It is a bold claim but not backed up by any convincing evidence.
I can believe that classical computing and in particular NN type AI can be speeded up by making some gross heuristic approximations that are usually true. Ignoring almost irrelevant noisy information may work.
Nothing can surpass an N bit quantum computer for factoring products of impossibly long primes. The whole of modern public key cryptography is predicated on that task being well beyond present day computing power. A decent length quantum register computer could change that overnight.
Dedicated hardware can always do better than general purpose computers at specific tasks but that is a different issue altogether.
Turing's Bombe or Collosus would have beaten anything less than a 386 PC at code breaking despite them having plug boards, paper tape, relays and valve logic. They were incredibly cunning designs able to short circuit the codebreaking by ruling out big chunks of the search space.