Cloud computing runs workloads in large, central data centres. Edge computing runs them on or next to the device that creates the data. Fog computing is the layer in between, such as a gateway or on-site server, and in 2026 most vendors simply call it the "near edge".
Updated October 2026
What changed in 2026
- Edge spending keeps growing. IDC says worldwide edge computing spending reached $265 billion in 2025 and forecasts about 15% annual growth to $450 billion by 2029 (IDC, February 2026).
- Edge hardware can now run generative AI. NVIDIA's Jetson AGX Thor, available since August 2025, offers up to 2,070 FP4 teraflops in a 130 W envelope, 7.5x the AI compute of Jetson Orin (NVIDIA).
- Cloud IoT services are shrinking, not growing. Google Cloud IoT Core retired on 16 August 2023 (RCR Wireless). AWS ended support for AWS IoT Analytics on 15 December 2025 (AWS docs). Azure IoT Central stopped new app creation on 23 September 2026 and ends on 20 September 2029 (Microsoft Learn).
- Telecom edge standards are maturing. ETSI's Multi-access Edge Computing group runs Phase 4 from 2024 to 2026, covering constrained devices and distributed edge networks (ETSI).
The short version
Think of three rings around a machine. The innermost ring is the device itself, or a box bolted next to it. That is the edge. The middle ring is a server in the plant, store or cell tower. That used to be called fog. The outer ring is a hyperscale data centre. That is the cloud.
Data moves outward. Decisions that must happen in milliseconds stay close. Work that needs huge storage or lots of history moves out.
Edge vs fog vs cloud: side by side
| Edge | Fog (near edge) | Cloud | |
|---|---|---|---|
| Where it runs | On the device or a gateway next to it | Local server, on-site cluster, telco site | Central hyperscale data centre |
| Typical latency | Milliseconds or less | Low milliseconds | Tens to hundreds of ms, network-dependent |
| Compute and storage | Small, power-limited | Moderate | Effectively unlimited |
| Works offline | Yes | Usually | No |
| Best for | Real-time control, safety, vision on the line | Aggregation, local AI, site dashboards | Training models, long-term analytics, fleet management |
| Example tech | NVIDIA Jetson, industrial PCs, smart cameras | Azure Local, AWS Outposts, on-prem Kubernetes | AWS, Azure, Google Cloud regions |
Latency ranges are typical, not guarantees. Your own network decides the real figure.
What is edge computing?
Edge computing processes data where it is created. A camera checks a weld. A sensor flags a vibration spike. A robot decides where to grip a part. None of these can wait for a round trip to a distant region.
Edge brings three benefits. It cuts latency. It cuts bandwidth, because you send results rather than raw video. And it helps with data residency, because sensitive data can stay on site.
The trade-off is scale. Edge devices have limited power, memory and cooling. You also have many of them, so updating and monitoring them is hard.
What is fog computing?
NIST defined fog computing in 2018 as "a layered model for enabling ubiquitous access to a shared continuum of scalable computing resources", placing fog nodes between devices and the cloud (NIST SP 500-325). The same document calls the thinnest layer, on microcontrollers, "mist computing".
Cisco researchers popularised the term in a 2012 paper (Bonomi et al., 2012). The OpenFog Consortium pushed it as a standard until it merged into the Industrial Internet Consortium in January 2019 (Edge IR).
Today the idea lives on, but the word has faded. Most product docs say "near edge", "on-prem edge" or "edge cluster" instead. If you read "fog" in an RFP, it usually means the site-level layer.
What is cloud computing?
Cloud computing rents compute, storage and services from a provider's data centres. It is still where most AI models are trained and where fleet-wide data is stored.
The cloud is also moving outward. AWS Outposts and Microsoft's Azure Local (formerly Azure Stack HCI) run cloud-managed infrastructure on customer sites. So the line between "cloud" and "fog" is now often one control plane.
Edge vs fog: is there a real difference?
In practice, very little. Both run outside the central cloud. The useful distinction is distance from the device:
- Far edge: on the machine or within metres of it.
- Near edge (fog): on the same site or campus, serving many machines.
Design by the decision, not the label. Ask how fast it must happen, what happens if the network drops, and where the data may legally live.
When to use each layer
| Decision you need | Put it at | Why |
|---|---|---|
| Stop a machine if a person enters a zone | Edge | Must work in milliseconds, even offline |
| Spot a defect on a moving line | Edge | Video is too heavy to ship; delay means scrap |
| Compare five lines in one plant | Near edge (fog) | Needs data from many devices, still local |
| Retrain a vision model on a year of images | Cloud | Needs large GPU clusters and storage |
| Roll out a new model to 300 sites | Cloud control plane, edge runtime | Central policy, local execution |
How AI changes the picture in 2026
The big shift is that models now run at the edge, not just sensors. Small language models and vision models fit on edge GPUs. That turns the edge from a data collector into a decision maker.
It also creates new work. Each site now runs models that drift, need updates and must be audited. That is an operations problem as much as a hardware one. Read more in our guide to MLOps for edge AI.
Where it's heading: 2026 → 2030
| When | What it looks like |
|---|---|
| Around 2020 | Edge collects data; cloud does the thinking. "Fog" still used as a separate tier (OpenFog/IIC era). |
| 2026 (now) | Generative and vision AI run on edge hardware such as Jetson Thor. Hyperscalers retire cloud-only IoT services (Google IoT Core, AWS IoT Analytics, Azure IoT Central). Global edge spend about $265B in 2025 (IDC). |
| By 2028 (expected) | One control plane manages cloud, near edge and far edge as one fleet (our view). |
| By 2030 (expected) | IDC forecasts $450B in edge spending by 2029. "Fog" disappears as a term. Teams plan a single edge–cloud continuum, placing each AI decision where latency, cost and regulation dictate (our view). |
How NexaStack fits
NexaStack deploys, updates and monitors AI models and agents across edge sites from one place. You set policy centrally; each site runs locally, even when the link drops. Every decision an edge model makes is logged and governed, so you can trace what ran, where and why. See also the NexaStack Edge AI solution and our guide to physical AI.
Frequently asked questions
Will edge computing replace cloud computing?
No. Edge handles fast, local decisions. The cloud still trains models, stores long-term data and manages fleets. Most 2026 architectures use both.
What are examples of edge computing?
Vision inspection on a production line, a robot planning its grip, a smart camera counting people in a store, and a gateway that shuts a valve when pressure spikes.
What companies use fog computing?
Few use the word today. The same pattern, a site-level server between devices and cloud, is common in manufacturing, energy, retail and telecom, often sold as Azure Local, AWS Outposts or on-prem Kubernetes.
Is Netflix edge computing?
Partly. Netflix caches video on servers inside internet providers' networks (Netflix Open Connect) so streams start close to viewers. That is content delivery at the edge; the main service runs in the cloud.
What are the latest edge vs cloud computing trends in 2026?
AI inference is moving to the edge on hardware like NVIDIA Jetson Thor. Hyperscalers are retiring cloud-only IoT services. And IDC expects edge spending to reach $450 billion by 2029.
Plan your edge-to-cloud architecture
Map which AI decisions belong on the device, on site and in the cloud, and see how NexaStack runs them from one control plane. Book a free edge-to-cloud architecture review →
