The Virtualization Tax
Think of a traditional hypervisor like a property manager who lives in one of the apartments they manage. They take up a unit, consume utilities, and are always present — even when tenants don't need them. In cloud computing terms, this "property manager" is the Hypervisor: the software layer that creates and manages Virtual Machines (VMs). On traditional infrastructure, this hypervisor runs on the same physical CPU as your application, consuming 10–30% of the machine's compute capacity just to handle network packets, storage I/O, and security enforcement. You pay for that capacity whether you use it or not. This is the virtualization tax.
For most of cloud computing's early history, this was accepted as an unavoidable cost of doing business. Every EC2 instance you launched was quietly sharing its CPU with AWS's own management software. A c4.large with 2 vCPUs wasn't really giving you 2 full vCPUs — it was giving you 2 vCPUs minus whatever the hypervisor needed at that moment.
The AWS Nitro System
AWS decided this wasn't good enough. Starting around 2017, they reinvented the EC2 architecture from the ground up to create the Nitro System — a purpose-built platform that physically moves all virtualization overhead off the main CPU and onto dedicated hardware.
The goal: Zero virtualization tax.
Instead of software running on the host CPU to manage networking and storage, AWS designed custom silicon cards (ASICs) that plug directly into the server. These cards handle all the infrastructure work. When you launch a c5.large today, the Intel or AMD CPU on that host is running only your code. AWS's management plane doesn't touch it.
This is why modern Nitro-based instances consistently outperform their predecessors on benchmarks — not because the CPUs got faster, but because the CPU is no longer being shared with the landlord.