The lossy JPEG compression does not merely remove small coefficients in higher frequencies. It encodes them with a precision relative to a (relatively crude) visual perception model; most notably, horizontal and vertical frequencies are not quantized with the same precision. And as in many compression formats, it essentially assumes that the data is locally stationary.
If you apply the DCT over the whole image, and you quantify DCT coefficients, this quantization will effect the whole image. Imagine an image with a background checkerboard pattern, and a small zebra in foreground. With a whole-DCT compression, the zebra is likely to loose its stripes, because their energy is negligible with respect to that of the checkerboard. Even more, as JPEG applies DCT on chrominances as well, with down-sampling, color coefficient quantization is likely to produce false colors at places where they do not belong. With block size larger than about $8 \times 8$, several meaningful image gradients (parts of object edges) are more likely to happen simultaneously in the same patch.
Meanwhile, one can get slightly better results with $16\times 16$ blocks. That is for the perceptual part.
There is a computational part too. Small non overlapping blocks are easier to process, and require less memory, a still expensive part of electronic devices. Early JPEG 2000, applying wavelets on the whole image, have failed adoption despite better results, partly because of their memory footprint. Now, with the advent of GPUs, processing blocks is becoming quite attractive. Last, a power of two ($8=2^3$) and clever tricks make the block DCT very efficient. In modern coders, standard often use different $2\times 4$, $4\times 4$... block sizes.
Last, but not least, it is not modular, as you need DCT implemetations for each image size. Although it is possible with software, chip makers are unlikely to like the hardware part.
So, applying DCT over the whole image:
- is detrimental to non-stationary images, and waster some local orientation,
- can be costly in terms of memory, and hardware,
- is not quite modular nor computationally efficient.
Thus being said, very stationary images would be better compressed with whole-size DCT.