BENGALURU: Vishal Sikka, founder and chief executive of Vianai, said that the current architecture behind large language models (LLMs) is fundamentally inefficient and will need to be replaced.
Speaking at the India AI summit, Sikka said that while artificial intelligence has shown “astonishing effectiveness”, the way today’s models are built and powered cannot continue at scale.
“This idea that I write a prompt, and these gazillion GPUs blast into existence to produce a response, and then I make a tiny change to that prompt, and then I do that again, it just seems like a completely absurd idea, especially to someone who has been around AI for such a long time.”
He said that each query to a large model can require significant computing power, and small changes in wording often trigger a full recalculation.
Sikka compared AI systems with the human brain. He noted that a person typically consumes about 2,000 calories a day, roughly equivalent to a 100-watt light bulb. Of that, the brain and nervous system use around 15 to 20 watts.
“That’s like when your laptop is in sleep mode,” he said. By contrast, modern AI systems depend on vast arrays of graphics processing units (GPUs) inside energy-hungry data centres.
“There are many zeros still to be removed from these models,” Sikka said, adding: “and the models themselves have to be removed.”
Hallucinations Blocking Enterprise Use
Sikka also highlighted technical limitations beyond energy use. He said hallucinations, when AI systems generate confident but incorrect information, remain a major obstacle to adoption in large organisations.
He added that safety is an “existential issue” and warned about the risks of autonomous AI agents acting unpredictably.
“Swarms of agents can be made to do completely reckless things, and we don't yet have ways to understand or deal with this,” he said.
He compared the need for safeguards in AI to the long-standing regulation of nuclear power, saying: “We can and we must do this with AI.”
AI Productivity
Despite his criticism of current architectures, Sikka described AI as “astonishingly effective” when used properly.
He gave an example of a friend who rebuilt a public service platform alone in 14 days using a generative AI coding tool. The same system had previously taken 15 engineers nine months to develop.
He also described a customer who used AI-driven simulations to decide within days to exit a country after a supplier shut down operations. According to Sikka, the same decision would previously have taken a year and involved external consultants.
However, he said there remains “a huge gap between LLMs and the business users inside enterprises”.
Bridging that gap requires systems that are “correct, trusted, verifiable, reliable” rather than simply powerful, he said.