Why AI's next battleground is efficiency, not trillion-dollar spending
As AI infrastructure spending approaches unprecedented levels, researchers and startup founders say the industry's next breakthrough will come from delivering more intelligence with fewer resources, enabling wider adoption across enterprises, devices and geographies.
5 Min Read
The artificial intelligence race has so far been defined by a simple formula: bigger data centres, bigger models and bigger investments. But some of the industry's leading researchers and startup founders believe the next phase of competition will be decided not by who spends the most, but by who can deliver the most intelligence at the lowest cost.
Speaking exclusively to CNBC-TV18, industry leaders argued that while hyperscalers continue to pour hundreds of billions of dollars into AI infrastructure, the real challenge is making advanced AI accessible, affordable and widely deployable.
"One major area of focus is recursive self-improvement. We also see that scaling loss remains unsaturated. Scaling loss and reinforcement learning are advancing rapidly, so there is heavy investment in these areas," Sagar said.
He noted that a single principle currently drives much of AI development: scale. Companies want larger data centres, larger models and larger datasets. Yet that approach alone may not determine who ultimately benefits from the technology.
"What this approach misses, however, is distribution. Intelligence per token will be the determining factor in who gets to use AI and who can truly harness it at scale," he said.
The shift in thinking comes as the world's largest technology companies continue to ramp up spending on AI infrastructure. Industry estimates suggest the biggest hyperscalers have collectively committed close to $1 trillion towards AI-related investments.
For many countries and companies, matching that level of spending is unrealistic. That is why efficiency is emerging as a critical area of research.
Sagar said researchers are increasingly focused on reducing the amount of computing power needed to generate useful outcomes.
"How do you reduce costs and improve efficiency? That's a major research area that isn't discussed enough, but it's one of the biggest bets we're making," he said.
The discussion around efficiency is already beginning to influence how enterprises think about AI deployment. As usage grows, businesses are becoming more sensitive to the cost of running increasingly sophisticated models.
According to Sagar, the future of AI may be measured not simply by model size, but by metrics such as intelligence per token, intelligence per watt and intelligence per dollar.
"For enterprises and individual consumers alike, if you want to realise the true impact of AI, the focus has to be on getting more intelligence per token, intelligence per watt, or intelligence per dollar," he said.
The drive towards efficiency is also reshaping research priorities. Instead of relying solely on ever-larger models, researchers are exploring ways to compress advanced capabilities into smaller systems that can operate on smartphones, personal devices and edge infrastructure.
"How do you condense large models into smaller ones? How do you make them run on phones and edge devices? That's the most exciting area for me and will be a major differentiator going forward," Sagar said.
The economics of AI are becoming increasingly important because the industry's monetisation model remains uncertain. Unlike previous technology waves, where businesses could more easily forecast adoption patterns and revenue opportunities, AI capabilities are evolving at an unprecedented pace.
Sagar noted that what appears to be a viable business model today may become obsolete within months as new models and capabilities emerge.
"It's a question that remains unsolved right now," he said, referring to AI monetisation.
That uncertainty is also creating opportunities for startups.
Karan Vaidya, Co-Founder of Composio, said enterprises are becoming more conscious of costs and increasingly wary of dependence on a single AI provider.
"We're also seeing that the token-maxing trend is beginning to reverse as companies become more cost-conscious," Vaidya said.
He added that helping enterprises avoid lock-in to a specific model provider could become a valuable business proposition in a world where organisations use multiple AI systems simultaneously.
According to Vaidya, the future is unlikely to belong to a single dominant model. Instead, enterprises will operate in a multi-model environment where companies choose among proprietary and open-source systems depending on cost, performance and use cases.
The focus on efficiency also extends beyond infrastructure and model development.
Tom Bradicich, Chief Product Officer and CTO at Arete, argued that productivity improvements delivered by AI are meaningful only when they improve broader business outcomes.
He said many organisations focus on making individual tasks faster without addressing bottlenecks across the entire workflow.
"The winners will understand entire workflows rather than optimising isolated subprocesses," Bradicich said.
His point highlights a broader lesson emerging across the AI industry: efficiency is no longer just about reducing computing costs. It is increasingly about extracting greater value from every unit of investment, whether that investment is measured in capital, energy, computing power or human effort.
