Edge Computing

A self-driving car processes 1 TB of sensor data per hour. It needs to make collision-avoidance decisions in under 10 milliseconds. Sending that data to a cloud data center 500 miles away, processing it, and sending the response back takes 50-100ms at minimum, just from the speed of light in fiber. By then, the car has traveled 3 feet at highway speed. That is too slow. People die.

An AR headset renders overlay graphics that must align perfectly with the physical world. Any perceivable latency, even 20ms, causes motion sickness. Apple's Vision Pro processes most spatial tracking locally on the M2 chip for exactly this reason. The cloud is 30-100ms away. The human brain notices mismatches above 7ms for visual-vestibular sync.

A factory robotic arm inspects parts at 60 frames per second on a production line. Each frame needs quality-control analysis and a pass/fail decision within 16ms (the time between frames). A round trip to the cloud takes longer than the gap between frames. By the time you get one answer, you have missed three parts.

Edge computing moves computation out of centralized data centers and closer to where data is generated. The destination varies: regional points of presence, 5G cell towers, on-premises servers, or the devices themselves. The goal is latency below what physics allows with centralized cloud, and local processing when bandwidth to the cloud is too expensive or too scarce.

"Edge" is not a single location. It is a spectrum of how close to the user computation happens:

Each step down the spectrum trades processing power for lower latency. The trick is putting the right computation at the right layer.

Cloud edge services you can use today:

Cloudflare Workers handles over 10 million requests per second across their network. Vercel runs Next.js middleware at the edge for every request to sites like Notion, Loom, and Washington Post.