Matilda Bailey is a distinguished networking specialist whose work sits at the intersection of cellular innovation, wireless infrastructure, and the next generation of computing solutions. As organizations grapple with the dual pressures of rapid Artificial Intelligence adoption and stringent environmental, social, and governance commitments, her expertise provides a vital bridge between high-performance tech and sustainable operations. In this discussion, we explore the shifting landscape of data center demands and the potential for edge computing to recalibrate how we process information without compromising the planet’s future.
The conversation covers the dramatic rise in power consumption driven by AI infrastructure and the specific role edge architectures play in reducing the energy tax of data transmission. We delve into the nuances of localized processing, the hidden costs of hardware underutilization, and the complexities of a fragmented NPU ecosystem. Finally, we examine the metrics that actually matter for green AI and how emerging trends like small language models and agentic AI are reshaping the very definition of efficient networking.
With data center power demand in the United States projected to more than double from 31 gigawatts in 2025 to 66 gigawatts by 2027, what is driving this unprecedented surge and how does it threaten corporate sustainability goals?
The primary engine behind this massive leap is the rapid buildout of infrastructure dedicated to AI. Currently, AI workloads account for approximately 14% of global data center demand, but that figure is expected to soar to 27% by 2027, creating a significant tension for organizations that have already signed public sustainability commitments. When you look at the sheer scale of moving from 31 gigawatts to 66 gigawatts in just two years, you realize that traditional centralized cloud facilities are struggling to keep up with the carbon targets they once thought were achievable. This surge makes it incredibly difficult to build a sustainable AI roadmap without fundamentally changing where our inference workloads actually run. Many companies are finding that their current AI usage is simply not aligned with their green energy goals, forcing a total re-evaluation of their hardware and deployment strategies.
Many experts suggest that moving AI to the edge could be a primary solution to these power issues, but what are the specific mechanisms that allow edge computing to offer a sustainability advantage over centralized clouds?
The sustainability advantage of edge AI is really rooted in the reduction of data transmission energy and localized processing efficiency. Every time you route raw data to a centralized cloud data center, you are consuming energy at every single network hop along that journey. Edge architectures mitigate this by processing data locally, meaning they only send summaries, exceptions, or specific events upstream rather than maintaining continuous, power-hungry raw data streams. Furthermore, while cloud GPU infrastructure is designed for massive training at scale, it is often vastly overpowered for repetitive inference tasks, leading to high overhead even when underutilized. At the edge, we use hardware specifically designed for low-power inference, which doesn’t require the massive, active cooling systems or water-cooled infrastructure that typically account for up to 30% of a large data center’s energy consumption.
While the benefits of localized processing seem clear, what are the risks of deploying overpowered or underutilized hardware at the edge, and how can that negate the environmental gains?
The “green” story of edge computing starts to fall apart very quickly if you don’t account for the full hardware lifecycle and utilization rates. If an organization simply deploys overpowered hardware at every remote site and runs those models continuously without a clear strategy, they may find they are actually increasing their total energy footprint. Underutilizing equipment while still carrying the full operational overhead is a common pitfall that makes the energy story much weaker. It is not just about moving the workload; it is about ensuring the hardware is right-sized for the task so that you aren’t wasting electricity on idle compute power. We also have to consider that edge devices, while they don’t need the 30% energy-intensive cooling of a data center, still represent a carbon cost in terms of manufacturing and replacement cycles that must be balanced against the savings in transmission.
When we look at the practical application of these technologies, what specific types of environments or workloads truly necessitate an edge-first architecture rather than a cloud-based one?
The decision to move to the edge should always be driven by the workload itself rather than a general preference for one architecture over another. Edge computing performs exceptionally well in environments that are latency-dependent, such as autonomous systems, or in situations where connectivity is limited, like industrial monitoring or remote manufacturing quality control. In these scenarios, the upfront cost of the hardware is usually offset by the significant reduction in ongoing cloud transmission fees and the energy required to move high volumes of data. Furthermore, for mission-critical infrastructure where data sensitivity creates compliance risks, processing at the edge allows the system to survive and function without a constant link to a centralized facility. It is less about asking if edge is “better” than cloud and more about identifying which environment the system needs to survive in to provide the most value.
