The funding round was led by VoLo Earth Ventures, which focuses on climate-related technologies. The company said the investment will help it grow its team, build its platform and strengthen talks with large technology firms.

Spending on data centres is expected to reach nearly $700 billion this year, driven largely by AI workloads. At the same time, energy use from these facilities is projected to double by 2028.

Refiant is trying to tackle that problem by making AI models smaller and less power-hungry. The company said it has already compressed a 120 billion parameter model so it can run on a standard laptop. Normally, such a model would require far more powerful hardware.

According to the company, the compressed version runs on a MacBook Pro with 12GB of memory. It keeps between 95% and 99% of its original performance. It can also run alongside another model on the same device.

Sid Gutta, co-founder of Refiant, said: “AI’s growing energy footprint is one of the most urgent and underappreciated challenges in the climate space. The industry’s default answer is to build more data centres and consume more power. Ours is to make the AI itself dramatically more efficient.”

The company, based in California, said it tested energy use inside a Faraday cage to ensure accurate readings. Under those conditions, the system reached about 3,000 tokens per kilowatt-hour.

That is up to 100 times more efficient than running the same model in a traditional data centre.

In practical terms, the energy used for one AI task on standard infrastructure could handle about 100 similar tasks using Refiant’s approach.

The founders argue that improving efficiency is a better long-term path than expanding infrastructure. Running AI locally on smaller machines could also help organisations avoid sending data to large cloud providers, which usually means higher costs and less management.

Recent developments from Google have pointed in a similar direction. Its TurboQuant compression method reduced memory needs significantly, reinforcing interest in making models leaner rather than simply scaling hardware.

Dr Viroshan Naicker, co-founder of Refiant, said: “The AI industry is spending hundreds of billions scaling infrastructure when the real breakthrough is the ability to do more with radically less. Nature doesn’t build by brute force. Evolution optimises. We’ve applied that principle to AI, and the results speak for themselves.”

The company believes its work could also help businesses balance AI adoption with environmental targets.

Mathew Haswell, another co-founder, said: “Those two mandates don’t have to be in tension. AI adoption and sustainability commitments can coexist, but only if the technology itself becomes more efficient. 

“Organisations shouldn’t have to choose between deploying AI and meeting their energy targets – and they shouldn’t have to send their data halfway around the world to do it.”

Joseph Goodman, managing partner at VoLo Earth Ventures, added: “AI’s biggest constraint isn’t demand, it’s energy. What’s been missing is a fundamentally more efficient way to compute.

“Refiant’s architecture replaces brute-force scaling with a far more efficient, nature-inspired approach that lowers energy use while increasing capability. That’s the kind of breakthrough needed to make AI sustainable on a global scale.”

Refiant said it is already in discussions with several multinational firms. It plans to push its technology further, with a focus on stronger compression, longer context handling and better tracking of how models operate.

Faced with surging power demands and an overstretched grid, Google confirmed it has entered formal “demand-response” agreements with Indiana Michigan Power and the Tennessee Valley Authority. 

Under these agreements, the company will temporarily reduce power consumption at its AI-heavy data centres during peak periods, when the grid is at risk of overloading.

These are the first such deals Google has signed specifically to control machine learning workloads, which are some of the most power-hungry processes in AI computing. 

The company had tested a similar system last year in a pilot with Omaha Public Power District, where it successfully reduced energy use across three grid stress events.

The decision comes as utilities across the U.S. report being overwhelmed by the scale of electricity needed to fuel the rapid growth of AI. In some areas, demand from data centres has already exceeded supply. 

The implications are serious as higher bills for everyday users, increased risk of blackouts, and mounting pressure on an already weak energy infrastructure.

We’re not talking about marginal energy use here. Data centres now account for 4.4% of the entire electricity consumption in the U.S., a figure expected to nearly triple to 12% by 2028 if current trends continue. In one extreme case, an AI facility planned in Wyoming is projected to consume more electricity than all the state’s residential homes combined.

Some AI server racks are reportedly drawing more power than entire steel plants. Training a single chatbot model can consume as much energy as 100 households use in a year. Between 2024 and 2025 alone, data centre expansion triggered an estimated $9 billion increase in electricity costs. 

Unfortunately, it’s not the tech companies footing the bill—it’s the public. In Virginia, for example, Dominion Energy is seeking regulatory approval to raise household electricity bills by $10.50 per month, noting increased demand from data centres.

With regulators watching, Google’s decision appears as much strategic as it is necessary. The Federal Energy Regulatory Commission held emergency hearings in June 2025 to examine whether the country’s power infrastructure can support the wave of AI projects being rolled out. The grid is under stress.

In a blog post, Google stated, “It allows large electricity loads like data centres to be interconnected more quickly, helps reduce the need to build new transmission and power plants, and helps grid operators more effectively and efficiently manage power grids.”

Historically, demand-response programmes have been associated with energy-intensive sectors like manufacturing or crypto mining, where businesses get paid or offered bill discounts in exchange for curtailing their energy use during peak hours. Google’s move is the first time this method is being systematically applied to artificial intelligence workloads.