Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Making Real-Time Drone Package Delivery Possible
Transcript
- Lucas: When you order a package online, the clock starts ticking the moment you click 'buy'. But the part of that journey that's hardest to speed up — the last mile — is also where edge computing is starting to have its most visible impact, especially with drone delivery. Luna: And when you say visible, you mean literally watching a drone drop a package in your backyard. That's the dream, right? But the tech behind it is intense. Lucas: Exactly. And if today's tech conversation gave you something usable — a new angle on edge or just a clearer picture of how this stuff works — a couple of dollars a month is genuinely what keeps these episodes going. Buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah, it really does make a difference. And we're not talking big pledges, just the coffee equivalent. Lucas: So back to drones. The key challenge is that a drone doesn't have the luxury of pinging a cloud server a thousand miles away every time it needs to make a decision. It has to decide now — is that a bird, a power line, or a landing pad? Luna: And that's where edge computing comes in. You're putting the compute on the drone itself, or on a local node near the launch site. Lucas: Right. The company that's arguably furthest along with this is Zipline. They started in Rwanda delivering blood and vaccines, and now they're operating in the US and Japan. As of 2025, they'd completed over one million autonomous deliveries — and every single one of those flights runs on edge compute. Luna: One million is a big number. What kind of compute are we talking about on the drone itself? Are we talking a full GPU, or something more modest? Lucas: More modest than you might think. Zipline's drones use a system-on-module with a dedicated neural processing unit — similar to what you'd find in a higher-end smartphone, but hardened for flight. The key is that the onboard AI model is optimized for a narrow task: identifying landing zones and obstacles in real time. Luna: So it's not running a general-purpose vision model. It's a specialized model that's been trained specifically for aerial views, with a lot of edge cases like shadows, glare, and sudden weather changes. Lucas: Exactly. And that model runs inference locally on the drone, so the latency between the camera capturing an image and the flight controller adjusting a propeller is under ten milliseconds. Compare that to even the best 5G connection to a cloud server — you're looking at maybe 20 to 30 milliseconds round trip, and that's optimistic. Luna: And in that extra time, the drone could have already hit a branch or drifted off course. So edge isn't just nice to have here; it's safety critical. Lucas: Exactly. But it's not just the drone itself. Zipline also uses edge nodes at their launch sites — think of them as mini data centers in a box. Those nodes handle things like route planning, fleet coordination, and processing data from multiple drones at once. They're running on standard server hardware with GPUs, but they're hardened for outdoor conditions. Luna: That makes sense because you can't rely on a stable cloud connection in every location, especially in rural or developing areas where these drones are often flying. Lucas: Right. In Rwanda, Zipline's distribution centers are often in areas with limited internet infrastructure. So the edge node at the center handles all the critical computations locally. It communicates with the drones over a dedicated radio link, not the public internet. That's a classic edge architecture: compute close to the action, with its own network. Luna: And what about the actual package delivery mechanism? The drone has to know exactly where to drop the package — that's a real-time decision too. Lucas: Yes. Zipline uses a tether and a winch system for deliveries. The drone hovers at altitude, and the package is lowered on a line. The edge compute on the drone has to identify the precise drop zone — often a small pad or a designated spot at a hospital. The onboard camera feeds into the edge AI, which recognizes the landing marker and controls the winch to release at the right moment. Luna: So if the wind shifts or someone walks into the drop zone, the drone has to adjust in real time. That's a lot of computation happening in a tiny device flying 80 meters up. Lucas: It is. And it's not just Zipline. Wing, which is owned by Alphabet, uses similar edge compute for their drone deliveries in Australia and the US. Their approach is slightly different — they use a hybrid model where some processing happens on the drone and some on a ground-based edge server that's within a few kilometers. The latency is still very low, but they offload heavier tasks like multi-drone coordination to the ground. Luna: So there's a design trade-off: do you put more compute on the drone, increasing weight and power draw, or more on the ground, which requires robust local infrastructure? Lucas: Exactly. And the answer depends on the use case. For long-range deliveries, you might want more onboard compute because you're out of range of ground nodes. For urban deliveries with lots of obstacles, you might rely more on ground-based edge to handle complex path planning. Luna: The regulatory side also pushes toward edge. Aviation authorities like the FAA want to see that the drone can operate safely even if it loses communication with the ground. So the onboard edge compute has to be capable enough for a 'lost link' scenario — return to home or land safely without any external input. Lucas: That's a great point. In fact, that's one of the reasons edge adoption in drone delivery has accelerated. Regulators require a certain level of autonomy, and autonomy requires edge compute. You can't have a drone that only works when the cloud is available. Luna: But all of this edge hardware isn't cheap. What's the cost picture? Is it scalable beyond medical supplies and premium deliveries? Lucas: The cost of edge compute modules has been dropping fast. A capable module that two years ago cost $500 can now be had for under $200. And the drones themselves are getting cheaper. Zipline's newest platform, the Platform 2, was designed with cost in mind — they use off-the-shelf components where possible. Luna: Still, the total cost of a drone delivery system — including the edge nodes, the drones, the launch infrastructure — is probably in the hundreds of thousands per site. That's not cheap, but if you compare it to the cost of building roads and maintaining a fleet of vans for last-mile delivery in remote areas, it can be competitive. Lucas: And the volume is starting to justify it. Zipline's one million deliveries is a milestone, but they're projecting two million by the end of 2026. As volume scales, unit economics improve. The edge compute cost per delivery is already very low — we're talking cents per flight. Luna: What about the energy draw? Edge compute on a drone is battery powered, and battery life is already a constraint. Lucas: It's a real constraint. The NPU on Zipline's drones draws about 10 watts during active inference. That's not nothing, but it's manageable — the total flight time is usually 20 to 30 minutes, so the compute energy is a small fraction of the propulsion energy. And the models are getting more efficient every year. Luna: I've also seen some companies using neuromorphic chips for edge AI on drones. Those promise even lower power — like milliwatts instead of watts. Lucas: That's a frontier. Intel's Loihi 2 and some startups are working on that. But for now, conventional NPUs are doing the job. The interesting thing is how the entire stack — from the silicon to the software — is being optimized for this specific use case. Zipline actually develops their own AI models and compiles them specifically for their NPU to get every last bit of performance. Luna: So it's not just about buying a chip off the shelf. There's a whole software pipeline to make it work. Lucas: Exactly. And that's where the edge computing value proposition really shines: when the hardware and software are co-designed for the task. Drone delivery is one of those applications where the constraints are so tight — latency, weight, power, cost — that you have to optimize everything. Luna: And we're still in the early innings. Most people have never seen a drone deliver a package, but that's changing fast. What do you see as the next big milestone? Lucas: I think the next milestone is beyond visual line of sight operations at scale. Right now, most drone deliveries are within visual range of a pilot or observer. When regulators allow widespread beyond visual line of sight flights, the reliance on edge compute will only increase — because the drone will have to handle even more scenarios autonomously. That's when the real volume will kick in. Luna: And when that happens, edge computing won't just be an enabler; it'll be the backbone of a whole new logistics infrastructure. Lucas: That's exactly right. The edge is what makes the drone smart enough to fly itself safely, and that's what makes drone delivery viable.