Edge computing has been “next year” for about a decade. AI inference may finally make it this year, but not in the way the original pitch decks promised. The real Edge is less about tiny data centers at cell towers and more about latency, interconnection, and who controls the path between a user and a model.
In Back to the Future Part II, 2015 came with hoverboards and flying cars. Edge computing has had a similar relationship with the future. Since roughly 2016, it has been perpetually about to arrive: autonomous cars, smart cities, and AR glasses were all supposed to need compute at the base of every cell tower. For years, Edge computing was the Half-Life 3 of infrastructure: endlessly anticipated, never quite shipped.
Most of that demand never materialized at scale. So, what is different now?
The short answer is AI inference. Training a model can happen anywhere with enough power. Running it, especially for real-time applications, is a different matter. When an AI agent handles a phone call, a vision model inspects parts on a production line, or a system analyzes live video, every millisecond between the user and the model shows up in the experience.
Why latency suddenly matters again
Physics sets the floor. Light in optical fiber covers roughly 200 kilometers per millisecond, so a round trip to a data center 1,000 kilometers away costs about 10 milliseconds before any routing, queuing, or processing. Add a few network hops and a busy GPU, and a “fast” response can drift past what a human conversation tolerates.
Three use cases make the point:
- Industrial AI: Machine vision, robotics, and quality control need fast, predictable responses, often on private 5G or campus networks.
- Video: Analyzing camera feeds centrally means hauling huge volumes of video across the network. Processing closer to the source cuts both latency and transport costs.
- AI agents: Voice agents live or die on turn-taking. A pause of a second feels like a dropped call.
None of this means every workload belongs at the Edge. As we explored in Edge Computing vs Cloud: Where Should Your Workloads Actually Live?, most applications are perfectly happy in a regional Cloud. The Edge earns its keep only where latency, bandwidth, or data locality makes the difference.
Interconnection is the real Edge
Here is the twist that took the industry years to accept: for most use cases, the useful Edge is not a shoebox data center at a tower. It is a well-connected metro data center where networks meet.
That is where interconnection lives. Peering lets networks exchange traffic directly instead of routing it through third parties. Internet exchange points (IXPs) concentrate hundreds of networks in one place. Cloud on-ramps give enterprises private, low-latency links into hyperscaler and regional Cloud platforms.
Europe is unusually strong here. Exchanges in Frankfurt, Amsterdam, London, and a growing list of secondary cities connect huge numbers of networks, and industry association Euro-IX tracks a dense, distributed IXP ecosystem across the continent. Put inference capacity in a facility with that kind of connectivity, and you are within a few milliseconds of tens of millions of users without building anything at a cell site.
The economics favor this model, too. A metro facility with strong interconnection can serve many customers from shared infrastructure, with proper power and cooling for accelerators. A thousand micro sites at the base of masts each need their own power upgrades, security, and maintenance visits, and most of them would sit underused. For AI inference, which needs GPUs that are expensive, hot, and scarce, consolidation into a few well-connected metro hubs is simply easier to finance. The “far Edge” still has a role in factories, ports, and campuses, but there it usually means on-premises equipment on a private network rather than a public Edge Cloud.
Telcos own the last mile; Clouds own the platforms
Edge is a team sport because nobody holds all the Infinity Stones.
Telcos own the access networks, the radio spectrum, the metro fiber, and the real estate close to users. They also hold deep relationships with enterprise customers and governments. What most of them lack is a developer-friendly Cloud platform that people actually want to build on.
Cloud providers, hyperscale and regional, own the platforms, the developer ecosystems, and the AI services. What they lack is the last mile.
The partnership models that have emerged reflect that split:
- Hyperscaler in the telco network: Cloud zones placed inside operator facilities, such as AWS Wavelength with Vodafone in Europe, bring Cloud services physically closer to mobile users.
- Telco as AI Cloud provider: Deutsche Telekom and NVIDIA opened their Industrial AI Cloud in Munich in February 2026, positioning a carrier as a sovereign AI platform for industry rather than just a connectivity supplier.
- Network as an API: Through the GSMA’s Open Gateway initiative, operators expose network capabilities such as quality-on-demand and location verification as standardized APIs that developers can call.
Who captures the value? That is the question that makes partnership dinners awkward around dessert. Telcos fear being reduced to “dumb pipes” while platforms collect the margin. Cloud providers fear depending on telco timelines and pricing. The deals that work tend to split the stack clearly: the telco monetizes guaranteed connectivity, location, and local presence, while the Cloud monetizes compute and services, with joint go-to-market for vertical customers.
The European angle
In Europe, the Edge debate is tangled up with two others: sovereignty and network investment.
On sovereignty, the Commission’s proposed Cloud and AI Development Act would make European control a procurement criterion for public sector Cloud. If inference for public services must run on sovereign infrastructure, the network path to that infrastructure matters too, and European carriers are natural partners.
On investment, the Commission proposed the Digital Networks Act in January 2026 to harmonize rules and encourage operators to scale across borders. Notably, the proposal did not include the “fair share” contributions from large traffic generators that some operators had lobbied for. The bill is now with the Parliament and Council, with BEREC set to play a stronger coordinating role.
The practical upshot: European telcos will need to earn their place in the AI value chain through products and partnerships rather than through regulation.
What to do now
For telcos and carriers:
- Map the metro sites where your network, power, and existing data center space overlap. Those are your Edge inference candidates.
- Package connectivity with compute for specific verticals (manufacturing, logistics, healthcare) instead of selling generic “Edge.”
- Expose network capabilities as APIs that developers can actually use.
For Cloud and data center providers:
- Treat interconnection density as a product feature. Measure and publish latency to key population centers.
- Partner with carriers for last-mile guarantees on latency-sensitive offerings.
- Place inference capacity in secondary European metros where demand is growing and power is easier to find.
The Edge was never really about where the server sits. It was always about how fast, how reliably, and how sovereignly the bits get there.
Telcos, IXPs, and Cloud providers all meet at one address in March. Join them at CloudFest Global, March 15–18, 2027. Registration is open.