One of the major hurdles you’ve mentioned in the past is the fragmented nature of the NPU ecosystem. How does this technical complexity affect the return on investment for companies trying to scale their edge AI?
The current state of the Neural Processing Unit (NPU) ecosystem is deeply fragmented, which creates a significant barrier for IT departments that lack specialized machine learning engineering teams. Because there is no dominant market standard for these purpose-built chips, AI models often have to be re-engineered or “hand-tuned” for every specific chip variant, which can absolutely kill the ROI case for a large-scale deployment. We see companies like General Electric’s rail division or Rentokil succeeding because they built their IoT infrastructure years ago and have the internal engineering depth to handle model updates, patching, and hardware replacement at scale. For smaller organizations, the need to manage model drift and hardware-specific optimizations across thousands of devices can become an operational nightmare. Until the ecosystem matures and becomes more standardized, the complexity of managing these specialized chips remains a major deterrent to widespread edge adoption.
How should organizations shift their performance metrics to accurately reflect the sustainability of an edge deployment, especially when traditional benchmarks like Power Usage Effectiveness might not apply?
Traditional benchmarks like Power Usage Effectiveness, or PUE, are excellent for measuring facility-level efficiency in a centralized data center, but they have almost no direct applicability to edge deployments where there is no facility overhead to measure. Instead, leaders should focus on energy consumption per inference, which measures the actual joules or watt-hours consumed per AI transaction. This allows for a direct before-and-after comparison of the energy required for a workload when it is moved from the cloud to the edge. Additionally, companies must track their total carbon footprint per AI workload, which includes the electricity source and the network transmission energy involved. It is also becoming critical to report on Scope 1, 2, and 3 emissions; while the SEC requires reporting on direct and purchased electricity emissions, Scope 3 extends to the entire supply chain, including the hardware manufacturing and the infrastructure of the cloud vendors themselves.
With the emergence of Small Language Models and Agentic AI, how do you see these trends influencing the balance between local and cloud-based processing in the coming years?
We are seeing a major shift toward a hybrid pattern where Small Language Models, or SLMs, handle the routine, repetitive, and event-driven work locally, while larger cloud models are reserved for tasks that require deep reasoning or massive context. These SLMs are compact and task-specific, making them perfectly suited for the lower-power hardware we find at the edge. However, the rise of Agentic AI—where autonomous agents run continuously and call models on every cycle—presents a new challenge because they can consume just as much compute locally as they would in the cloud. The sustainability benefit of these agents really depends on technical discipline: using smaller models whenever possible, caching results to avoid redundant compute, and escalating to the cloud only when absolutely necessary. This shift will likely pull infrastructure investment toward distributed colocation environments that are physically closer to existing enterprise systems, which is inherently more efficient than the old hyperscale model.
As regulatory pressures like the EU’s AI Act and various climate disclosure rules become more prominent, what is your forecast for how organizations will approach AI infrastructure in the near future?
My forecast for the intersection of AI and sustainability is that we are moving toward a period of intense workload-level accountability. In the very near future, the standard question will no longer be a simple “Can we use AI for this task?” but rather a much more complex calculation: “What specific model are we running, where is it located, what is its total energy profile, and what is the definitive business value it provides?” The regulatory environment, particularly with the EU’s Corporate Sustainability Reporting Directive and the AI Act, is making energy and carbon disclosure a business necessity rather than a voluntary goal. Organizations that are already tracking their carbon footprint per workload will have a massive compliance head start, while those that continue to route everything to the cloud without a strategy will find themselves struggling to justify the energy costs. We will see a significant shift toward “green AI” strategies where the choice of architecture is dictated by a strict balance between compute needs and environmental impact.
