There seems to be an ever-growing number of “AI companies.”

There’s an AI tool for your meetings, your inbox, your sales team and probably whatever task you’re avoiding right now. Put together a demo, give it a memorable name and promise to save someone six hours a week.

But underneath every one of those products is a computer doing the work.

Someone has to build that, too.

At the University of Minnesota, two startups are taking on that assignment right here in Minnesota. Besimax and D-Rez are developing hardware that could make AI computation substantially more energy-efficient. Their ambitions involve new ways of handling information and some very cold chips.

It’s an intriguing role for Minnesota: a potential underdog in the AI boom, with a history that makes the “underdog” label a little misleading.

In 1964, Minnesota-based Control Data introduced the CDC 6600, widely recognized as the first successful supercomputer. Seymour Cray and his team then went on to develop it at the company’s laboratory in Chippewa Falls, Wisconsin. It became the world’s fastest computer and helped put the region at the center of an emerging industry. Computer History Museum, NSF NCAR.

In today's compute architecture, significant amounts of energy are spent transferring data between memory and logic.

Jian-Ping (J.P.) Wang, founder of Besimax and the inventor behind Midwest unicorn Niron Magnetics, explains the idea with a kitchen analogy. Cooking a meal means moving back and forth between the pantry and the stove, fetching ingredients at each step. Now imagine bringing the stove into the pantry, so everything happens in one place. That’s the idea behind Besimax.

Besimax is commercializing research by UMN professor J.P. Wang and his collaborators into computational random-access memory, or CRAM. The idea is to perform computation inside memory itself

For Besimax cofounder, Marie Rahne, years of research are meeting a moment when the need feels urgent.

“Honestly, it’s the timing,” she said when asked what excites her about the approach. “AI keeps demanding more, and you can see the strain, in the grid, in the water, in the neighborhoods near the data centers.” Her argument is that reducing those demands could change how people feel about AI itself. It is easier to get excited about a technology’s possibilities when its costs feel manageable.

“We believe in the promise of AI, but that promise cannot come at the expense of the places we live,” Rahne said. 

Besimax’s next step is a prototype chip for prospective partners and customers. Then come manufacturing relationships, scaling production, and fitting the technology with today's software. There is no “launch website” button for that part.

Besimax isn’t the only university team doing the work. A few labs away, D-Rez is going after a different source of waste: the circuits themselves. Co-founder Katelyn Espe explains it in terms any Midwesterner will get.

“Try to imagine ice skating on gravel,” she said. “You wouldn’t get very far due to friction.”

In her analogy, conventional semiconductor circuits are the gravel: electrical resistance turns some of the energy flowing through them into heat, adding to power demands and limiting performance. Superconductors, cooled to extremely low temperatures, provide the smooth ice, allowing current to flow without electrical resistance.

D-Rez is developing chips that use these properties to perform AI computation. According to Espe, the company’s prototype can already perform tasks at significant energy and speed advantages. The challenge now is scaling it to handle more complex work and fit with today's infrastructure. 

Still, the prospect is compelling enough to spend a lot of time getting very small things very cold. And Minnesota provides the equipment to do it.

Espe said the team can build chips at the Minnesota Nano Center, then take them to its UMN lab for testing in specialized cooling systems.

That detail makes the local connection tangible. Developing this kind of company requires access to facilities, specialized equipment and people who know how to use them. For D-Rez, those resources are here.

Two startups do not make Minnesota the next capital of AI hardware. Both companies have substantial work ahead before their technology can reach widespread use.

But they do make for an interesting answer to a question that can get lost amid the parade of new AI products: Who is working on the infrastructure all of this depends on?

Here, that work may be found growing out of university research. The founders’ accounts point to something more specific than hometown pride: years of scientific work, places to fabricate chips and facilities to test ideas that cannot be tested on a laptop.

Rahne sees those assets as part of a longer Minnesota story.

“The talent, the university research, and the manufacturing culture that produced Cray and Control Data are still here,” she said.

The opportunity, in her view, is to turn those strengths into companies whose growth benefits the region. That means jobs, expertise and intellectual property taking root locally. It also means giving hardware companies the support to get through the difficult stretch between a promising result and something a customer can actually use.

Rahne said that requires investors, government, higher education and the broader business community to participate.

“No single company builds a leadership position alone,” she said. “It comes from a region deciding together that this is worth doing here.”

For now, both companies need to raise capital, build chips, and prove their technology can meet tomorrow’s computing demands. Those are difficult, concrete milestones, and how they deliver will matter more than comparisons to Minnesota’s computing glory days.

But there is something fun about the possibility.

The next time another AI app appears in your feed, spare a thought for the people figuring out what it might run on someday.

Some of them are in Minnesota, building chips, and, in at least one lab, making sure they’re very, very cold.