Artificial intelligence is expanding at remarkable speed, with adoption, investment, and technical capability all moving upward. Stanford HAI's 2026 AI Index Report reports that organizational AI adoption reached 88% in 2025, while global corporate AI investment more than doubled to $581.69 billion. Generative AI also reached an estimated 53% population adoption within three years, suggesting that AI is becoming part of everyday digital and organizational activity at an unusually rapid pace.
That expansion brings a physical question into view. The International Energy Agency's Key Questions on Energy and AI projects global data-center electricity consumption could rise from 485 terawatt-hours in 2025 to approximately 950 terawatt-hours by 2030. A 2026 report from the United Nations University Institute for Water, Environment and Health also examines AI's associated electricity, water, and land requirements, highlighting the physical infrastructure supporting increasingly digital experiences.
Amid this landscape, Ankit Dheendsa, CTO of Morphos AI, an edge AI infrastructure company focused on autonomous robotic systems and defense applications, sees a question extending beyond model capability: How efficiently can intelligence be delivered? As AI systems become more capable, the computing, memory, electricity, cooling, connectivity, and physical infrastructure supporting them may become increasingly important considerations.
Dheendsa says, "This raises a broader question. What happens when the economics and resource requirements of continued centralized scaling begin to influence where and how AI can be deployed?"
His perspective points toward a different infrastructure model, where some forms of intelligence can operate locally, using smaller models and more efficient information systems. Morphos began exploring this direction before infrastructure constraints became a prominent part of mainstream AI discussion, according to Dheendsa. The company's work has consequently focused on making useful AI capabilities practical on constrained hardware, including machines operating with limited power, memory, connectivity, and physical space.
"The vast majority of use cases can actually be achieved using a small language model in conjunction with our technology," Dheendsa states. That observation reframes the role of scale. For some applications, the relevant question may become the amount of intelligence required for a particular task, rather than the maximum size of the model available.
Morphos' Green Vectors technology illustrates that premise through information organization. Traditional AI vectorization can produce large collections of numerical representations as information becomes searchable. According to Dheendsa, Green Vectors is a system for organizing related information into compact semantic structures, allowing substantial knowledge collections to occupy less memory. In the company's Project Gutenberg benchmark, an index containing roughly 15 million conventional vectors was reduced from approximately 260 GB to 1.3 GB, while accuracy went up to the 99th percentile and query speed increased by 32%.
The significance extends beyond storage. "If useful knowledge can occupy a smaller computational footprint, AI may become practical on hardware traditionally associated with lightweight computing," Dheendsa states. Morphos has demonstrated the underlying concept by running a searchable Wikipedia knowledge base alongside a small language model on a Raspberry Pi-class computer. The example suggests a broader possibility: intelligence can be placed closer to the point where information is generated, and decisions are required.
That idea becomes particularly relevant for autonomous machines. Morphos' DarkStar software architecture and EdgeRunner hardware platform are designed around local processing, with mission-specific intelligence intended to operate on devices such as drones, robots, vehicles, vessels, and remote sensors. For these systems, connectivity can vary considerably, while power, memory, weight, and heat remain persistent engineering considerations.
"The machine should be able to take its intelligence with it," Dheendsa remarks. In practical terms, that could allow a robotic system to retain searchable knowledge locally, process sensor information on-device, and receive mission-specific intelligence without requiring every interaction to travel to a distant data center.
The implications reach into business and national technology strategy. Smaller organizations could gain additional options for deploying AI where large-scale computing infrastructure is difficult to justify. Industries operating underground, underwater, offshore, or in remote environments could have greater scope for locally processed intelligence. Governments may also increasingly consider how computing efficiency, energy requirements, and infrastructure independence factor into AI policy and investment.
The Stanford HAI report's findings on AI investment and infrastructure make this question especially timely. As capability advances, the capital required to participate in the AI economy is also expanding. That creates an opening for efficiency to become an important dimension of technological progress.
Dheendsa's broader thesis is about expanding the range of architectures available. "We have to rethink the entire approach from the ground up," he notes. "We need to look at systems where intelligence can be distributed closer to the machines, people, and environments that use it."
AI's next phase may consequently involve a broader architectural shift, from concentrating intelligence in increasingly large centralized systems toward distributing appropriately sized intelligence across a wider range of devices. For Morphos, that means exploring how compact models, efficient memory structures, and purpose-built edge hardware can work together.
As AI becomes more deeply embedded in physical systems, the quality of the intelligence may remain important, while the efficiency of delivering that intelligence could become equally consequential. The longer-term question may be about how much useful intelligence can be delivered within the resources available.
© Copyright IBTimes 2026. All rights reserved.