glTF doesn’t change anything on the GPU memory…
This isn’t strictly true — there are techniques to reduce GPU memory usage like quantized attributes and compressed textures. But neither are official glTF features quite yet — texture compression support is actively being developed, quantized attributes may be in the future.
Regarding “JPEG for 3D”, this only is meant for file compression…
We use the phrase more broadly than that — think of PSD vs JPEG, rather than BMP vs JPEG. A photoshop file preserves lots of data that’s useful while editing a file, so it’s a “good” format, but you wouldn’t write a JS parser and embed photoshop files on your website. The format has much more complexity than is needed just to view an image, so we export to JPEG or PNG. The same is true of COLLADA and FBX — they have many features, and many ways of representing the same thing, which is all great for content pipelines but makes it hard to render consistently. By exporting to glTF you strip this down to the subset needed to render. For a concrete example, a glTF file cannot contain Ngons — the model must be triangulated at time of export.
But of course, glTF does also offer (optional) compression, which gives faster download at the cost of decompression time.