Indian Firm Runs AI in Air Conditioners Without Internet

While tech giants race to build bigger cloud-based AI systems, an Indian company is going the opposite direction and solving real problems in the process.

CoRover AI, led by CEO Ankush Sabharwal, has built an AI model that runs inside an air conditioner on just 4GB of RAM with zero internet connection. The system understands voice commands in five languages and controls the AC without ever hallucinating or making mistakes because it's designed to do exactly one thing perfectly.

At DevSparks 2026 in Bengaluru, Sabharwal challenged the tech industry's automatic assumption that AI needs cloud infrastructure. "Begin with purpose in mind," he said, arguing that the industry's reflex to build everything in the cloud since 2015 deserves closer examination.

The practical benefits are showing up in critical environments. Banks across India now use CoRover's conversation alert system running entirely on-premises, avoiding the compliance nightmares that cloud migrations trigger. Defense installations operate secure AI systems that can't be hacked from outside because they're completely isolated from the internet.

Hospitals have an even more urgent need. Cloud providers guarantee 99.9% uptime, which sounds impressive until you calculate that it means eight hours of potential downtime each year. In an operating room, Sabharwal notes, eight hours down could be catastrophic. The 2024 CrowdStrike outage forced Boston hospitals to abandon their digital systems and return to pen and paper.

CoRover's approach actually delivers better performance than cloud alternatives for specific tasks. Smaller, focused models respond faster and more accurately than giant general-purpose systems because they're designed and tested for exact use cases rather than trying to do everything.

The hardware is catching up to make this possible at scale. NVIDIA's DGX Spark, a desktop-sized supercomputer, allowed CoRover to train models with up to three billion parameters entirely without cloud infrastructure. Intel's AI PC architecture combines multiple processors on a single device, running different parts of AI pipelines simultaneously.

The Ripple Effect

Only 7% of companies have successfully scaled AI into their actual workflows, according to MIT research that Sabharwal cited. The other 93% are stuck experimenting or running limited pilots, often because they chased trendy technology instead of solving concrete problems.

CoRover's success shows a different path forward. By starting with real human needs in constrained environments, they're bringing AI to places that matter most: rural infrastructure, factory floors, medical facilities, and areas where internet connectivity can't be assumed. These aren't experimental deployments or temporary fixes but full-scale systems serving millions of users across India.

The lesson extends beyond AI architecture. Technologies that work in the most challenging conditions, that respect privacy and security by design, and that solve specific problems exceptionally well often matter more than systems chasing the biggest possible scale.

Smart technology isn't always about being bigger or more connected; sometimes it's about being precisely right for where it's needed most.