Watch out Nvidia: AI startup bets big on reversible computing

  • The CEO of Vaire Computing says reversible computing is inevitable
  • Reversible computing chips generate virtually no heat
  • Of course, chips that generate no heat would throw a dent in the liquid cooling industry’s growth plans

If you’ve been in the industry long enough, you know naysayers have been warning for years now that chipmakers are about to run up against the limits of Moore’s Law. But while there’s not a clear consensus on exactly when that will happen, the day is coming. And startup Vaire Computing will be ready and waiting with its reversible computing technology.

In fact, the CEO of Vaire Computing is so confident that reversible computing will power artificial intelligence (AI) in the future that he’s actually gone out and hired pretty much every expert in the field. By securing an early monopoly in this area, he told Fierce Network, the company is aiming to become the next Nvidia.

“Our perspective, and I know this sounds outrageous, is that in 15 years, every computer chip is going to be reversible. It is inevitable, there is no escaping,” Vaire CEO Rodolfo Rosini said. “I think we’ve figured it out, and everyone else has gotten it wrong, including Nvidia, including Intel.”

What on earth?

Follow us into the weeds for a moment so we can explain the logic behind Rosini’s statement.

Moore’s Law is more an observation than a hard and fast rule of physics, and it basically states that the number of transistors on an integrated circuit will double every two years without a dramatic rise in cost. The number of transistors matter for chip companies like Nvidia that are trying to push the boundaries of computing for use cases like AI.

Despite what headlines would have you believe, performance gains have actually been slowing down (we heard about this from an exec at Nokia acquisition target Infinera a while back), and the focus is starting to turn to energy efficiency.

That’s where reversible computing comes in.

In traditional chips, energy is used to change the state of gates on the silicon, and the waste byproduct is heat. In reversible computing, though, that waste energy from each gate change is recycled to power the next gate change. In practice, this means reversible computing chips are slower than traditional chips but they are wildly energy efficient and generate virtually no heat.

Reversible computing is actually not a new technology. It dates back to the 1970s, with the first universal reversible gates built in 1982 – but Vaire is only now optimizing the technology with a multi-core design to make it viable in an AI-driven world.

Those who have been paying attention know heat and power are two key concerns for the compute industry, given growing chip densities and power grid constraints in the U.S. and abroad. Of course, chips that generate no heat would throw a dent in the liquid cooling industry’s growth plans.

“In four years, there will be a reset of the slate where no one can get performance unless they change the architecture,” Rosini said. “At that point there will be an opportunity to build another trillion-dollar company over a decade or so by completely controlling this technology.”

“So, the bet is not being acquired by Nvidia, but being the next Nvidia,” he continued.

Rosini said Vaire already spent three years in stealth mode doing research and development, and it just raised $4 million in seed funding to bring its total raised to date to $4.5 million.

Vaire is preparing to launch an initial version of its silicon in early 2025. The CEO said at first its chips will be designed for edge deployments, but within four to five years it expects to have chips robust enough for data centers.

Outside perspective

So, what does the analyst community make of Vaire and it's claims?

Well, Gartner VP Analyst Gaurav Gupta told Fierce he isn't familiar with Vaire, but noted Rosini's point about Moore's Law breaking down is true. The cadence of transistor doubling has slowed and costs are no longer holding steady. In fact, Gupta said, the cost of leading edge wafer chips is actually exploding. 

That said, Gupta noted chipmakers are making improvements in other areas, for instance through packaging techniques. 

While he couldn't speak to Vaire specifically, Gupta also pointed out there's a big difference between doing R&D work – as Vaire has now been doing for several years – and commercializing a product at scale. The latter, of course, is much harder.

Fierce also took the question of whether Vaire's plan is feasible to AvidThink Founder Roy Chua, who is one of apparently a few analysts who are familiar with both the company and its technology. The takeaway? He seems to think Vaire has a huge lift ahead of it.

"Vaire's initial raise, which is relatively small for a silicon play, especially one that is potentially revolutionary, to get to a demonstrable milestone indicates renewed interest in this field," Chua said, adding the early 2025 target for Vaire's first silicon "seems aggressive but exciting if it happens."

He continued that it will take "a significant amount of effort and time" to displace today's compute architecture, which is complex and would require thoughtful reconsideration of things like the nature of how logic gates are constructed, manufacturing processes, EDA process for designing circuits at scale and system architecture, among other things. And it's not entirely clear how Vaire's solution could directly plug into existing workflows, Chua added. 

But if Vaire is able to demonstrate the feasibility and commercial viability of its solution? Well, then "there will be a lot more money pouring into the field given the need for increased computing resources to power all manner of workloads, including AI/ML. I'd expect their next raise, if this is successful, to be significantly larger to help drive scale," Chua concluded.

For the record, Fierce reached out to Nvidia for its thoughts but the company declined to comment on whether it has explored reversible computing, whether it is worried about the limits of Moore’s Law or whether it believes anyone can unseat it as the dominant force in the AI compute space. Instead, a spokesperson pointed Fierce to this blog post.

But it’s worth noting that the company isn’t just relying on transistors to boost chip performance – it’s also tweaking its software to deliver gains well after each generation of chips is released. You can read more about that in our story here.