At the core the mechanism explained in this article can determine fraud with an acceptable level of accuracy.

To handle latency and heavier traffic, a machine with more processing power/GPUs would do the job.

You can also have multiple auto scaled instances of the applications deployed to cater to this scenario.