Access Partnership: Hybrid AI Infrastructure Needed by 2030 to Meet AI Compute Demand

12 November, 2025 (Washington D.C.) – Access Partnership recently released a landmark report, AI Beyond Data Centers: On-device, On-demand, Everywhere, revealing that the world’s current AI infrastructure is on track to fall dramatically short of projected compute demand by 2030. It argues that a “hybrid” approach that combines cloud, edge, and on-device computing – training models in data centers and running more inference on user devices – is the most practical way to close the gap.

The study projects a global inference compute shortfall of 18.7 QFLOPs, a way of measuring compute power, by 2030, equal to about 130% of the capacity currently planned by operators. Over the same period, compute demand for AI is expected to rise 125-fold from 2024 levels, as usage spreads across consumers and enterprises.

The report argues trying to serve this demand through data centers alone would be resource-intensive. It estimates an extra $2.8 trillion in capital spending would be required, alongside large increases in electricity and water use for cooling – pressures that are already delaying data center projects in several markets.

The report models the resource savings from shifting a significant share of inference to devices such as smartphones, PCs, and connected sensors. It finds power use could fall by up to 90%, water efficiency could improve by 96%, annual water savings could reach 315 billion liters (equivalent to about 126,000 Olympic swimming pools), and grid load could drop by 507 TWh — roughly South Korea’s yearly electricity generation.

Regional disparities are stark. Sub-Saharan Africa and Latin America could face unmet inference demand of up to 98% by 2030, the analysis suggests. Access Partnership argues that device-side inference can broaden access where cloud AI capacity is limited, while freeing data centers to focus on training the largest models.

“AI has reached an inflection point where infrastructure, not imagination, is the constraint,” said Abhineet Kaul, Director at Access Partnership. “The path forward is not simply to build more data centers — it’s to distribute artificial intelligence intelligently. On-device and hybrid AI architectures can meet compute demand sustainably, securely, and at scale.”

The report calls for policymakers to plan for a hybrid AI ecosystem: support technical standards and labelling for device-level AI, fund R&D on lightweight models, and run regulatory sandboxes in sectors such as finance and health. These policy actions, it says, would improve resilience, reduce latency and bandwidth needs, and help organizations comply with privacy rules by keeping more data on the device.

This research comes amid growing concerns over the energy consumption of data centers for AI use. The International Energy Agency forecasts global data center electricity use is set to double to 945 TWh by 2030 thanks to AI, while the EU announced plans to address energy efficiency of data centers.

For media enquiries, please contact:

Harry Curtis, Head of Communications, Access Partnership – [email protected]

Notes to editors

To download the report, click here.

What the terms mean

The analysis expresses compute in QFLOPs. Floating-point operations per second, or FLOPs, are a way of counting how many mathematical calculations a computer can perform each second. A QFLOP stands for quettaFLOP, equal to 10³⁰ FLOPs. This approach allows demand and planned capacity to be compared on a common basis.

Training refers to building and refining models – typically done in data centers -while inference is the day-to-day use of trained models to generate outputs for users. At scale, inference becomes the dominant workload and can be distributed across devices.

About Access Partnership

Access Partnership works with businesses, governments, and multilateral institutions to unlock new markets, drive commercial success, and create solutions that advance humanity. Our multidisciplinary team includes geopolitical analysts, regulatory specialists, and communications strategists, as well as economists, engineers, physicians, and scientists. We help ambitious organisations turn ideas into real-world outcomes quickly and decisively. Whether you’re expanding into new markets, launching new solutions, or tackling global challenges, we combine commercial insight, public-policy and regulatory analysis, and deep networks to help you move with confidence and clarity. Find out more at  accesspartnership.com