[throwaway76543]

There's no need to store anything. You have a list of hashes you'd like to reverse. You compute the entire search space and compare each.

No storage, other than the hashes you're attempting to reverse -- which for a data dump from even a large site like Yahoo wouldn't be very large at all. Megabytes.

[Retric]

Not at the same hash rate. Sure you can do binary search, but 100^14 or even 30^14 such searches is not fast. So, it's about as fast if you want 2 passwords but not 2 billion passwords.

[TheCoelacanth]

It wouldn't be a binary search, the values are already hashes, so you could do a hash table lookup very cheaply. MD5 isn't great by cryptological standards, but it is extremely robust by hash table standards.

While no extra resources might not have been strictly accurate, the lookup would be practically free compared to the time it takes to compute the hash.

[Retric]

In the example I gave the hashes don't fit on GPU's and for your hash table lookup you would need ~64GB of ram to do the hash table lookups. You can scale this across multiple machines but even the ideal case of 2 lookups to main memory * 100^14 is slow and thus expensive.