As the AI race enters its next phase, industry leaders suggest that success may depend less on who builds the biggest model and more on who can make intelligence cheaper, more accessible and more useful at scale.
Speaking exclusively to CNBC-TV18, industry leaders argued that while hyperscalers continue to pour hundreds of billions of dollars into AI infrastructure, the real challenge is making advanced AI accessible, affordable and widely deployable.
Adithya Sagar, Head of AI Research at Meta, said frontier AI research remains focused on achieving Artificial General Intelligence (AGI) and eventually Artificial Superintelligence (ASI). However, he believes the industry's obsession with scale overlooks a crucial factor.
"One major area of focus is recursive self-improvement. We also see that scaling loss remains unsaturated. Scaling loss and reinforcement learning are advancing rapidly, so there is heavy investment in these areas," Sagar said.
He noted that a single principle currently drives much of AI development: scale. Companies want larger data centres, larger models and larger datasets. Yet that approach alone may not determine who ultimately benefits from the technology.
"What this approach misses, however, is distribution. Intelligence per token will be the determining factor in who gets to use AI and who can truly harness it at scale," he said.
The shift in thinking comes as the world's largest technology companies continue to ramp up spending on AI infrastructure. Industry estimates suggest the biggest hyperscalers have collectively committed close to $1 trillion towards AI-related investments.
For many countries and companies, matching that level of spending is unrealistic. That is why efficiency is emerging as a critical area of research.
Sagar said researchers are increasingly focused on reducing the amount of computing power needed to generate useful outcomes.
"How do you reduce costs and improve efficiency? That's a major research area that isn't discussed enough, but it's one of the biggest bets we're making," he said.
The discussion around efficiency is already beginning to influence how enterprises think about AI deployment. As usage grows, businesses are becoming more sensitive to the cost of running increasingly sophisticated models.
According to Sagar, the future of AI may be measured not simply by model size, but by metrics such as intelligence per token, intelligence per watt and intelligence per dollar.
"For enterprises and individual consumers alike, if you want to realise the true impact of AI, the focus has to be on getting more intelligence per token, intelligence per watt, or intelligence per dollar," he said.
The drive towards efficiency is also reshaping research priorities. Instead of relying solely on ever-larger models, researchers are exploring ways to compress advanced capabilities into smaller systems that can operate on smartphones, personal devices and edge infrastructure.
"How do you condense large models into smaller ones? How do you make them run on phones and edge devices? That's the most exciting area for me and will be a major differentiator going forward," Sagar said.
The economics of AI are becoming increasingly important because the industry's monetisation model remains uncertain. Unlike previous technology waves, where businesses could more easily forecast adoption patterns and revenue opportunities, AI capabilities are evolving at an unprecedented pace.
Sagar noted that what appears to be a viable business model today may become obsolete within months as new models and capabilities emerge.
"It's a question that remains unsolved right now," he said, referring to AI monetisation.
That uncertainty is also creating opportunities for startups.
Karan Vaidya, Co-Founder of Composio, said enterprises are becoming more conscious of costs and increasingly wary of dependence on a single AI provider.
"We're also seeing that the token-maxing trend is beginning to reverse as companies become more cost-conscious," Vaidya said.
He added that helping enterprises avoid lock-in to a specific model provider could become a valuable business proposition in a world where organisations use multiple AI systems simultaneously.
According to Vaidya, the future is unlikely to belong to a single dominant model. Instead, enterprises will operate in a multi-model environment where companies choose among proprietary and open-source systems depending on cost, performance and use cases.
The focus on efficiency also extends beyond infrastructure and model development.
Tom Bradicich, Chief Product Officer and CTO at Arete, argued that productivity improvements delivered by AI are meaningful only when they improve broader business outcomes.
He said many organisations focus on making individual tasks faster without addressing bottlenecks across the entire workflow.
"The winners will understand entire workflows rather than optimising isolated subprocesses," Bradicich said.
His point highlights a broader lesson emerging across the AI industry: efficiency is no longer just about reducing computing costs. It is increasingly about extracting greater value from every unit of investment, whether that investment is measured in capital, energy, computing power or human effort.
As the AI race enters its next phase, industry leaders suggest that success may depend less on who builds the biggest model and more on who can make intelligence cheaper, more accessible and more useful at scale.
(Edited by : Prashant)
Source link
