Everyone's requirements are different, and everyone can expect different levels of concurrency and throughput from a single grid node, depending on their test plan complexity and of course how well their target site performs under load.
For that reason, we recommend you first run your load test on a single grid node in a geographic region that is close to your target site. We do have recommended numbers of users per node, but this can be higher or lower depending on the test plan tool, complexity and target site performance. To find this number for your specific circumstances, run a few tests, increasing the number of users you are running gradually while watching your grid's resource utilization to ensure that CPU and memory utilization aren't constantly above 80%.
We call that process baselining a single node. Once you have an idea of a single node's performance, you can then plan to scale out with as many nodes in as many grids or regions as you require.
A single node's performance is representative of a node in a grid of up to 90 nodes. We've designed grid nodes to be autonomous, loosely coupled, and shared nothing, which affords you the greatest unhindered scale when planning your load tests.