Artificial intelligence model compression startup Refiant AI said today it has raised $5 million in seed funding from VoLo Earth Ventures to try to put an end to the “arms race” that has ignited a multibillion-dollar data center boom.

Refiant’s goal is simple, yet extremely ambitious: it wants to convince the AI industry to stop spending billions of dollars on building massive, power-hungry data centers and instead focus on making models smaller and more efficient. To do this, it has developed “nature-inspired” compression algorithms that it says can slash the energy requirements of most models by more than 80%.

The startup believes the current AI infrastructure boom is simply unsustainable. Big technology giants such as Amazon.com Inc., Microsoft Corp., Google LLC, Meta Platforms Inc. and Oracle Corp. have committed collectively to spending almost $700 billion on data centers this year, while their hardware demands have led to crippling shortages of key components like memory chips.

The reason for the AI infrastructure boom is the broad consensus that the most powerful models can only be run on massive clusters of graphics processing units. They need to be chilled by specialized, power-intensive cooling systems that suck up enormous amounts of energy.

Refiant says this has led to a situation where only the richest companies are in a position to host the most advanced large language models today. That forces everyone else to send their sensitive data to cloud-hosted servers owned by those tech giants.

The startup believes this not only creates a privacy risk, but encourages greater inefficiency for AI workloads, as the cloud giants generate immense profits from it. Moreover, there are concerns that the AI industry is rapidly approaching a ceiling, as the energy requirements of data centers outstrip power availability.

Model compression offers a tantalizing solution to this. Refiant’s secret sauce is the way it handles model weights and retraining. The startup explains that traditional compression techniques have to sacrifice AI model’s intelligence and accuracy, and to get around this it has developed a novel mathematical approach that mimics biological optimization.

Co-founder Viroshan Naicker is an experienced mathematician who has spent years researching networks and quantum systems. He argues that it’s mathematically possible to do a lot more with a whole lot less energy. “Nature doesn’t build by brute force. Instead, evolution optimizes,” he said. “We’ve applied that principle to AI, and the results speak for themselves.”

The company offered evidence of this claim when it recently demonstrated how it compressed a 120 billion-parameter model to run on a standard MacBook Pro laptop with just 12 gigabytes of random-access memory. Ordinarily, such a model would require a minimum of 80 gigabytes of high-end VRAM.

Impressively, virtually nothing was sacrificed while doing this. According to Naicker, the model retained about 95% to 99% of its original fidelity, and was able to process around 3,000 tokens per hour. It’s almost 100 times more energy-efficient than the standard data center configuration used to host models of that size.

Demonstrations are one thing, and Refiant’s goal now is to show it can successfully scale its mathematical approach to model compression. It said it’s holding talks with a number of technology companies that want to run AI models on their own on-premises hardware to maintain “sovereignty” and avoid the escalating costs of cloud-hosted AI infrastructure. Meanwhile, it’s working to apply its technique to even higher compression ratios and longer context windows.

“AI’s biggest constraint isn’t demand, it’s energy,” said VoLo Earth Managing Partner Joseph Goodman. “What’s missing is a fundamentally more efficient way to compute, but Refiant’s architecture replaces brute-force scaling with a far more efficient nature-inspired approach that lowers energy use while increasing capability. This is the kind of breakthrough needed to make AI sustainable on a global scale.”

Image: Refiant