Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Enabling Drone Package Delivery at Scale
Transcript
- Lucas: So earlier this year, Wing Aviation — which is Alphabet's drone delivery subsidiary — hit a milestone that barely made headlines outside the trade press. They passed 400,000 commercial deliveries in the US alone. That's up from around 100,000 in 2023. But here's the thing nobody talks about: none of those deliveries would work without edge computing. Luna: Because the drone can't just phone home to the cloud for every decision, right? It has to process a lot locally. Lucas: Exactly. The latency window for obstacle avoidance — if a drone is flying at 30 miles per hour and needs to dodge a bird or a power line — is under ten milliseconds. Round-trip to a cloud server, even on a good 5G network, you're looking at 20 to 30 milliseconds minimum. That's the difference between a safe maneuver and a crash. Luna: And I imagine the compute requirements don't stop at collision avoidance. They've got to handle route optimization, wind adjustments, battery management — all in real time. Lucas: Right. And this is where the edge architecture becomes really specific. Most people think of edge computing as what CDNs do — cache content closer to users. But drone delivery needs a different kind of edge: compute that can run multiple neural networks simultaneously, handle sensor fusion from cameras and LiDAR, and do it on a device that draws under 50 watts because it's battery-powered. Luna: So it's not just about location; it's about power efficiency. The chip choice matters a lot. Lucas: Hugely. And I want to focus on one company that I think illustrates this better than most — Zipline. You might know them for medical deliveries in Rwanda, but they've since expanded into retail and food delivery in the US. Their Rwanda operation, which has been running since 2016, is a masterclass in edge computing for logistics. Luna: What's the setup? I heard they use a distributed network of launch sites, not just a central hub. Lucas: That's right. In Rwanda, Zipline operates from multiple distribution centers, each equipped with a local edge server. These servers handle everything: flight path planning, weather data ingestion, real-time telemetry from the drones, and the coordination of multiple drones in the same airspace. The central cloud in Kigali handles fleet-level analytics and inventory management, but every time-critical decision happens at the local node. Luna: And that's been running for almost a decade? That's pretty impressive. What kind of hardware are we talking about? Lucas: They've evolved over time, but the early systems used Intel Xeon processors with NVIDIA GPUs for inference. Today they're likely on something like an NVIDIA Jetson platform — those are designed for exactly this use case: embedded, low-power, but capable of running multiple AI models. Each distribution center's edge server can process data from up to 20 drones simultaneously, with a local latency of under two milliseconds. Luna: So the edge node is basically the air traffic controller. That's a lot of responsibility for a box in a warehouse. Lucas: It is. And speaking of responsibility, let me step back for a second. I want to make a point about why this show exists — we cover edge computing in-depth because we believe it's going to underpin so much of the infrastructure we rely on. If today's conversation gave you something usable, a new way to think about compute architecture, honestly, if it was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. That's what keeps this podcast ad-free and lets us drill into these technical details without chasing clicks. Luna: Yeah, I think it's a good reminder that real-world engineering decisions — like how much compute to put on the edge versus the cloud — have massive cost and safety implications. And we only get to explore those because listeners support the show directly. Lucas: Exactly. So back to Zipline. One of the interesting challenges they faced early on was that Rwanda has a lot of hills and changing weather. The edge server had to process regional weather models and update flight paths in real time, because a storm could blow in over a ridge in minutes. That's not something a centralized cloud could handle quickly enough. Luna: And I assume they also have to handle communication between drones. If two drones are approaching the same drop point, the edge node coordinates the handoff. Lucas: Exactly. The edge server assigns time slots and altitude bands so that multiple drones can operate in the same airspace without conflict. It's essentially a very localized version of what the FAA does for commercial aviation, but running in real time on a single server. Luna: So how does this scale? If you go from 20 drones per node to 200, does the architecture change? Or can you just add more compute at the edge? Lucas: That's the million-dollar question. For now, most commercial drone delivery operations are still relatively small. Wing's 400,000 deliveries happened over years, across multiple cities. But as the industry scales, you hit a limit. The edge node can only handle so many simultaneous drones before the compute budget runs out. At that point, you need to either distribute the computation across multiple edge nodes — which means designing a local mesh — or you offload some non-time-critical tasks back to the cloud. Luna: So the edge-cloud split becomes dynamic, not static. That's a much harder engineering problem. Lucas: Much harder. And it's where things get interesting from a networking perspective. Today, most drone delivery systems use a cellular link — 4G or 5G — to communicate between