Building AI’s $400B backbone

Powering the Future: Unpacking AI’s $400 Billion Infrastructure Revolution

At Web Summit Vancouver 2026, Editor-in-Chief Harold Munro moderated a panel on the $400 billion AI infrastructure opportunity. A critical challenge highlighted was the escalating power demand, with global data center consumption projected to surge by 165% by 2030. This necessitates innovative solutions for scaling AI, moving beyond traditional approaches to meet the immense energy needs of advanced computing.

Mr. Rodrigo Liang, Co-founder & CEO of SambaNova Systems, detailed his company’s focus on power-efficient AI chips for inference. SambaNova aims to deliver significantly more performance per rack at 10 kilowatts, compared to typical 100-kilowatt Nvidia racks. Their specialized chips, designed for large-scale inference, enable “premium inference” for even trillion-parameter models, offering both speed and intelligence without compromise, as supported by a Stanford study.

Mr. Darrick Horton, CEO of TensorWave, explained how the power crisis drives diversification in AI GPU clusters. The difficulty in finding large, centralized power sources means utilizing multiple smaller data centers. While this poses challenges for training, it benefits inference by bringing computing closer to users. TensorWave provides a unified control plane to seamlessly manage these distributed resources, abstracting complexity for businesses.

Ms. Ami Badani, CMO at ARM, emphasized ARM’s pervasive presence, with 350 billion chips shipped globally, powering virtually every smartphone. ARM’s new AGI CPU is designed for power-efficient agentic workloads. Ms. Badani predicted AI’s inevitable shift from cloud to edge devices, driven by cost, security, and privacy concerns, akin to the evolution from mainframes to personal computing. This requires a common software architecture across the entire compute spectrum.

The concept of AI sovereignty was explored, defined as nations developing their own AI models and infrastructure to safeguard data privacy and reflect unique cultural, linguistic, and legal frameworks. Governments often fund local entities to build this infrastructure. Mr. Horton stressed that sovereign AI is crucial for governments using AI for classified information, ensuring control over data and infrastructure, and requiring investment in local data centers and AI labs.

The panel also stressed the importance of an educated workforce, with ARM actively training local populations to manage AI systems, aligning educational programs with future job demands. On preventing monopolies, Mr. Liang advocated for an open ecosystem, while Mr. Horton noted TensorWave’s partnership with AMD to counter Nvidia’s historical dominance in AI chips, promoting competition and innovation in the AI ecosystem.

Regarding inference’s architectural impact, Mr. Liang highlighted its CUDA-free nature and reliance on open Pytorch models, fostering diverse hardware. He envisions heterogeneous data centers mixing various chip types for optimal efficiency across different workloads. The balance between CPUs and GPUs is also evolving, with a rising demand for CPUs in agentic workflows. While GPUs remain essential for training, CPUs will increasingly manage the growing number of agents.

Looking to the future, Mr. Horton predicted a massive global upgrade in power infrastructure, driven by the AI boom, leading to cleaner and larger energy sources. Ms. Badani identified alternative power, innovative capital partnerships, and supply chain resilience as key. Mr. Liang concluded that AI will become real-time, 24/7, and constantly running, making efficiency, cost, and speed paramount for the next, even faster wave of technological adoption.