It is possible to provide true randomness from a chip, but it's probably cheaper not to.
http://spectrum.ieee.org/semiconductors/processors/behind-in...
http://en.wikipedia.orghwiki/RdRand
I suspect they just didn't consider RNG quality as a big factor in the usefulness of the design. TPM pre-dates the recent advances in on-die TRNG design. Plus the chip tends to live on a low-speed bus on the system which makes it inherently a bit crappy as a place to keep a high-bandwidth application like an RNG.
Personally I'd trust the random data supplied by a third party web service less than the pseudo-random data that the software on my local machine produces. Especially if you mix in data from a TPM, especially if you mix in data from an Entropy Key as well.