When DeepSeek’s R1 model startled Silicon Valley last year, the takeaway went beyond lower costs. It showed what can happen when a team challenges received wisdom, and finds a better way.
A Bengaluru-based research lab now claims to have done something similar - not with a chatbot, but with a fresh approach to how AI itself is built.
Fresh approach
Bud Ecosystem has developed a new architecture, Resource Aware Attention (RAA). Its first application, a suite of AI safety models called Bud Sentinel, delivers striking gains: on a standard consumer laptop, it runs more than twice as fast as comparable systems deployed on high-end GPU servers.
Safety checks are an essential but often overlooked layer of AI. Every chatbot or automated system relies on guardrail models to screen inputs and outputs for harmful or inappropriate content.
Bud Ecosystem logo
Questioning assumptions
Most of today’s leading systems are designed for expensive GPU infrastructure. While they respond in about 18–19 milliseconds on a high-end server, performance slows sharply - down even to 300–800 milliseconds on the CPUs that power most enterprise systems, making real-time use difficult. Bud Ecosystem questioned that premise.
RAA rethinks the core mechanism by which AI models process information, tailoring it to the strengths of widely available CPUs rather than specialised hardware. Bud Sentinel is built on this approach, and can classify requests in about 8.4 milliseconds on a standard laptop. This is faster than rival systems running on far more expensive machines.
Democratising AI
The company says it also surpasses current benchmarks for accuracy across multiple independent tests. “Democratising AI means making it affordable and accessible on everyday devices, without dependence on specialised hardware,” said Jithin VG, Founder and Chief Executive of Bud Ecosystem.
Rethinking design
“After 18 months of pushing CPU performance, we realised optimisation alone was not enough. We had to rethink the design itself,” he told businessline. The company describes Sentinel as only a first step. It plans to extend RAA across the broader generative AI stack, with the aim of enabling high-performance systems to run on ordinary hardware rather than costly accelerators.
DeepSeek’s breakthrough mattered not just for its price, but for its willingness to question long-held assumptions. Bud Ecosystem’s work poses a similar challenge. Whether it will reshape the economics of enterprise AI remains to be seen. But it begins with the right question, says Jithin.
Published on April 8, 2026