
A notebook by Saurav Chowdhary
Compute in the Wild
Where edge computing runs, what it costs,
and who’s making it work.
The shifting center of computing
Mainframes, personal computers, and cloud platforms each changed where computing happens. Edge computing puts some of that work near the devices producing the data, especially when sending it elsewhere would take too long or use too much bandwidth.
- 01
Mainframe
Centralized
1960s–1970s
Shared processing, accessed through terminals.
- 02
Client–server
Distributed
1980s–2000s
Work is shared between personal computers and servers.
- 03
Mobile–cloud
Centralized
2000s–2010s
Mobile applications draw on shared cloud infrastructure.
- 04
Edge intelligence
Distributed
2020s onward
More sensing, inference, and action happen near devices.
In his 2017 a16z Summit talk, Peter Levine described a shift toward distributed computing. His model places sensing, inference, and action at the edge, with learning and storage in the cloud.
Read The End of Cloud Computing ↗Peter Levine, The End of Cloud Computing · a16z Summit 2017 talk · Published November 15, 2019.
I start with the task, then ask where it needs to run.
Compute, closer to the work
Edge computing places processing and storage near the people, devices, and operations that use them.
Processing might happen inside a camera, on a factory server, or in a nearby network facility. The location depends on response time, connectivity, and operating requirements.
Time to respond
Processing data locally can shorten the time between detecting an event and responding to it.
Data movement
Filter or analyze data close to its source, then send what is useful elsewhere.
Continuity
Keep essential functions running during an internet outage, provided they can operate without remote services.
Local constraints
Account for data handling rules, network access, space, and power.
Where edge computing runs
Computing can run on a device, at a business site, or in a provider’s network. Each location has different limits on capacity, response time, and who manages the equipment.
Premises edge
In a factory, a gateway can connect machines and translate their protocols. A local server can run inspection models, databases, and site applications. They have different jobs, even when one appliance combines them.
Ask: What must keep working if the site loses its internet connection?
What runs at the edge
Edge systems can run business applications, process sensor data, or run AI models. The examples below show how those tasks might be divided between a device, a site, and central systems.
Follow the work
A camera captures each part
The camera sends images of each part to a nearby server for a quality check.
The question I’d askHow quickly must a defect be spotted before the part moves on?
Illustrative examples
Device → Site → Central systems
| Setting | Work at the edge | What to evaluate |
|---|---|---|
| Manufacturing | Machine monitoring and visual inspection | Response time, integration, production uptime |
| Retail | Store applications and inventory analysis | Service continuity, privacy, support across sites |
| Healthcare | Local processing of device and imaging data | Data handling, availability, clinical validation |
| Transport | Onboard perception and telemetry processing | Connectivity gaps, power, environmental conditions |
| Energy | Remote asset monitoring and anomaly detection | Bandwidth, maintenance access, resilience |
I’m interested in where AI helps and where simpler software is enough.
When AI runs where the data is
Edge AI runs models near the data they analyze. Examples include checking camera images for defects, recognizing speech on a device, and detecting unusual machine vibrations.
Running a model locally can help when a remote service is too slow, unavailable, or unsuitable for the data. Training, updates, and analysis across sites can still happen centrally.
A site can also run databases, application services, and rule-based processing without using AI.
What makes it worth deploying?
Before deploying, I’d ask these five questions about the benefit, cost, and day-to-day operation.
- 01Outcome
What needs to improve?
Define the operational problem and how you’ll measure the result.
A measurable target - 02Local need
Why here?
Identify the response-time, continuity, or data-handling need that makes local processing useful.
A reason to run locally - 03Cost
What will it cost over time?
Include equipment, software, integration, connectivity, power, support, and replacement.
A lifecycle cost estimate - 04Operations
Who keeps it working?
Assign responsibility for monitoring, security, updates, recovery, and on-site service.
A named operations owner - 05Scale
What changes at 100 sites?
Test provisioning, configuration, workload updates, and recovery across a fleet.
A repeatable rollout
About this space
I’m Saurav. I work in enterprise tech, GTM, and partnerships. Here, I explore edge computing and the strategy behind it.
Explore Insights →