Herb Zien joins Sunny Ray to explain why the AI boom's real bottleneck is not chips but heat. Zien previously co-founded a district energy company, the largest owner of district energy systems in the United States, moving heat through downtown utility plants in eleven cities before selling for roughly $800 million to the Harvard endowment. He later applied that same physics to servers, cofounding LiquidCool Solutions, which seals server components inside a standard chassis filled with dielectric fluid, combining total immersion with the targeted cooling of direct to chip cold plates. Zien walks through why air cooling, essentially an insulator moved only by energy hungry fans, is now obsolete, citing NREL and Lawrence Berkeley Lab testing that found LiquidCool dramatically more efficient. He details a federal case study that cut rack count, space, energy, and maintenance sharply, and a utility pilot that led to a 40 server fleet order. Along the way he discusses his 72 patents, stepping aside as CEO, and why physical infrastructure founders must expect resistance from incumbents that would rather not change. Zien argues the shift away from air cooling is inevitable and already visible in the market.
Herb Zien of LiquidCool Solutions explains why AI's real infrastructure bottleneck is heat, not chips.
Why should people care about liquid cooling, and how does it connect to the AI that everyday people are using?
AI is software sitting on a hardware platform that's largely obsolete. Air cooling was always a bad idea from a physics standpoint, and it can't keep up anymore. That's why liquid cooling is becoming necessary and why it's in the headlines, because it affects water supplies and power supplies. The amount of energy and money going into data centers right now is incredible.
You spent 35 years moving heat around cities before you moved it around servers. What transferred and what didn't?
We owned the downtown district energy systems in eleven US cities, including Boston, Baltimore, Kansas City, St. Louis, and Los Angeles, running fifteen central utility plants. We sold that company because we got an offer we couldn't refuse, but also because I became concerned it wasn't the most efficient way to move heat. That got me interested in the huge inefficiency I saw in data centers, and I've spent my time since trying to solve it.
For years total immersion cooling was called impractical. What was actually wrong with it, and what did you fix?
Traditional immersion means dunking servers in a horizontal tank, an inconvenient form factor with a maintenance mess, and it doesn't target hot chips specifically. Direct to chip cold plates solve that but only remove about 70 percent of the heat, leaving two cooling systems. We combine both into one sealed chassis. NREL tested us for two years and called us the gold standard, and Lawrence Berkeley found us 30 times more efficient than the Navy's tropical climate cooling.
Why the chassis rather than the tank? What does sealing it at that level buy you?
A standard chassis in a standard rack means you can replace one server at a time using equipment you already have. In a federal agency case study, our servers replaced racks at a four to one ratio. Rack count went from 115 to 25, space dropped 70 percent, energy was cut roughly in half, and maintenance dropped 90 percent because the equipment that would otherwise be needed just isn't necessary.
You co-founded a district energy business that sold for around $800 million. What did that teach you about infrastructure businesses that a software founder might never learn?
That business had fixed assets already in place, with long term contracts because customers couldn't open their doors without us, so it was really a utility offering. LiquidCool Solutions is the opposite. We're selling equipment, not long term supply contracts, so it's a completely different kind of business even though both are infrastructure.
Everyone talks about a chip shortage. Can you make the case that it's actually a heat problem?
It's absolutely a heat problem. Air is an insulator with no thermal mass, so the only way to move it is with fans, and fans burn a lot of energy. A 2016 Facebook data center in Iowa moved 6.5 million cubic feet of air per minute between its central system and rack fans just to remove heat. That's the scale of waste we eliminate.
With 72 patents, does IP still defend anything in a market moving this fast?
The patents came out of products we developed during our science phase before we shifted to sales about three years ago, and now we sell an off the shelf catalog. We're tracking seven infringements right now. Going after infringers isn't really our business model, but the portfolio gives us a moat, and we even have patents filed in China even though they'd be hard to defend there.
What advice would you give a founder building physical infrastructure in a software-obsessed funding market?
It's expensive, partly because you're fighting large companies with inertia who don't want to change unless they have to, and eventually that catches up with them. Our approach is pilot projects. One utility bought a small rack for about $60,000, tested it in a coal fired power plant for four months, we bought it back, and they ended up ordering 40 servers for their fleet.
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