True randomness is hard. (Technically, impossible on a deterministic machine.)
If you really need to generate a megabytes of random bits, there are ways to do it locally (of varying quality). Unix systems provide /dev/random, which is cryptographically secure.
You probably won't be able to draw megabytes from /dev/random in any amount of time anyway (unless the configuration is screwed with) as entropy is difficult to collect on your average system because estimates are conservative and there aren't *that* many entropy sources in the system. If you really need "real" random bits, there are small, fairly affordable USB devices that use signal noise or quantum physics to produce large amounts of it. Google for "true random number generator" or "hardware random number generator" (HRNG).
But like others have said above, you seem to be misguided. The output of cryptographically secure random number generators is designed to be computationally indistinguishable from a stream of "truly random" bits. And when you think about it, that's all you need; doesn't matter if the output bits aren't "really" random, there is no test anyone without the original key can perform in reasonable time to detect that they aren't! So is there really any observable, measurable difference? And they are much cheaper to generate (since you only need a few dozen truly random bits in total to seed the CSPRNG and generate a virtually infinite pseudorandom stream of bits) which is the whole point.
In fact, and I like to point this out to people, the probability that someone somehow manages to distinguish between the output of a CSPRNG and a truly random stream of bits is actually far, far lower than the probability of your hardware number generator failing and producing correlated, non-random bits (and also lower than the probability of the computer generating the CSPRNG output failing). So there is in any case a physical limit to how reliable a process can be made to be, and CSPRNG's happen to fall below that threshold.