Serverless computing promises developers a world where hardware is irrelevant. Write your function, deploy it, and let the cloud handle the rest. It is a compelling abstraction—but it rests on an inconvenient truth: serverless is still powered by servers. And as computing moves toward the network edge, physical infrastructure is becoming more important, not less.
The Serverless Paradox
The cloud has done an excellent job of hiding complexity. Virtual machines, containers, and now serverless functions let developers focus on application logic rather than rack-and-stack operations. Google’s Kelsey Hightower, among others, has pointed out with characteristic wit that this abstraction does not make the underlying hardware disappear—it just relocates the concern to the cloud provider.
For most workloads, that trade-off is entirely reasonable. But a new class of applications—real-time IoT, autonomous vehicles, distributed machine learning, augmented and virtual reality—cannot tolerate the round-trip latency of sending every request to a centralized data center hundreds of miles away. At that point, the abstraction breaks down, and physical geography re-enters the conversation.
Why the Speed of Light Is an Engineering Constraint
Don Duet, who served as co-head of technology at Goldman Sachs before joining edge computing startup Vapor IO, framed the problem bluntly: centralized data centers are simply too far away for latency-sensitive workloads. Light travels through fiber at roughly two-thirds the speed of light in a vacuum. Even that theoretical maximum means a round trip from a device to a data center 500 miles away and back introduces measurable delay—delay that is unacceptable when an autonomous vehicle needs to react in milliseconds or a real-time sensor network must respond instantly.
Duet’s conclusion: the only engineering answer is to push cloud capabilities out to the edge of the wireless network, adjacent to the devices generating and consuming data.
“Edge computing has sparked a lot of interest as the Internet undergoes a massive transformation to support new, low-latency applications, including real-time IoT, distributed machine learning, and autonomous vehicles. These applications require cloud capabilities that can respond within milliseconds, and centralized data centers are too far away. The speed of light is too slow. The only solution is to push cloud capabilities out to the edge of the network.”
Don Duet, Vapor IO
Cloud Becomes a Multi-Tier Fabric
In this model, the cloud is not a single location—it is a tiered fabric stretching from a central data center through regional hubs down to micro data centers at the very edge of the wireless network. Each tier handles the workloads best suited to its latency profile and compute capacity:
- Central data centers handle batch processing, long-term storage, model training, and workloads where latency is not critical.
- Regional data centers aggregate traffic, cache content, and serve as coordination points between the center and the edge.
- Edge nodes run latency-sensitive workloads directly adjacent to end devices, with direct cross-connections to the wireless network.
Duet argues that realizing this architecture requires securing physical real estate—literally. Vapor IO partnered with Crown Castle, a shared wireless infrastructure company, to gain access to more than 40,000 cell tower locations and over 60,000 miles of metro fiber. That kind of footprint cannot be conjured through software; it requires long-term agreements over physical land, conduit rights, and urban infrastructure.
What True Edge Cloud Actually Means
Major cloud providers have entered the edge space. AWS Greengrass and Microsoft Azure IoT both extend cloud functionality toward the network edge. But Duet’s critique, at least as of the time this discussion was published, was that these approaches remain too centralized in practice—they focus on edge gateways and device management while resolving IP addresses and workloads back in central office locations.
A genuine edge cloud, by contrast, would provide:
- Full cloud workload execution at the physical edge location, not just data forwarding.
- IP addresses presented and handed off at the edge node itself, not resolved back at a central location.
- Elastic, automatically provisioned compute capacity at the edge—not fixed-capacity gateway hardware.
- Direct cross-connection to the wireless network, with seamless tiers of service back to regional and central data centers.
- The reliability, security, and redundancy expected of production cloud infrastructure.
This is a significantly higher bar than simply placing a small server near a cell tower and calling it edge computing.
Micro Data Centers and the Hardware Reality
To meet these requirements, Vapor IO developed what it calls “Vapor Chambers”—cylindrical micro data centers approximately nine feet in diameter that house 130 to 160 kilowatts of compute capacity and are fully self-contained and remotely operable. They are designed to be deployed at or near cell tower sites without requiring a dedicated building or on-site personnel.
The design philosophy reflects a broader point: as much as the industry talks about abstracting hardware concerns away from developers, the physical servers, fiber runs, and real estate deals that underpin even the most “serverless” workloads do not disappear. Someone has to own them, maintain them, and position them correctly relative to the end user.
Implications for DevOps and Platform Engineering Teams
For DevOps practitioners, the rise of genuine edge computing introduces a new tier of infrastructure to reason about. Several practical considerations follow:
- Deployment topology matters again. Where your workload runs affects end-user latency in ways that choice of programming language or container runtime does not. Edge-aware deployment pipelines need to place workloads at the correct tier automatically.
- Observability becomes more complex. Monitoring a distributed fabric of edge nodes, regional data centers, and central clusters requires consistent telemetry pipelines that can function across all tiers, including locations with intermittent connectivity.
- Configuration management at scale. Managing hundreds or thousands of remotely operated edge nodes demands infrastructure-as-code discipline and strong GitOps or policy-based configuration management.
- Security surface increases. Edge nodes deployed in semi-public locations like cell tower bases present physical and network security challenges that a walled data center does not.
- Latency SLOs become first-class requirements. For IoT and real-time applications, latency service level objectives belong alongside availability and throughput in system design conversations from day one.
The Bottom Line
Serverless and cloud-native abstractions are powerful and genuinely reduce cognitive load for application developers. But those abstractions are built on, and increasingly constrained by, the physical world. The next major frontier in cloud computing—edge infrastructure capable of supporting autonomous vehicles, real-time IoT, and sub-millisecond response applications—is as much a civil engineering and real estate problem as it is a software problem.
As Don Duet put it, the land grab for the next generation of computing literally involves land. For engineers who assumed hardware was someone else’s problem, edge computing is a useful reminder that physics and geography are constraints that no abstraction layer fully eliminates.
This article draws on reporting originally published by TechRepublic and statements made by Don Duet, co-founder of Vapor IO, circa 2017.