The Cooling Point: Re-Architecting Data Centers for AI - Six Five Connected with Diana Blass

AI is running hotter than ever—and air cooling is hitting its limit. How are evolving liquid cooling architectures reshaping the future of dense, high-performance AI infrastructure?

On this episode of Six Five Connected, host Diana Blass is joined by Solidigm’s Avi Shetty, and other industry experts at SC25 for their perspectives on the evolving ecosystem for AI liquid cooling at scale. They explore how direct-to-chip, hybrid precision cooling, and full immersion liquid cooling technologies are redefining data center design to address performance, energy, and space challenges.

Key Takeaways:

🔹 Liquid Cooling Is Now Essential: AI workloads are pushing heat and power density beyond what air cooling can sustainably support, making liquid cooling a core infrastructure requirement.

🔹 Storage Has Moved Into the Critical Path: Storage now feeds GPUs continuously, forcing tighter integration between compute, storage, and cooling to maintain performance and thermal stability.

🔹 Multiple Architectures, Real Tradeoffs: Direct-to-chip, hybrid precision, and immersion cooling each offer different balances of performance, deployment complexity, and total cost of ownership.

🔹 Immersion Enables Thermal Stability: Full immersion cooling delivers consistent temperatures, reduces thermal stress, and extends component lifespan by eliminating hot spots and airflow constraints.

🔹 Power and Space Are Driving Adoption: Rising energy costs and space limitations—especially in high-density regions—are accelerating liquid cooling adoption as a pillar of scalable AI infrastructure.

‍

Learn more at Solidigm.

And subscribe to our YouTube channel, so you never miss an episode.

‍

Listen to the audio here:

‍Disclaimer: Six Five Connected is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

Diana Blass:

Is the industry putting enough focus and investment towards their storage solutions right now?

‍

Avi Shetty:

Clearly not.

‍

Diana Blass:

One year later, that's changed.

‍

Avi Shetty:

We are in production. We are ramping the demands there.

‍

Diana Blass:

AI has changed the data center, pushing far more data through the system, much faster and for much longer periods of time. Take storage. What used to sit off to the side now feeds GPUs continuously. Systems becoming far more compact because AI needs data instantly. But packing more power into a tighter space creates one thing.

‍

Francesca Cain-Watson:

Chips that are running hotter, creating a lot more heat, using a lot more energy and a lot more water.

‍

Diana Blass:

Raising the question, what's the best way to cool all this?

‍

Multiple Speakers and John Griffith:

Liquid cooling. Liquid cooling. Liquid cooling. Liquid cooling. We're running an Xbox right out of the fluid here.

‍

Diana Blass:

Heat is a constant problem inside data centers. GPUs draw enormous amounts of power to function at full speed, and nearly all of that energy ends up as heat. Traditional air-cooled facilities use massive fans, HVAC systems, and chiller towers that evaporate enormous amounts of water to dial down the temperature. It's effective enough for high-performance computing and early AI, but as AI workloads grow hotter and denser,

‍

Hecheng Han:

Everything is getting more and more power hungry to the point that we're seeing that, especially with air cooling, it's starting to reach its limit.

‍

Diana Blass:

Enter liquid cooling, systems that pump liquid, usually a special dielectric fluid, into parts of the system, or in some cases, the entire system, to pull heat away. The liquid behaves like a mineral oil, safe to touch and non-conductive.

‍

Dr. George Zhang:

We take the flame on the fluid. It's not actually going to light up the fluid at all. So that's the safety factor of it.

‍

Diana Blass:

Today Valvoline provides a liquid used in DLC, hybrid and immersion cooling. The three architectures of liquid cooling, each with their own history and unique use. DLC was a first, emerging inside data centers around 2015. Instead of cooling the whole server, it targets only the hottest components, pushing liquid through cold plates mounted directly on top of the chips. The liquid absorbs heat at the source and carries it out through a closed loop. It remains the easiest and most widely adopted approach because it fits standard racks with minimal changes. And that makes it a prime space for innovation. with vendors now looking to cool more than just the chips, including new solutions that cool the SSDs themselves.

‍

Avi Shetty:

The cold plate touches the one side of our SSD, and we have thermal mechanisms inside of our SSD, which allow cooling on both sides, thus bringing the overall thermal profile down.

‍

Diana Blass:

But it also has a clear limitation.

‍

Neil Edmunds:

You still need fans and cold plates together to produce enough cooling to cool the IT down.

‍

Diana Blass:

Iseotope specializes in precision liquid cooling, which is essentially a hybrid between DLC and a full tank immersion system, which we'll explore next.

‍

Francesca Cain-Watson:

We wanted to be able to keep the customer experience so that they're able to service the systems the same way that they usually do, in a way that's very energy efficient and doesn't require any fans.

