At Sunday’s Startup & VC Reception one theme swallowed the room whole: data center power and cooling. Founders from around the world are trying to solve the biggest data center struggle and its biggest environmental concerns all in one pass.
And they just might.
That might sound like a side quest compared to the usual AI conference chatter about models, agents, and who just raised an amount of money best described as “small-nation-ish.” It wasn’t. In that room, power and cooling felt like the main plot.
The crowd was several hundred deep, mostly founders and investors, and the founder open mic leaned heavily toward companies trying to solve the infrastructure bottlenecks behind the AI boom.
That wasn’t just a stage pattern. Nearly every conversation I had throughout the evening circled back to the same pressure point, and not just from a U.S. perspective. Whether people were talking about the U.S., Australia, Sweden, or Greenland, the question was basically the same: how do you power and cool the next wave of AI infrastructure fast enough to matter?
Publicly, AI is still sold through chatbots, copilots, and demos that make it look like the future arrives through a clean little text box. But in rooms like this one, the obsession is increasingly the physical layer underneath it all. The glamour has moved downstream. The bottleneck is no longer just better models. It’s whether the grid, the cooling stack, and the power architecture can keep up with the appetite of the machines we’re building.
That framing lines up with the broader market reality. We’re already in the phase of the AI buildout where power constraints are shaping where and how data centers get built, and where the industry’s biggest ambitions increasingly sound less like software roadmaps and more like utility planning documents.
The hyperscaler arms race is still about compute, obviously. But compute now drags a giant tail behind it: substations, transformers, thermal management, backup systems, permitting, and a lot of engineers who are suddenly much cooler at parties than they were two years ago.
A few companies captured that shift especially well. Some from the Founder’s Open Mic hosted by the NVIDIA Inception Program, others from conversations I had over the course of the event.
- Zap Power was focused much more directly on the power-and-cooling problem inside AI infrastructure itself. The company is building a cryogenically cooled DC power architecture for AI data centers, with the pitch that reducing energy loss at the source can ease both electricity demand and the downstream cooling burden. Its materials frame that as a foundational infrastructure play for the AI era, arguing that conventional AC-based power chains are becoming a bigger liability as rack densities climb.
- Scalvy was focused on a much nearer-term constraint: power electronics. Its pitch centers on modular, distributed power conversion for AI data centers and other high-demand environments, replacing bulkier centralized architectures with what it calls intelligent “Power Neurons.” Translation: the less glamorous but extremely real machinery layer that determines whether more power can be delivered efficiently, reliably, and in a footprint that doesn’t eat the very rack space companies are trying to monetize with GPUs.
- Reliability Engine came at the problem from the cooling side, with a focus on reliability and uptime in advanced liquid-cooling environments. That may sound narrow, but it hits one of the most important shifts in AI infrastructure right now: liquid cooling is no longer a weird edge-case conversation. It is rapidly becoming core infrastructure, which means the industry now needs better systems for monitoring, maintaining, and preventing failures before they turn into expensive chaos. In other words, cooling has graduated from facilities problem to boardroom problem.
- Bluefin ESI represented the grid-facing side of the story. The company works across electric power delivery markets, including utility and large-load environments, which made it a natural fit for a room increasingly preoccupied with what happens when AI demand collides with real-world electrical systems. As GPU clusters scale, the challenge isn’t just generating enough power in theory. It’s delivering it cleanly, predictably, and without turning the grid into a stressed-out raccoon. That’s where firms like Bluefin start to matter a lot more.
- IPE Solutions spoke to a different but equally important layer: the operational reality of sustainable data-center design. Its public positioning is around sustainability consulting, liquid-cooling guidance, operator education, and risk reduction; basically, helping data center teams avoid getting sold a shiny cooling story that doesn’t survive contact with deployment. In a market suddenly full of urgency and vendor noise, that kind of practical translation layer feels increasingly valuable.
- Project Ohm brought one of the more globally resonant ideas in the room: use AI workloads themselves as a flexible energy instrument. The company is building a decentralized, energy-aware compute platform designed to run AI and HPC workloads next to stranded or surplus renewable energy, dynamically scheduling jobs around availability and price. It’s a very 2026 idea: don’t just feed AI from the grid; let AI infrastructure adapt to the grid’s weirdness. That helps explain why conversations about Australia, Scandinavia, and other energy-rich regions all kept converging on the same thing. The next infrastructure gold rush may be about who can turn mismatched energy geography into usable compute.
The deeper signal here is that AI’s next wave of startup activity is starting to look less like “another model wrapper” and more like picks, shovels, pumps, pipes, and power-routing software.
Considering OpenAI CEO Sam Altman’s recent remarks about AI as more of a utility, this added a lot of the perspective he’s seeing.
That doesn’t mean the model race is over. It means the model race has summoned an infrastructure race behind it. And that second race may end up shaping the market just as much as anything happening at the application layer.
This illustrates the reason why companies have spent the past year focused on locking down compute for several years into the future. They took the buy early and buy a lot approach in an effort to be well-positioned in the coming sprint.
That’s also why events like this are so revealing. Official conference narratives tend to orbit the cleanest and most legible parts of the ecosystem: big product launches, flashy demos, benchmark wins, enterprise partnerships. But the founder-and-investor room tells you where the discomfort is. And right now, the discomfort is physical. Too much demand. Not enough power. Too much heat. Too little time.
There’s a reason this keeps surfacing across the industry, from nuclear and alternative-energy conversations around AI infrastructure to increasingly gigantic buildout plans that sound more like industrial policy than startup strategy. The AI boom is becoming a contest over energy logistics, site selection, cooling architectures, and power density, not just model quality. That may be less cinematic than a new frontier model launch, but it’s where a huge amount of actual value will get created.
Also had maybe the most Silicon Valley moment ever: a frantic founder sprinting around the room looking for an engineer-first investor and insisting they’d be rich by morning. Equal parts great and terrible. In other words, a perfect mixer.
The big takeaway: the AI boom story is increasingly an infrastructure story. Beneath all the model talk, power and cooling are becoming the main characters.
