24th November 2025
This article comes from our ‘AI Infrastructure as a Trojan Horse for Climate Infrastructure’ whitepaper, published October 2025.
TL;DR
AI is shifting from training to inference, and by 2030 most workloads will need smaller, faster, distributed data centres close to users and energy.
Modular inference centres cut latency and environmental impact while enabling heat reuse, lower water use, and seamless integration with local renewables.
With intentional design, these centres become climate infrastructure, anchoring clean power, supporting low-carbon materials, and delivering real community benefits.
The Arrival of the Inference Era
For the past few years, “AI infrastructure” has mostly meant one thing: massive clusters built for training ever-larger models. These billion-dollar compute farms and trillion-parameter breakthroughs have dominated the headlines. But they’re only half the story — and increasingly, the smaller half.
By 2030, as much as 75% of AI workloads are expected to come from inference — the real-time, user-facing side of AI that powers everything from autonomous vehicles and robotics to healthcare and live video applications. This shift is already clear in trends like the rise of test-time scaling and edge AI. Training creates intelligence; inference makes it useful in the world.
And critically: inference is reshaping the physical architecture of AI.
Unlike training, which benefits from scale and centralisation, inference demands something entirely different. It must be close to users, close to energy, and close to communities. It must respond in milliseconds, operate continuously rather than in scheduled cycles, and run inside applications where reliability is non-negotiable. These characteristics are explored in depth in this overview of the differences between training and inference. This shift is driving a quiet revolution: a move toward right-sized, modular, and distributed data centres.
Deep Learning Explained | Source
Why Inference Wants to Be Distributed
Inference behaves differently from training: it’s latency-sensitive, continuous, and embedded directly in user-facing applications. These characteristics make it far better suited to a distributed model, as outlined in this comparison of the deep learning workflow.
1. Lower latency, higher resilience
Putting compute closer to users cuts response times and reduces the impact of outages. A distributed footprint also keeps more data local, strengthening privacy and sovereignty, as captured in this overview of edge computing.
2. Turning externalities into community benefits
Modular centres can integrate circular systems from day one. Dry cooling cuts water use by up to 70%, highlighted in this analysis of circular cooling solutions, while heat reuse projects can feed pools, greenhouses, or district networks, as shown in examples of data-centre heat reuse.
3. Built to partner with renewables
Smaller centres can colocate with geothermal, hydro, or curtailed solar and wind, improving both economics and grid flexibility. Research on data centres in future energy systems shows how they can align workloads with renewable availability.
4. Faster, cheaper, lower-risk deployment
Modular designs can be deployed in weeks rather than years, scaled incrementally, and financed with lower upfront capital. This reduces stranded-asset risk, explored in this piece on the environmental costs of unused infrastructure.
5. A demographic and geopolitical imperative
High-growth regions like Africa and South Asia are seeing rapid population expansion, with projections from the UN on global demographic shifts and analyses of age distributions. Yet cloud infrastructure remains unevenly distributed, as shown in global population projection maps and data centre density comparisons. Distributed inference centres can help close this gap.
Top: UN population growth projections (2015-2100) (39). Bottom: number of data and cloud centres. Countries that have a higher number have a darker red colour. We see that the geography of data and cloud centres is uneven, with significant heterogeneity across counties. From both figures it is evident the difference between future growing demand and current deployment of data centres. | Top Source | Bottom Source
Mind-map of the seven pillars of data centres as climate Trojan horses | Source: Opna
Read more about creating AI infrastructure as climate infrastructure in our latest whitepaper, or contact us directly at hi@opna.earth.
