Dr. Sakyasingha Dasgupta is the founder and Chief Executive Officer at EdgeCortix Inc., a leading fabless semiconductor company from Japan.
getty
For many leaders, artificial intelligence (AI) still brings to mind vast data centers filled with power-hungry GPUs. Although that picture isn't entirely wrong, it’s become increasingly incomplete. A new generation of AI is emerging beyond the cloud or traditional data centers. This new class of AI—AI agents—won't just passively respond to queries; they're built to act, making decisions and executing tasks in the real, physical world.
Today, AI agents are appearing on factory floors, in robots, inside telecom base stations, aboard drones and on satellites. They're really good at planning what needs to be done, using the right tools and making decisions quickly. What's interesting is that I've found that these AI agents work better when they're close to the real world, rather than being far away in cloud servers like traditional AI systems. They can react faster and make better decisions because they're right where the action is happening.
As a result, AI agents depend not only on massive cloud infrastructure but also on power-efficient processors operating at the edge or the "thick edge," meaning the next era of AI may be defined less by raw computing power and more by how much intelligence it can deliver per watt.
Scenarios like a quality-control system on a manufacturing line that detects defects in real time or a drone navigating through unpredictable conditions can't afford the latency of sending every frame to a remote cloud. Even a 100-millisecond response delay can be the difference between a safe maneuver and a collision in these examples.
Running AI locally changes the equation. Decisions happen instantly, sensitive data stays near its source and systems continue operating even when connectivity becomes unstable.
The cloud will remain essential for training models and orchestrating fleets of devices. But when perception, reasoning and action converge, the most valuable computation increasingly happens onsite.
Real-world action introduces constraints like power budgets, thermal envelopes, memory bandwidth and environmental stress. Engineers describe this as size, weight, power and cost (SWaPC). Any system that flies, drives, sails or orbits must operate within tight limits. Every watt saved extends endurance. Every gram removed enables additional payload or longer operation.
In these environments, the meaningful metric isn't spec-sheet TOPS (trillions of operations per second) but how many useful decisions a system can make per joule. Models that perform perfectly in a lab may throttle inside a 45-degree telecom cabinet next to a vibrating motor. That failure is rarely about the model itself. System design is crucial. Hardware, software and thermals must be treated as a unified whole from the start.
AI-enabled laptops and smartphones are only visible on the edge. Cameras, robots, industrial kiosks, vehicles, satellites and tens of thousands of telecom sites are all becoming intelligent endpoints. But each environment brings its own constraints. Managing these heterogeneous fleets quickly becomes more complex than building any single device.
Success in edge AI depends on more than fitting a large model into a small box. Organizations must deploy systems safely, update them reliably and observe behavior across the entire stack.
Some of the clearest insights come from the hardest of constraints—space and defense. Onboard AI reduces the need to transmit raw data back to Earth, identifying events in real time and enabling autonomy when communication delays make human intervention impossible. These same capabilities translate directly into commercial applications: anomaly detection in manufacturing, navigation in degraded environments and logistics under severe weather.
Telecommunications networks are also becoming a proving ground. Operators are exploring AI-enhanced radio access networks where the same hardware accelerates both radio functions and AI processing, especially inference. This convergence reduces power consumption across thousands of sites and embeds intelligence directly into the infrastructure that connects everything else.
For technology leaders, a few themes are starting to emerge:
1. Workloads must be designed with locality in mind, so real-time inference happens where data is created.
2. Power, thermals and memory bandwidth are no longer secondary concerns but are defining what an AI agent can reliably do in the field.
3. AI agents need built-in observability, guardrails, security and reliable rollback paths because agents acting in the physical world must be monitored and controlled with precision.
4. The metrics that matter have now shifted—decisions per joule, real‑world latency and operational reliability across hardware and software reveal far more than model size or peak benchmarks.
The next decade of AI will belong to systems that can make intelligent decisions closer to where data is created and with far less energy. In the end, real competition will be about who can turn millijoules of energy into reliable, real-world actions on a global scale.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。