true, for extremely large rainbow tables. SHA1 tables are around 20-60GB depending on how large your base character set is. If you shoved all this data into a giant database, query speed is still under a few milliseconds. In general, rainbow tables can be sharded fairly easily, so if your data set is a few hundred terabytes, just split it across a few machines and you'll retain the millisecond query times. Storing and querying easily partitioned data will usually be faster than a brute force calculation.
Calculating it is like saying you want to find the fibonacci number for any given N, and you have a really fast processor to calculate it to that N, but if you just persisted pre-calculated values up to C, you'd only need to calculate N-C hashes. So even if you are bruteforcing the password, it is still faster to have rainbow tables up to a certain length.
What I say is true for any size of rainbow table. It seems you forget that RT lookups require CPU resources in addition to mere I/O resources. There is always a number of hashes beyond which brute forcing them is faster than RTs. Sometimes this number is very high (billions of hashes), sometimes it is lower (thousands of hashes). It depends on many factors: RT chain length, speed of the H() and R() functions, speed of the brute forcing implementation, etc.
To take your example of a small SHA1 rainbow table of 20GB, assuming it has a chain length of 40k, looking up a hash in it will require on average 200M calls to the SHA1 compression function (assuming a successful lookup). A modern CPU core can do about 5M calls per second. Therefore looking up one hash will take at least 40 sec, and looking up these 6.5M LinkedIn hashes would take 8.2 years! (This is just counting CPU time, I assume the RT is loaded in RAM for a negligible I/O access time to its data.) A RT of this size would cover a password space of about 2^44. For comparison a decent GPU can brute force this many hashes concurrently at a speed of roughly 500M per second (see oclhashcat perf numbers on an HD 7970). Covering the same password space would take only 9.8 hours. Compare 8.2 years vs. 9.8 hours: obviously the LinkedIn hashes that have been cracked so far have been brute forced, not looked up in RTs!
And even if you leveraged GPUs to perform RT lookups, they would speed up the computations by roughly a factor 100x, reducing the 8.2 years down to 30 days, still unable to match the short 9.8-hour brute forcing session. (My friend Bitweasil is doing research on GPU-accelerated rainbow tables, see cryptohaze.com)
Calculating it is like saying you want to find the fibonacci number for any given N, and you have a really fast processor to calculate it to that N, but if you just persisted pre-calculated values up to C, you'd only need to calculate N-C hashes. So even if you are bruteforcing the password, it is still faster to have rainbow tables up to a certain length.