In an exclusive conversation with CNBC-TV18, industry experts conveyed that while hyperscalers continue to invest hundreds of billions in AI infrastructure, the real challenge lies in making advanced AI both accessible and cost-effective for widespread implementation.
Adithya Sagar, Head of AI Research at Meta, stated that the focus of frontier AI research is on achieving Artificial General Intelligence (AGI) and eventually Artificial Superintelligence (ASI). Nonetheless, he believes that the industry’s fixation on scale disregards a vital aspect.
“One significant focus area is recursive self-improvement. We also observe that scaling loss remains unsaturated. There’s rapid progress in both scaling loss and reinforcement learning, leading to substantial investments in these fields,” Sagar explained.
He highlighted that a single principle currently underpins much of AI advancement: scale. Companies strive for larger data centers, more colossal models, and bigger datasets. However, relying solely on this approach may not guarantee who ultimately benefits from the technology.
“What this methodology overlooks is distribution. Intelligence per token will be the key factor in determining who can utilize AI and who can genuinely tap into its potential at scale,” he added.
This shift in perspective emerges as the largest tech firms continue increasing their spending on AI infrastructure. Industry estimates suggest that the top hyperscalers have collectively pledged nearly $1 trillion for AI-focused investments.
For many countries and organizations, matching that level of expenditure is simply unattainable. Thus, enhancing efficiency is becoming a pivotal research focus.
Sagar emphasized that researchers are pivoting towards minimizing the computing power needed for effective outcomes.
“How can you cut costs and enhance efficiency? This is a major area of research that isn’t sufficiently discussed, but it’s among our most crucial bets,” he stated.
The increasing dialogue on efficiency is already shaping enterprise strategies regarding AI deployment. As usage surges, businesses are becoming increasingly mindful of the costs associated with operating advanced models.
According to Sagar, the future of AI might be evaluated not just by model size, but by metrics such as intelligence per token, intelligence per watt, and intelligence per dollar.
“For both enterprises and individual users, to truly harness AI’s benefits, the emphasis must be on maximizing intelligence per token, watt, or dollar,” he noted.
This drive for efficiency is also influencing research priorities. Rather than depending exclusively on larger models, researchers are investigating methods to integrate advanced functionalities into smaller systems capable of operating on smartphones, personal devices, and edge systems.
“How do we shrink large models to smaller versions? How do we enable them to function on phones and edge devices? That’s the most thrilling domain for me and will serve as a significant differentiator moving forward,” Sagar remarked.
The economic aspects of AI are increasingly pivotal as the industry’s monetization framework remains ambiguous. Unlike earlier technological waves, enterprises find it challenging to predict adoption trends and revenue potentials due to the rapid evolution of AI capabilities.
Sagar noted that current viable business models could become outdated within months as new advancements are introduced.
“This question regarding AI monetization remains unanswered,” he remarked.
This uncertainty also opens avenues for startups.
Karan Vaidya, Co-Founder of Composio, indicated that enterprises are growing more cost-conscious and cautious about relying on a single AI provider.
“We are witnessing a reversal of the token-maxing trend as businesses become increasingly mindful of expenses,” Vaidya mentioned.
He added that helping companies avoid dependency on specific model providers could emerge as a significant business opportunity in a multi-AI system landscape.
Vaidya predicts that the future will not be dominated by a single model, but will involve enterprises operating in a multi-model environment where they select among various proprietary and open-source systems based on cost, efficiency, and application.
The emphasis on efficiency extends beyond just infrastructure and model development.
Tom Bradicich, Chief Product Officer and CTO at Arete, contended that the productivity gains from AI are only valuable if they enhance broader business outcomes.
He noted that many organizations tend to optimize individual tasks without addressing larger workflow bottlenecks.
“The winners will be those who understand entire workflows instead of merely improving isolated subprocesses,” Bradicich asserted.
His comments reflect a larger insight emerging in the AI sector: efficiency is increasingly about maximizing value from every investment, whether that investment pertains to capital, energy, computational power, or human resources.
For the complete interview, watch the accompanying video
As the AI race progresses into its next stage, industry leaders propose that success may hinge less on the size of models and more on the ability to make intelligence cheaper, more accessible, and widely useful.
