161702*x3])) +
(-0.00198044 + Log[0.000588107*x1])*
(808.576*(-0.00198044 + Log[0.000588107*x1]) -
162.12*(0.00183562 + Log[0.00128038*x2]) +
0.752924*(0.000608031 + Log[0.000161702*x3])) +
(0.000608031 + Log[0.000161702*x3])*
(0.752924*(-0.00198044 + Log[0.000588107*x1]) -
37.947*(0.00183562 + Log[0.00128038*x2]) +
71.447*(0.000608031 + Log[0.000161702*x3])))*Max[0, -8665.59 + x1 + x2 + x3])/
(x1*x2*x3), {x1, 0, Infinity}, {x2, 0, Infinity}, {x3, 0, Infinity}]
Creating a table of values in the ranges of 0 to 10000 in steps of 2000 for each variable will show that the result is not implausible. So either there are bad evaluations e.g. from cancellation error in machine arithmetic, or else the underlying integral is not what was wanted. Changing the values from machine numbers to exact nearby rationals, and evaluating at high precision, I get results consistent with the machine precision results. This makes me suspect numeric error is not the issue.