‍

Diana Blass:

It sits in the same design family as DLC because both approaches push liquid cooling inside a standard server chassis. But it differs in how the liquid touches the hardware and how it moves.

‍

Neil Edmunds:

We apply dielectric fluid into the chassis and there's about one or two inches in the bottom and that's actually circulated by small pumps through the system and to the CPUs, RAM, GPUs, storage, all the components get actively cooled by pump fluid.

‍

Diana Blass:

Liquid outperforms air because it pulls heat directly off the chip at the point of contact, instead of relying on airflow to push heat away, the way the traditional systems do. The next idea pushes this concept even further.

‍

John Griffith:

With Midas, all the fluid is in direct contact with all the board components.

‍

Diana Blass:

Instead of circulating fluid through parts of the system, Immersion dunks all components into a tank of fluid. At Supercompute, we saw examples of Immersion solutions that use Midas tanks, Hypertech servers, Valvoline fluid, and embedded storage solutions by Solidime.

‍

Hecheng Han:

So you can see on top here is where you have all of your NICs, peripherals, as well as all the components that are meant to be hot swappable, right? Like in this case, all the times E1.L drives. The other thing as well, if you look at our compute nodes, each single one of them are individual in respect to each other. So you're able to pull them out without any issue. And obviously with the bigger servers, we have a crane that we can pull up, let it trip, and then you're able to service. It's very easy to do so.

‍

Diana Blass:

By submerging all data center components, the fluid cools everywhere continuously, preventing localized spikes, and in this case, keeping SSD temperatures flat and predictable for long stretches.

‍

Alan Bumgarner:

These drives have been in the tank, I think, since Sunday night, and they've never deviated from the 5 degrees Celsius temperature. The fluid just starts to run faster, and everything stays the same temperature.

‍

Diana Blass:

That stability brings another benefit.

‍

Hecheng Han:

There's an increase in longevity of the components.

‍

Diana Blass:

Electronics fail from stress, thermal stress and physical stress. Immersion removes both. The fluid keeps temperatures steady, preventing the heat spikes and hot spots that fatigue components over time. And with no air moving through the system, there's no dust to create new heat traps.

‍

Dr. George Zhang:

One of the servers we have in our lab, we have been running for about 18 months, 100% capacity, non-stop. And after 18 months, compared to the beginning fluid, along the way, we have a lot of data, right? We chart that 18 months, no change.

‍

Diana Blass:

That level of efficiency isn't just a technical win, it's becoming a necessity. In regions like Europe, where space is tight and power costs soar, hybrid and immersion cooling methods are enabling operators to squeeze maximum performance out of every watt and every rack they have.

‍

John Griffith:

They just don't have the square mileage to be able to build new facilities up. So they're having to figure out how to utilize their existing facilities in a denser environment. The other big issue is power. The EU, especially during the wintertime, has power costs just spike very, very much higher than what we're used to in the States. So might as well save them 40% of the power utilization that an air-cooled environment would take to cool the exact same workload from that perspective.

‍

Diana Blass:

And that combination of power savings and a tiny physical footprint is opening up new market opportunities in the U.S. too. Both at the edge.

‍

Neil Edmunds:

From a lab to a workshop to a school.

‍

Francesca Cain-Watson:

And on-prem. We're hearing that customers are wanting to start bringing their compute back on-prem because of security reasons.

‍

Diana Blass:

There's no denying the cost that comes with solutions like these. After all, you're looking at new hardware that likely comes with a learning curve. But that investment can pay itself back quickly.

‍

Neil Edmunds:

So if you can save 80% of the cooling energy, you can use that saving towards higher power compute, which you can monetize. You can't really monetize the cooling aspect, right?

‍

Diana Blass:

But the big question remains, how do you know which architecture to deploy? We turned to Josh Grossman with Supermicro.

‍

John Grossman:

So most people that are going to do a greenfield, and by that I mean a new deployment, are going to be using something that's liquid cool because they see the kind of the increase in the power consumption and the TDP. and all the things that require liquid cooling. But there's a lot of folks who have legacy data centers that are air-cooled, and so they might be looking for something like the Sidecar Solution, which is liquid to air, or they might want to stay with an air-cooled solution. You basically want to make it the least cost possible to get to the client that needs the AI service.

‍

Diana Blass:

Because today, your workload truly drives your architecture. And that shift points to something even bigger. Think back to the demos we filmed. Immersion, DLC, even air-cooled racks, different cooling, different power. But everyone featured a Solidime SSD. Because modern AI architectures aren't built around a single component. They're built around the flow of data end-to-end. Signing off for Six Five Media, I'm Diana Blass.

‍