Bulk data dumps for each species are available via http here.
The data are in HDF5 format. HDF5 is ideal for storing and and accessing large and complex datasets. Within each of the HDF5 files, there will be three objects. "Bigmatrix" is the main data object and contains the gene expression data. In this type of data, the rows represent SRA runs, and the columns represent genes. The other two objects are the column names and row names. Changing format to HDF5 provides a few key benefits. Firstly, the old format was computationally very expensive to convert from long to wide format for downstream analysis. HDF5 allows for random access without loading the entire object into memory. This will be a great benefit for end-users who want to analyse subsets of the bulk data. Also, HDF5 occupies less disk space. For kallisto counts (floating point), the saving is ~25%, but for STAR counts it can be up to 75% disk space saved. This change will allow us to continue adding samples to the database without the need to purchase additional cloud storage.
Corresponding metadata can be obtained here via http here. The files with the ".tsv" suffix are obtained from SRA and describe each run. The files with the ".tsv.cut" are a reduced metadata, which just contains the corresponding accession numbers and QC summary.
Data for studies with more than 4 runs are packaged as zip files and available here.