Databricks’ ex AI chief aims to slash energy costs by 1,000x

The relentless pursuit of the next breakthrough in AI has funded numerous ambitious initiatives, yet one startup is seizing this opportunity to reconstruct computing architecture from scratch.
Under the leadership of Naveen Rao, former head of AI at Databricks, Unconventional AI aims to drastically improve the energy efficiency of inference processing. Their key innovation: a novel oscillator-based computer architecture.
On Thursday, the company unveiled its inaugural AI model, Un0, an image-generation tool that demonstrates for the first time how their technology can mirror conventional AI systems. In a companion research paper, the team explains how they developed a fully functional image generation model using a software simulation of this new architecture, achieving performance comparable to state-of-the-art diffusion models.
“This represents the ‘hello world’ of a new computing paradigm,” Rao told TechCrunch. “Expect some compelling developments in this space over the next year.”
While the output from Un0 resembles that of models like Stable Diffusion or OpenAI’s GPT Image 1, the distinction lies in its underlying mechanism. The model relies on an oscillator-based architecture, fundamentally different from the chips powering traditional computing and large language models. Although the benefits of this approach are intricate, Rao predicts it could reduce power consumption by up to 1,000 times.
However, much of the necessary infrastructure remains under development. The current Un0 version operates on a software simulation of Unconventional’s oscillator chips, with plans to release schematics for actual hardware soon. The long-term vision involves constructing a complete inference stack from the ground up, allowing Unconventional AI to provide compute capacity akin to other cloud providers.
“We are building a new system architecture centered on our proprietary chips,” Rao explains. “AI models will run on this infrastructure, with a network interface handling incoming prompts and outgoing inferences, all while consuming just 1/1000th of the power.”
This is an exceptionally ambitious objective for a company with fewer than 50 employees. Nevertheless, given the massive scale of AI expansion and the rising costs associated with meeting inference demands, this approach may be among the few capable of addressing the problem’s magnitude. Rao identifies power availability as a critical bottleneck for future AI growth, positioning Unconventional as one of the few projects equipped to tackle this constraint.
“Scaling AI is difficult due to energy constraints, which will likely become the fundamental limit in the coming years. We simply cannot bypass it; ultimately, this will be an energy-limited challenge,” he notes.
Related article
Anthropic Launches Claude Haiku 5.5, Cutting Small Model Costs by 75% and Letting Users Adjust Inference Intensity
On October 8th, Anthropic introduced its most efficient and capable model to date: Claude Haiku 5.5. Positioned as the fastest and most capable small model in Anthropic’s lineup, it is engineered for high-frequency, cost-sensitive applications such a
Claude Unveils First Full Ultraviolet Sky Map, Charting 119 Million Stars
With the powerful assistance of Claude, Anthropic’s large language model, astrophysicists have successfully generated the first complete ultraviolet all-sky map in history. This vast star chart, featuring 119 million stars, was largely completed by C
How AWS and NVIDIA Are Powering Agentic and Physical AI
NVIDIA CEO Jensen Huang (Credit: Tomohiro Ohsumi/Getty Images)Jensen Huang’s NVIDIA and AWS scale their global partnership, deploying 2M GPUs to power next-generation agentic and physical AI enterprise infrastructureAWS and NVIDIA have announced a ma
Related Special Topic Recommendations
Comments (0)
0/500

The relentless pursuit of the next breakthrough in AI has funded numerous ambitious initiatives, yet one startup is seizing this opportunity to reconstruct computing architecture from scratch.
Under the leadership of Naveen Rao, former head of AI at Databricks, Unconventional AI aims to drastically improve the energy efficiency of inference processing. Their key innovation: a novel oscillator-based computer architecture.
On Thursday, the company unveiled its inaugural AI model, Un0, an image-generation tool that demonstrates for the first time how their technology can mirror conventional AI systems. In a companion research paper, the team explains how they developed a fully functional image generation model using a software simulation of this new architecture, achieving performance comparable to state-of-the-art diffusion models.
“This represents the ‘hello world’ of a new computing paradigm,” Rao told TechCrunch. “Expect some compelling developments in this space over the next year.”
While the output from Un0 resembles that of models like Stable Diffusion or OpenAI’s GPT Image 1, the distinction lies in its underlying mechanism. The model relies on an oscillator-based architecture, fundamentally different from the chips powering traditional computing and large language models. Although the benefits of this approach are intricate, Rao predicts it could reduce power consumption by up to 1,000 times.
However, much of the necessary infrastructure remains under development. The current Un0 version operates on a software simulation of Unconventional’s oscillator chips, with plans to release schematics for actual hardware soon. The long-term vision involves constructing a complete inference stack from the ground up, allowing Unconventional AI to provide compute capacity akin to other cloud providers.
“We are building a new system architecture centered on our proprietary chips,” Rao explains. “AI models will run on this infrastructure, with a network interface handling incoming prompts and outgoing inferences, all while consuming just 1/1000th of the power.”
This is an exceptionally ambitious objective for a company with fewer than 50 employees. Nevertheless, given the massive scale of AI expansion and the rising costs associated with meeting inference demands, this approach may be among the few capable of addressing the problem’s magnitude. Rao identifies power availability as a critical bottleneck for future AI growth, positioning Unconventional as one of the few projects equipped to tackle this constraint.
“Scaling AI is difficult due to energy constraints, which will likely become the fundamental limit in the coming years. We simply cannot bypass it; ultimately, this will be an energy-limited challenge,” he notes.
Anthropic Launches Claude Haiku 5.5, Cutting Small Model Costs by 75% and Letting Users Adjust Inference Intensity
On October 8th, Anthropic introduced its most efficient and capable model to date: Claude Haiku 5.5. Positioned as the fastest and most capable small model in Anthropic’s lineup, it is engineered for high-frequency, cost-sensitive applications such a
Claude Unveils First Full Ultraviolet Sky Map, Charting 119 Million Stars
With the powerful assistance of Claude, Anthropic’s large language model, astrophysicists have successfully generated the first complete ultraviolet all-sky map in history. This vast star chart, featuring 119 million stars, was largely completed by C
How AWS and NVIDIA Are Powering Agentic and Physical AI
NVIDIA CEO Jensen Huang (Credit: Tomohiro Ohsumi/Getty Images)Jensen Huang’s NVIDIA and AWS scale their global partnership, deploying 2M GPUs to power next-generation agentic and physical AI enterprise infrastructureAWS and NVIDIA have announced a ma





Home






