The problem is not the transistor size, but the transistor cost.
It is easy to make transistors as small as on silicon on most other semiconductors, but the cost of the final product would be many times greater.
One important reason is that silicon is made into huge 12-in wafers, while most other semiconductors are made into small 2-inch to 4-inch wafers, like for silicon several decades ago. At each processing step on a silicon production line one machine processes 9 to 36 times more transistors than if another semiconductor were used. Large wafers also waste much less area when making big dies.
In general, for silicon there exist very big processing machines with very high productivity, while for other materials there is little difference between lab equipment and what can be used for commercial production. An integrated circuit made on a non-silicon material also requires more processing steps and more expensive materials.
In order to keep increasing the performance of CPUs and GPUs, the replacement of silicon with another semiconductor is unavoidable, perhaps in a decade from now. However, that will be done only after all other possibilities of improving silicon devices will be completely exhausted, in order to avoid the increase in production costs.
> It is easy to make transistors as small as on silicon on most other semiconductors
Yikes, [citation needed] here. No, it absolutely is not. All the etch and litho chemistry is highly specific to the substrate and dopants. You can't just feed a germanium crystal through an etcher tool in a TSMC fab and get anything but a brick out the other side.
I'm not aware of anyone anywhere doing low-nm lithography on anything but silicon (even in a demo context, or even announcing plans for the capacity), but I'm willing to be educated.
Nobody has done low-nm lithography on anything but silicon, on integrated circuits.
The reason is that such integrated circuits could not compete in cost, so developing all the fabrication equipment for them would not be worthwhile.
On the other hand, on special discrete devices, e.g. microwave transistors, and on experimental devices, "low-nm" (which means tens of nm for the most advanced devices) has been done for a long time.
The low fabrication yields, which would be unacceptable for the mass production of integrated circuits, have much less relevance for small and expensive discrete devices and for experimental devices.
It is easy to make transistors as small as on silicon on most other semiconductors, but the cost of the final product would be many times greater.
One important reason is that silicon is made into huge 12-in wafers, while most other semiconductors are made into small 2-inch to 4-inch wafers, like for silicon several decades ago. At each processing step on a silicon production line one machine processes 9 to 36 times more transistors than if another semiconductor were used. Large wafers also waste much less area when making big dies.
In general, for silicon there exist very big processing machines with very high productivity, while for other materials there is little difference between lab equipment and what can be used for commercial production. An integrated circuit made on a non-silicon material also requires more processing steps and more expensive materials.
In order to keep increasing the performance of CPUs and GPUs, the replacement of silicon with another semiconductor is unavoidable, perhaps in a decade from now. However, that will be done only after all other possibilities of improving silicon devices will be completely exhausted, in order to avoid the increase in production costs.