cBioPortal takes the variant calls from the projects that they host. They just perform their own annotation for consistency.
The TCGA provisional datasets are directly from TCGA data center partly via Broad Firehose which are updated regularly. We are also actively curating datasets from literature. Studies from literature were curated from the data published with the papers. We sometimes reach out to the investigators to additional data such as clinical attributes. All the mutation data (VCF or MAF) were processed through an internal pipeline to annotate the variant effects in a consistent way across studies.