What edge computing is — why servers move closer to users, and why AI wants them
Edge computing means doing computation on servers spread out close to users instead of in a few huge central data centers. It began with content delivery networks, which kept copies of web pages and video near viewers; Akamai, which pioneered the field, had more than 4,300 points of presence in over 130 countries at the end of 2025. Distance matters because of physics: light in optical fiber travels about 200 kilometers per millisecond, so a round trip between London and San Francisco can never take less than roughly 86 milliseconds. AI is reviving interest in the edge, because tasks that need fast answers, local data or huge numbers of short jobs can run faster and cheaper near users
The three lines
- Definition — computation on servers near users rather than in central data centers; it grew out of content delivery networks
- Why — light covers about 200 km per millisecond in fiber; London to San Francisco round trip has a floor of roughly 86 ms
- AI — suits fast responses, local data and many short agent tasks; training giant models stays in central data centers
Key questions
- What is edge computing
- **Computing close to the user.** | | Central cloud | Edge | |---|---|---| | Servers | a few giant data centers | hundreds to thousands of small sites | | Distance to user | can be another continent | same city or carrier network | | Strength | massive compute, economies of scale | low latency, local processing | | Weakness | latency grows with distance | each site is small | | Typical use | model training, big analytics | web and video delivery, security filtering, fast inference | The "edge" is the **outer rim of the network, where users are.** The two are complements, not rivals.
- Why is edge computing faster
- **Because even light has a speed limit.** | Route | Approx. distance | Round-trip floor in fiber | |---|---|---| | within a city | tens of km | under 1 ms | | New York–Washington | about 330 km | about 3 ms | | London–Frankfurt | about 640 km | about 6 ms | | New York–London | about 5,600 km | **about 56 ms** | | London–San Francisco | about 8,600 km | **about 86 ms** | Light in fiber covers **about 200 km per millisecond.** Real routes are longer and pass through equipment, so actual latency is higher. Tasks that exchange messages many times multiply the delay. **Moving the server is the only way to cut this time.** Distances are straight-line approximations.
- Why does AI need edge computing
- **Some AI work is not one big slow computation but many small ones.** | AI task | Why the edge helps | |---|---| | voice conversations, live translation | a tenth of a second feels awkward | | agent tool execution | short jobs like reading pages and running code, repeated many times | | security and bot filtering | bad traffic should be stopped near its source | | data residency | personal data should not cross borders | **Training large models** still needs tens of thousands of chips in one place, so it stays central. Anthropic's 11.6 billion dollar deal with Akamai on September 24, 2026 is for CPU capacity to run agents, not GPUs for training.
The hardest delay to remove on the internet is the time light takes to travel. Edge computing attacks it by moving servers toward users. It has made the web and video faster for more than 25 years, and AI is now returning to it for the same reason.
1. Definition — computing at the network's edge
| Central cloud | Edge | |
|---|---|---|
| Servers | a few hyperscale data centers | hundreds to thousands of small sites |
| Location | where land and power are cheap | in cities and inside carrier networks |
| Good at | large-scale computation | fast response, local processing |
| Limit | slower with distance | each site is small |
The edge is the network's outer rim, the user end. If the central data center is headquarters, the edge is the local branch: complex work goes to headquarters, quick work to the branch.
It started with CDNs (content delivery networks). In the late 1990s popular websites crashed under traffic. The fix was to copy pages and video to servers worldwide and send each user to the nearest copy. Akamai pioneered the approach; at the end of 2025 it had more than 4,300 points of presence in about 700 cities across more than 130 countries.
2. Why closeness matters — the speed-of-light wall
Light in optical fiber travels about 200 km per millisecond.
| Route | Straight-line distance | Round-trip floor |
|---|---|---|
| within a city | tens of km | under 1 ms |
| London–Frankfurt | 640 km | about 6 ms |
| New York–London | 5,600 km | about 56 ms |
| London–San Francisco | 8,600 km | about 86 ms |
| Seoul–San Francisco | 9,000 km | about 90 ms |
Real networks are slower, because cables are not straight and traffic passes through equipment. What matters more is the number of round trips. Opening a single page involves connecting, negotiating encryption and requesting files. Five round trips at 86 ms is almost half a second. Nothing but moving the server removes that time.
3. What the edge does — from delivery to computation
| Generation | Edge role | Examples |
|---|---|---|
| 1 | copy and deliver files | web images, video streaming |
| 2 | security | DDoS blocking, bot filtering, web firewalls |
| 3 | run code | short programs near users |
| 4 | AI inference and agent execution | fast-response AI, agent tool calls |
The edge began as a nearby warehouse, became a security checkpoint that filters attack traffic close to its source, and then a small computer that runs short code. It is now expanding to generating AI answers and running agents' tasks.
4. Why AI is coming back to the edge
AI computation splits roughly in two: training and inference.
| Task | Nature | Best location |
|---|---|---|
| Model training | tens of thousands of chips for months | central |
| Large-model inference | needs big GPU clusters | mostly central, some regional |
| Real-time voice, translation | 0.1 s delay is noticeable | edge |
| Agent tool execution | many short CPU jobs | distributed |
| Personal-data processing | must stay in-country | local |
Agents amplify this. One agent task can involve dozens or hundreds of short steps: read a page, open a file, run code. Each step is light, but there are many. Those steps run on CPUs rather than GPUs, and spreading them out can be more efficient than concentrating them.
That is the context for Anthropic's seven-year, 11.6 billion dollar agreement with Akamai on September 24, 2026, which supports Anthropic's CPU workload demand on Akamai's distributed infrastructure. It is the first large test of an edge company as a major AI compute supplier.
5. Limits and counterpoints
| Limit | Detail |
|---|---|
| small sites | the largest models do not fit in one site |
| hard to run | thousands of sites must be maintained and secured identically |
| cost claims | "edge is cheaper" comes mainly from providers |
| smaller models | may move onto phones and PCs (on-device AI) instead |
The edge does not replace the central cloud. The center thinks big; the edge moves fast. As AI shifts from thinking to acting, the share handled by the part that acts grows, and the industry is now measuring how large that share will be.
6. Quick answers
| Question | Answer |
|---|---|
| Is a CDN the same as edge computing? | a CDN is the first form; edge computing adds computation |
| Is it related to 5G? | "mobile edge" puts servers near cell towers |
| How is it different from on-device AI? | on-device runs on your device; edge runs on a server near you |
| Who runs edge networks? | CDN and cloud providers operate sites in most major markets |
7. What remains unconfirmed
- Latency figures are computed physical floors, not measurements.
- Akamai's footprint is as of end-2025.
- Independent cost comparisons for edge AI inference are scarce.
Sources
- Akamai — Form 10-K FY2025 (global platform: 4,300+ points of presence)
- Akamai — What Is Distributed Cloud Inference?
- Akamai — Why AI Inference is Driving the Shift from Centralized to Distributed Cloud Computing
- Akamai — Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic
- TechCrunch — Anthropic to pay Akamai $11.6 billion over seven years in cloud deal