Vaire Computing aims to tame the AI energy crisis

A new startup is building an energy-efficient chip based on the principle of physical reversibility. Could it open a multi-trillion-dollar market with a positive environmental impact?

AI has an energy problem.

As companies like OpenAI try to train more ambitious systems like ChatGPT, they use more and more electricity, so much that their CEO, Sam Altman, is exploring new nuclear power plants to cover the growing demand—as are their competitors at other major AI players like Google, Meta, and Microsoft. The computer chips used for AI also generate so much heat that data centers often rely on evaporative cooling systems that consume surprisingly large amounts of water just to keep cool, sometimes straining local water resources.

But all that heat and power might not be necessary.

Ideas from physics suggest that computers could use far less energy and generate far less heat, especially in the kinds of parallel computations common in AI. Now, a startup hopes to bring those ideas out of simulations and into the real world, making more efficient chips for the future of computing.

Harnessing the laws of physics for energy efficiency

In 2021, Hannah Earley had just finished a PhD in Applied Mathematical and Theoretical Physics at the University of Cambridge, where she used ideas from both physics and computer science to study unconventional methods of computing.

“I saw the world around me and wanted to know how it worked, and got into science that way,” said Earley. “That extended to computers.”

From the beginning of her PhD, she was curious about an idea called reversible computing. In physics, most laws are time-symmetric—they work the same backwards and forwards in time. If you ran the orbits of the planets or the current in a circuit backwards, you’d find their new motion still obeys the laws of physics.

The exception comes from thermodynamics. There are only a few ways for matter to be orderly and useful, like an egg on a plate, and many ways for it to be messy and useless, like a cracked egg on the floor. While a cracked egg could, in theory, reassemble itself by chance, it is statistically improbable. That’s entropy: systems naturally tend toward disorder. And when they do, useful energy (or eggs) turns into useless heat (or messes to clean up).

In 1961, physicist Rolf Landauer applied this idea to computing. When a normal computer performs a calculation, it deletes information. Add two numbers together, and you get only one number out: you can’t tell anymore which two numbers you started with. Normal computers are irreversible, and that irreversibility has a real physical cost, as Landauer calculated: it turns energy into heat.

But what if you didn’t have to erase anything?

That’s the idea behind reversible computing: design a computer that keeps all its information, so it never needs to erase—and never needs to pay the heat penalty.

“You start to think, where will our civilization go if it becomes more technologically advanced?” said Earley. “How will it get the computing power it needs given the fundamental thermodynamic constraints?”

The birth of Vaire Computing

Vaire Computing co-founders Hannah Earley and Rodolfo Rosini (Credit: Intel Ignite)

In 2021, Earley met Rodolfo Rosini. A tech entrepreneur, Rosini was watching AI become more and more energy-hungry, and saw a crisis on the horizon.

“He saw that a big hardware crunch was coming. And he was convinced that, not only was reversible computing the future, but it was the future now,” said Earley.

Together, Earley and Rosini founded Vaire Computing. Named after the goddess Vairë the Weaver from Tolkien’s The Silmarillion, Vaire aims to build reversible computer chips that can be mass-produced, making energy efficiency practical.

Part of that efficiency comes from new ways of handling signals in computer chips. Even if a computer didn’t delete data, it would still release heat as it quickly turned signals on and off. This waste heat can be reduced by making the signals slower, with gentler changes in voltage. Slowing things down gets you closer to what is called adiabatic behavior, a thought experiment of a system that changes so slowly that it doesn’t release heat at all.

Of course, slower chips usually mean slower performance. But AI often relies on parallelism—many chips running simultaneously doing different pieces of a calculation.

“If you slow down the clock by a factor of two, you’ve reduced your energy by a factor of four. This is where reversible computing gets really interesting,” said Earley.

If each chip runs twice as slow but uses four times less energy, you can use twice as many chips, getting the same speed for half the power bill (though you obviously have to buy more chips). Slower, cooler chips can also be packed more densely without overheating, potentially unlocking new levels of performance.

Vaire's first tests and exciting potential

In 2024, Vaire made a major leap. They hired Michael Frank, a researcher at Sandia National Laboratories in the US. The year before, Frank had registered a patent for a circuit that could retain energy used in a reversible chip, using a resonator, the first such design that looked to be capable of tackling the problem in a practical setting. With that key technology in hand, they raised more money from investors, reaching $4.5 million in seed funding, as well as being selected for the UK’s ChipStart semiconductor incubator and the Intel Ignite startup accelerator.

“Every computer is constrained by heat and data centers by energy demands. Reversible computing and Vaire could transform computing with near zero-energy chips, extending Moore’s Law limits,” said Jude Gomila, one of the investors from Vaire’s 2024 seed round. “This technology has market potential in the multi-trillion dollars and, once commercialized, will positively impact the environment.”

Theoretical estimates suggest reversible chips could be up to 4,000 times more efficient than irreversible ones. By 2027 or 2028, Vaire plans to have a chip that’s commercially competitive, either for high-power applications like AI or for devices like cell phones where battery power is limited.

Vaire Computing prepares to test their new chip. (Credit: Vaire Computing)

For now, they’ve designed a prototype chip, one that they hope will be several times more efficient than chips on the market today, “showing that not only do the principles work on paper in theoretical physics but we’ve got actual silicon that actually does this,” explained Earley. “We don’t want to just be a couple or a few times better than what else is out there.”

Matt von Hippel is a science journalist based in Copenhagen with a background in particle physics. He blogs weekly at 4gravitons.com, and has written for Quanta Magazine, Scientific American, and Ars Technica.