# Why Chip Designers Are Adopting Liquid Cooling

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 16 Jun 2026, 6:56.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0054.mp3

Episode 54 of The Hardware Podcast dives into the quiet shift from fans to liquid cooling in high-performance chips. Lucas and Luna explore why a single AI accelerator now draws over 1000 watts, how direct-to-chip liquid cooling works, and what the transition means for data center design, energy costs, and future chip architectures. Using specific examples from recent hyperscaler deployments and chipmaker roadmaps, they unpack the thermal limits that are forcing the industry to rethink decades-old cooling assumptions. #LiquidCooling #ChipCooling #DataCenter #ThermalManagement #AIAccelerators #DirectToChip #Hyperscaler #EnergyEfficiency #HeterogeneousIntegration #AdvancedPackaging #PowerDensity #CoolantDistributionUnit #SinglePhaseCooling #TwoPhaseCooling #ImmersionCooling #Technology #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: If today's tech conversation gave you something usable, you might appreciate what one listener said to me last week: they asked why we don't talk more about how chips actually stay cool. And they're right — we've covered design, packaging, lithography, but the thing that physically keeps billions of transistors from melting is something most of us take for granted.
Luna: And what's interesting is that the answer used to be simple — a fan and a heatsink — but that's changing fast.
Lucas: Really fast. We're at a point where a single high-end AI accelerator can draw over 1000 watts. Think about that — a chip consuming more power than a hair dryer, and all that energy ends up as heat inside a package smaller than your hand. Traditional air cooling just can't move that much heat fast enough.
Luna: So the industry is turning to liquid. But this isn't your car radiator. The kind of liquid cooling we're seeing in data centers right now is direct to chip — a cold plate mounted right on the processor, with coolant flowing through microchannels that snake directly over the hotspots.
Lucas: Exactly. And it's not just experimental anymore. Google, Microsoft, Amazon — all the hyperscalers are deploying direct liquid cooling at scale. The coolant is usually a dielectric fluid — it doesn't conduct electricity, so if it leaks, it doesn't short out the board. And it flows through a coolant distribution unit, or CDU, that manages the temperature and pressure.
Luna: There's a key number here: the thermal design power, or TDP. A typical server CPU today might have a TDP of 200 to 300 watts. Air cooling handles that fine. But once you cross about 400 or 500 watts per chip, the fans get enormous, the noise goes up, and the airflow just isn't enough.
Lucas: And we're well past that. NVIDIA's latest GPU modules — the ones powering the biggest AI clusters — are rated at 700 to 1000 watts per accelerator. Some estimates put the next generation at 1500 watts. So liquid cooling isn't optional anymore. It's a requirement.
Luna: And it's not just about raw power. It's about density. A single rack of servers with air cooling might handle 20 or 30 kilowatts. With direct liquid cooling, that same rack can handle 100 to 150 kilowatts — because the heat is carried away by liquid instead of air.
Lucas: That density changes the economics of data centers. A few years ago, most data centers were built around air-cooled designs. Now, new builds are including liquid cooling infrastructure from day one. The retrofit market is growing too — companies are installing CDUs and cold plates in existing facilities.
Luna: One thing I find fascinating is that there are actually two main approaches: single-phase and two-phase cooling. In single-phase, the coolant stays liquid all the way through. In two-phase, the coolant boils inside the cold plate, absorbing latent heat, then condenses back to liquid outside the server.
Lucas: Two-phase is more efficient because it uses that phase change to carry away more heat per volume of coolant. But it's also more complex. You need precise pressure control, and the system has to handle vapor. Most of the hyperscalers are using single-phase for now, but two-phase is getting a lot of R&D attention.
Luna: And there's a whole other category — immersion cooling. Where you literally dunk the entire server into a tank of dielectric fluid. It's not as common for high-density compute yet, but some crypto mining operations and a few HPC clusters use it.
Lucas: Immersion is great for simplicity — no pumps or cold plates on each chip — but it makes maintenance tricky. You can't just swap a drive without pulling the whole board out of the tank. For hyperscalers doing rapid hardware iteration, direct to chip seems to be the sweet spot.
Luna: Quick honest thing — a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many of these episodes possible. No ads, no sponsors, just people who find the deep dives useful. If that's you, it makes a real difference.
Lucas: Yeah, it's a small group that keeps the whole thing ad-free. And we appreciate it. So back to cooling — one of the less obvious challenges is reliability. Pumps and valves have moving parts, and in a data center with thousands of cooling loops, failures happen.
Luna: Right. A single pump failure can take down an entire rack if the chip temperatures spike. So the industry is building redundancy — dual pumps, leak detection sensors, and automated shutoffs. The coolant itself has to be maintained, too. It degrades over time, absorbs particles, and can grow bacteria.
Lucas: There's a whole supply chain emerging for coolant quality management. Companies like CoolIT and Boyd Corporation make CDUs, and the fluid manufacturers — like 3M with their Novec line, though they exited the PFAS market recently — are developing new dielectric fluids with lower environmental impact.
Luna: That's a big issue. Some of the older fluids are fluorinated compounds with high global warming potential. The industry is moving toward more environmentally friendly options, but it's a trade-off between performance and sustainability.
Lucas: And the chip designers themselves are adapting. When you have liquid cooling, you can push the chip harder — higher clock speeds, more voltage. But you also have to design the chip's thermal interface material, or TIM, to work with the cold plate. That's a whole other optimization.
Luna: There's a great example from a recent IEEE paper: a research team showed that by optimizing the microchannel pattern in a cold plate, they reduced thermal resistance by 15 percent compared to standard designs. That's a big deal when you're trying to squeeze out every watt of performance.
Lucas: And that's the kind of incremental improvement that compounds. So the bottom line is: liquid cooling is no longer a niche technology for supercomputers. It's becoming mainstream for any chip that draws more than about 400 watts. And with AI models growing, that threshold is being crossed more and more.
Luna: So the next time you run a prompt through a large language model, just remember: somewhere in a data center, there's a cold plate with fluid flowing over a chip that's burning through a thousand watts to give you that answer.
Lucas: And that fluid is quietly carrying away the heat of the most compute-intensive workload humanity has ever built. It's a fascinating piece of infrastructure that most people never see.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.