Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / Why Edge Compute Needs a New Cooling Architecture
Transcript
- Lucas: So the conversation around edge computing usually centers on latency, bandwidth, and security. But there's a quiet bottleneck that's actually harder to solve than any of those: heat. Luna: Heat? I mean, servers get hot, sure. But isn't that a solved problem from the mainframe days? Lucas: It is — for a traditional data center with raised floors, precision cooling, and a dedicated HVAC team. But edge nodes live in places like a cell tower hut, a retail back room, or a factory floor. You don't have twenty tons of cooling with a backup chiller. Luna: So what's the typical thermal load in one of those edge locations? Lucas: A single edge server rack can pull anywhere from two to ten kilowatts. And the enclosure might be a nema rated cabinet with zero active cooling. If ambient hits 40 degrees Celsius and the equipment is dumping heat, you get thermal throttle within minutes — or a hard shutdown. Luna: And that defeats the whole purpose of having compute at the edge if it's not reliable. Lucas: Exactly. So the industry is starting to look beyond fans and passive heat sinks. direct to chip liquid cooling and immersion cooling are moving from hyperscale data centers into these distributed environments. Luna: Immersion cooling — where you dunk the whole server in dielectric fluid. That sounds expensive and messy for a five-kilowatt edge node. Lucas: It can be. But the cost argument shifts when you factor in reliability and space. A liquid-cooled edge node can run at full load in a sealed cabinet with no dust, no humidity issues, and virtually no noise. Nokia and Intel have been piloting immersion-cooled edge servers for 5G base stations. Luna: And what about the CDN side? Those Points of Presence are everywhere — in every major city, often in shared colo space. Lucas: CDNs are actually the early adopters here. They have thousands of edge nodes worldwide, many in older colocation facilities with limited power and cooling capacity. A company called Iceotope has deployed chassis-level liquid cooling in CDN nodes that reduces cooling energy by up to 90 percent. Luna: Ninety percent? That's massive for the total cost of ownership. Lucas: It is. And it lets them double the compute density in the same footprint. So instead of one server per cabinet, you can put two — which is crucial as edge AI inference starts demanding more GPU power. Luna: That's actually a great segue. If today's conversation about edge cooling gave you something useful, the reason this show stays ad-free is listener support. You can keep it going at buy me a coffee dot com slash fexingo. Lucas: Yeah, it's a small gesture that makes a big difference for us. Helps us keep digging into these niche infrastructure topics. Luna: Alright, back to the heat. Lucas, you mentioned AI inference — how does that change the cooling equation compared to a typical edge workload? Lucas: It's a step change. A standard edge server doing data aggregation might pull 200 to 300 watts per CPU. Add an NVIDIA Jetson or an Intel Movidius for inference, and you're at 500 to 700 watts per node — but the heat density is concentrated in a small chip. That's where direct to chip liquid cooling shines. Luna: So you're cooling the chip directly rather than the whole cabinet. Lucas: Right. A cold plate sits on the processor, and a coolant loop carries the heat to a small radiator on the cabinet. No fans needed. The whole thing fits in a weatherproof enclosure that can sit outside. Luna: That sounds like the future of edge deployment. But what's the catch? I'm guessing cost per node is still higher than air cooling. Lucas: It is — about 20 to 30 percent more upfront. But the total cost of ownership over three to five years often comes out lower because you avoid downtime, extend component life, and dramatically cut the power used for cooling. And in remote locations where human intervention is expensive, reliability is the priority. Luna: So the tipping point comes when you can't afford a truck roll to replace a fried server. Lucas: Exactly. That's why oil and gas companies, mining operations, and telecoms are among the first adopters. They'd rather pay a premium upfront for a sealed liquid-cooled system that runs for five years without a site visit. Luna: What about the standardisation piece? Are there industry specs emerging for edge cooling? Lucas: The Open Compute Project has a group working on an edge cooling specification. They're defining standard liquid interfaces, coolant types, and form factors so that different vendors' gear can share the same cooling loop. It's early, but it's moving. Luna: That would be a game-changer. Right now, every vendor has its own plumbing. Lucas: It is a fragmented mess. But the big cloud providers — AWS, Microsoft, Google — are pushing standardisation because they want to deploy edge hardware at scale without custom cooling for each site. Luna: So we're at the point where cooling is becoming a strategic decision, not just an operational one. Lucas: Absolutely. And the choice you make today — air-cooled, direct to chip, or immersion — will determine what workloads you can run, where you can place them, and how much they cost to operate. It's a fundamental architectural decision. Luna: And one that doesn't get nearly enough airtime in the edge computing conversation. Lucas: Which is exactly why we wanted to cover it. Next time you see a cabinet on a street corner or a rack in a retail store, think about what's happening inside — it's getting hot in there. Luna: And if they're smart, it's getting liquid-cooled.