Your VM host simply must instantly inject unique entropy into each VM guest instantly on creation, duplication and start. Otherwise you're breaking your own security.
On a recent Linux environment those problems are solved however. Use proper random drivers for your virtualized environment and use getrandom() in your application.
When you duplicate any VM guest, all sensitive key material should be replaced instantly (preferably never copied to begin with, the host should generate and inject fresh key material).
Refilling the entropy pool alone isn't enough when reuse-insecure ciphers like stream ciphers are actively used, for the reasons stated above. Some ciphers are more tolerant of reuse (see AES-GCM vs AES-GCM-SIV).
How? And why 128 only?
The acceptable safe maximum length isn't quite infinite, it's basically 2^64 rounds (not necessarily bits or bytes) for 128 bits of entropy. But that's basically arbitrarily long in practice as anyone cares.
So with original entropy x, evaluate G(x) => x', c. Output the random bits c and repeat with x = x'. The sequence of c is psuedorandom.
k=128 is the key length of the generator. We define an efficient program to be one that runs in some time polynomial in k. How big of a k you need is based on how powerful computers are at present.
So then the question becomes, what's the largest file you can encrypt with a single key? And the answer is any file you want. There's practically no size limit when you're encrypting files.
128 bits is usually the minimum recommended key size, not because of anything to do with the amount of randomness you can produce, but because you always have worry about an attacker guessing your key by brute force. Brute forcing a key is the same as just trying all possible keys to find one that works. (If there's another way to brute force the key, again you have a bad cipher.) We can estimate how big a key needs to be by looking at the fastest computers that exist: https://en.wikipedia.org/wiki/Supercomputer The fastest computer in the world does something like 2^55 operations per second. There are 2^25 seconds in a year, so if we generously assume that checking a key is just one operation, then a supercomputer like that could check about 2^80 keys per year. 128 is just the first nice power of 2 that's much larger than 80.
Often (e.g. in the Signal protocol), a short secret key is extended using a HKDF (RFC 5869) to create enough bits for a symmetric key and an IV or a nonce (and a MAC key if needed).
So for randomness, take 128 bits, use HKDF to extend to 192 bits, use 128 bits of output as AES-128 key, use next 64 bits of output as nonce, use 64 bits counter, produce AES-128(key, nonce||counter) for every value of counter until it overflows as randomness. 2^64 of 2^4 byte blocks is 2^68 bytes of randomness - enough for most purposes. An Intel chip will produce it at 1GB/sec per core.
Though there will still be people doing it for various reasons, one of them "because we can".
1. Seed itself reliably.
2. Do nothing unless seeded.
3. Re-seed itself regularly, for forward secrecy.
4. Have good performance.
How difficult is #1? It's as simple as gathering 128 bits of entropy. If you have 128 bits, you an expand them into an infinite sequence of random numbers, securely, forever.
I think this is the essence of your comment. The topic is important, so it's worth being clear.