He was wrong then and now, but the crucial speedup of Mathematica vs Maxima was difference in modeling of data and how Mathematica could piggy back on fortran libraries for numerical solving, while Maxima focused on symbolic solving.

This meant that for many cases where numerical method was beneficial, Mathematica easily won on speed.

The predecessor of Mathematica, Wolfram and Cole's SMP, was 10-100x faster than the best Lisp based symbolic math programs at the time on what Wolfram and Cole considered to be small to medium research physics problems. On large problems they were unable to compare because the Lisp based programs needed more memory than the VAX 11/780 at Caltech High Energy Physics had.

That's hilarious, because the Wolfram Language is actually quite lispy. Or, at least you can write it in a very lispy style.

Now Wolfram Language uses LLVM framework for compilation. SBCL-compiled Maxima is also quite fast .

That's probably because Stephen Wolfram in the 80s knew little about Lisp. Chances are he still knows little about it today!