[EJCK3HEJOVOF]

NAT Gateway and a bunch of our other fancy networking gadgets are based on a thing publicly called Hyperplane, and at the end of the day it's just EC2 instances. I'm skeptical that we'd ever use T series instances for this kind of thing for multiple reasons, but there are efforts to move internal systems to Graviton.

The reason for the hourly cost is a bit subtler than that, but we are working on the cost of this stuff.

(Source: Work at AWS.)

[RexM]

Thanks for chiming in, I had no idea that was backed by actual EC2 instances and the charges make more sense now.

So that means all external traffic in a private subnet with a NAT Gateway is routed through a single instance? I’ll have to read up more on Hyperplane.

[nodesocket]

Can you say what instance types are currently used for NAT gateways then? m5.large?

[EJCK3HEJOVOF]

I'm a little removed from where I'd need to be to answer that very precisely, and probably shouldn't anyway.

They're not small- these are large, multitenant fleets handling huge numbers of NAT gateways all at once. The system has several layers, that scale on different dimensions and have different requirements. And those "requirements" can be very weird.

The part that actually moves packets around is probably mostly full-size c5n's, for the bandwidth.