the drone and the edge server. But if you have a mesh of edge nodes, you might need low-latency links between them. That starts to look like a private 5G network or even a wireless mesh. And that's expensive. Luna: Right, because you're not just buying servers; you're buying spectrum or leasing private network capacity. That changes the unit economics completely. Lucas: Exactly. And that's where the comparison to CDN edge breaks down. A CDN edge node serves static content — images, video, web pages — to potentially millions of users. The throughput is high, but the compute per request is low. A drone delivery edge node serves a handful of drones, but each drone requires intense, low-latency computation. So the cost per operation is much higher. Luna: So the business model has to account for that. Which is probably why we're not seeing drone delivery in every city yet. The infrastructure is just not cheap enough. Lucas: Right. And that brings me to another example that I think illustrates the edge computing challenge in a different way. There's a company called Flytrex that's doing drone delivery in North Carolina and Texas. They use a slightly different architecture — instead of a fixed edge server at a distribution center, they rely on a combination of on-drone compute and a cloud-based decision engine, with LTE for connectivity. Luna: So they push more compute onto the drone itself. That's a different trade-off — more expensive drones but potentially cheaper ground infrastructure. Lucas: Exactly. Their drones have onboard processors that handle the real-time control and obstacle avoidance, while the cloud handles route planning and fleet management. The latency is a bit higher, but they say it's acceptable for their current operating environment — suburban backyards, not dense urban airspace. Luna: So it's a tiered approach. The more complex the airspace, the more you need local compute on the ground. For simpler routes, you can get away with on-drone compute and cloud. Lucas: Exactly. And that's probably how the industry evolves — a spectrum of architectures depending on density and regulatory requirements. But here's the thing: every architecture eventually runs into the problem of coordination. If you have multiple drone operators in the same airspace — say, Amazon, Walmart, and Zipline all operating in the same city — who manages the edge nodes? Do they share a common infrastructure? Or does each operator build their own? Luna: That's a huge question. It reminds me of the early days of cellular networks, when each carrier built their own towers. Eventually they started sharing. I could see a similar pattern here — neutral host edge infrastructure that serves multiple drone operators. Lucas: That's exactly what some startups are betting on. There's a company called Edgegap that's building a distributed edge network specifically for latency-sensitive applications — gaming, but also drone coordination. And there are telcos like Verizon and AT&T that are positioning their 5G edge compute as a platform for drone operators. The idea is that you pay for compute by the millisecond, and the edge node is managed by the network operator, not the drone company. Luna: That makes a lot of sense economically. But then you have to trust that the edge node is reliable and secure. If a single node goes down, you could have drones in the air with no coordination. Lucas: Which is why redundancy is built into the system. In Zipline's case, each distribution center has a backup edge server that takes over if the primary fails. And the drones themselves have onboard fallback — if they lose connection to the edge for more than a few seconds, they autonomously return to a designated safe zone. Luna: So the edge isn't a single point of failure, but it's still critical infrastructure. This is where I think the comparison to cloud computing falls short. Cloud outages are annoying — you can't access your email. Edge outages in drone delivery could mean property damage or worse. Lucas: Exactly. And that's why the reliability requirements for edge compute in logistics are much higher than for, say, a CDN. A CDN node can go down and traffic reroutes to another node with no real-world consequences. A drone edge node going down means you have to ground every drone in that airspace immediately. Luna: So we're talking about five-nines reliability, maybe even six-nines. That's expensive. And it's one reason why drone delivery has been slower to roll out than the hype suggested. Lucas: Absolutely. But the pieces are coming together. Edge hardware is getting cheaper and more powerful. NVIDIA's Jetson Orin, for example, can run multiple vision transformers in real time at under 15 watts. Cellular networks are adding low-latency features. And regulators are starting to approve beyond visual line of sight operations, which is the real bottleneck. Luna: And we're already seeing the results. Wing's 400,000 deliveries, Zipline's 500,000 plus across Rwanda and Ghana — these aren't pilot programs anymore. They're production systems. And they all rely on edge computing in one form or another. Lucas: Right. And I think the next five years will be about standardizing the architecture. Right now, every company builds its own edge stack. But eventually, we'll see common protocols for drone to edge communication, shared edge infrastructure, and maybe even a standard for edge node hardware. That's when the industry really scales. Luna: It's a fascinating space. And a reminder that edge computing isn't just about faster Netflix — it's about enabling entirely new categories of autonomous systems. Lucas: Exactly. And that's what we'll keep covering on this show.