In this episode of The Sunny Ray Show, host Sunny Ray talks with Scott Sickmiller about why immersion cooling may solve the biggest efficiency problem facing AI data centers. Sickmiller's company, Midas Immersion, began as a landlocked Austin data center operator struggling to cool dense racks, then pivoted after building and patenting its own immersion tanks in 2012. He explains how submerging servers in dielectric fluid lets facilities reach a power use efficiency near 1.05, compared to 1.4 to 1.55 for typical air cooled facilities, while also cutting water use and eliminating most air conditioning infrastructure. Sickmiller details lessons learned from early tank failures, why liquid moves heat far more effectively than air as AI racks push past 100 kW, and how nearly 4,000 tanks deployed largely in Texas, many for Bitcoin mining, prove the technology at scale. He also discusses the shift of power capacity from Bitcoin to AI workloads, partnerships with renewable fluid suppliers like Oleon, and why he believes hyperscalers will eventually adopt immersion as physics forces the industry's hand.
Immersion cooling pioneer Scott Sickmiller explains why liquid, not air, is the physics backed future of AI data center cooling.
What are you building, and why should the world care?
We live in a world obsessed with AI, and that comes with huge energy and water costs, mostly from cooling. Every time you convert energy to compute, you generate heat. Immersion lets us convert that energy far more efficiently using liquid instead of air as the cooling medium. That means less energy consumed overall and a big reduction in the water use and greenhouse gases tied to cooling infrastructure.
Midas started as a data center operator, not a cooling company. What went wrong inside your own facility that made you build a tank?
We were space and power constrained in Austin and had a niche cooling hard to cool IT, which back then meant about 25 kW a rack. We tried adopting immersion from another provider, but like a lot of early technology it did not perform the way we needed as an operator, so we built and used our own system until 2016. Ultimately rising power costs and hyperscaler competition priced us out of staying a data center operator.
What did you get wrong in the first generation of tanks that the current XCI system fixes?
Those first tanks were purchased units from a great company, but we were their first client and there were real disagreements. We thought as an operator the system should behave one way, and the originating engineers thought it should go another way. The client always wins, so we decommissioned those tanks, built our own solutions, and patented the design in 2012. It came from adopting very early technology without enough deployed field experience.
You say physics are in your favor. What is the one number that ends the argument, like liquid moving heat about 1,200 times better than air? Why is the industry still debating this?
Air was all we ever needed until recently. A typical rack used to run 8 to 12 kW, which air can dissipate, just not efficiently. Once you push past 25, 40, 50 kW, and now AI has us at 100 to 300 kW a rack, air simply cannot move enough molecules to grab that heat. Liquid is far denser, so we capture more heat and can actually reuse it, even for growing starfish for protein in one European project.
Killing the server fans takes a 1,000 watt box down to about 800. What else disappears that nobody counts?
Removing server fans gets us 15 to 20 percent efficiency immediately, but that is just the server level. At the facility level, typical power use efficiency today runs 1.4 to 1.55, meaning 400 to 500 watts wasted per kilowatt of compute on air conditioning and peripherals. In immersion, PUE is around 1.05, so only about 50 watts are wasted per kilowatt. That lets you downsize generators and UPS systems and often eliminate air conditioning entirely.
Where is the average rack today?
That is a broad question. Globally the average rack is still only about 10 to 12 kW, because most data centers are not AI focused. But AI racks today run between 50 and 120 kW, with visions of 150, 200, even 300 kW, and people in the industry are talking about megawatt racks being physically possible. AI is really the most addressable market for alternative cooling, and that is squarely the 50 to 120 kW range today.
You partnered with Oleon on fluids. Why does the fluid supply chain decide who wins this market?
We are fluid agnostic, with 15 approved providers including Oleon, and the fluid is really where the rubber meets the road. We handle the physics and the container, but if the fluid does not have the right material compatibility and thermal absorption at the server level, none of it works. What makes Oleon interesting is it is one hundred percent renewable, plant based, an ester, rather than coming from the petroleum chain like many other fluids.
What do you believe about data center cooling that the hyperscalers would argue with you about?
That their existing roadmap is not future proof. Immersion is different today, but eventually an immersed server environment will become part of the accepted reference architecture, because physics is on our side. At some point states like Texas or Pennsylvania will require data centers to prove they are using the most efficient methods and taking less water from the community. We already have a large opportunity in Alabama moving toward immersion as the accepted standard.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interview