The modern tech stack is being pushed to its limits as organizations confront a new generation of high-density, high-heat infrastructure demands.

Artificial intelligence is accelerating this pressure, forcing teams to rethink everything from thermal dynamics to component layout. What was once a conversation about raw performance is now a race to engineer systems that can survive — and thrive — under extreme conditions. Hypertec Group has been building immersion-ready platforms since 2018 to meet that reality, optimizing for thermal density, tiered configurations and long-term hardware resilience, explained Hecheng Han (pictured, right), product manager of high-performance computing and AI at Hypertec.

“For the full AI stack, we’re not just talking about the server; we’re talking about the entire solution, especially when it comes to immersion,” Han said. “We have the server, the hardware on board, the fluid, as well as the tank. It’s a whole ecosystem that encompasses not just the server.”

Han, alongside Scott Sickmiller (middle), chief executive officer of Midas Immersion Cooling, and Z. George Zhang (left), vice president of R&D at Valvoline Global Operations, spoke with theCUBE’s John Furrier at SC25, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how hardware, fluid science and system design are converging to power the next decade of the AI tech stack. (* Disclosure below.)

The modern AI tech stack has moved beyond air cooling

If one message has rung clearly, it’s that air cooling can’t keep up with modern AI infrastructure demands. Air is an extremely poor thermal conductor. With global data center power usage effectiveness averages at 1.55, more than half the energy used for compute is burned just to cool it, according to Sickmiller. As racks push past 100–200 kW — and megawatt designs emerge — air becomes physically incapable of pulling heat away.

“[Water cooling] does a job on about 80% of the chassis, and then we move into the full immersion fluid or immersion cooling, which now cools the full chassis, it’s touching everything,” he said. “It’s removing all of that heat, and it’s giving an environment where the server, the AI server, the IT just wants to live. It’s clean, environmentally safe, with no oxidation and full material compatibility.”

Water usage is an often-overlooked factor. Evaporative cooling systems tied to air-cooled data centers pull massive amounts of groundwater, while immersion eliminates that dependency, according to Zhang.

“For us, the use case is how we can improve efficiency,” he said. “How do we take the heat away? When you do the AI calculation, you have the chipset generating so much heat. How do we manage that? We work with our partners, again, Hypertec and Midas, and we design the fluid, which is the lifeblood in the system.”

Valvoline’s role in the ecosystem centers on engineering the fluid — the lifeblood of the immersion system. For compatibility, every hose, connector, server board and tank component must withstand long-term exposure without degradation or chemical reaction. In terms of safety, high flashpoints must make the fluid inherently safe for data center environments. To maintain sustained performance, viscosity and flow must be tuned to match AI workloads and rising GPU heat output, Zhang emphasized.

“If we design the immersion system, we have to make sure from day one or even before day one, we have every compatibility resolved,” he said. “We have the tank from Midas. We must ensure that the hose and all components in the tank are compatible with our fluid. We have the Hypertec server. We have all these components on the server. It has to be compatible too.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of SC25:

(* Disclosure: Solidigm sponsored this segment of theCUBE. Neither Solidigm nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE