Do the cancer samples in the database include tumor exomes, or is this from germline samples only?
All of the "cancer" samples in the current release of ExAC are blood ("germline") samples from TCGA. We excluded any sample labeled as tumor. However, note that some sample/label swaps may have occurred in TCGA; in addition, it is possible that in some patients the blood samples are contaminated by circulating tumor cells.
So I believe the "ExaC non-TCGA" term comes from taking out the TCGA cohort of 7,601 samples from the total of 60,706 for population-based variant filtering purposes.
ExAC is only using normals from TCGA. The non-TCGA subset excludes those samples.
Since the TCGA normals are normal tissue from cancer patients, those samples are not really healthy. That is why you may want to exclude them. They also provide a non-psych cohort for a similar reason.