The Gist: Executive Overview

Multiple developments signal that energy access and thermal management now constrain AI deployment more than semiconductor supply. SoftBank is building 10GW of dedicated generation capacity alongside its Ohio data centre. Nvidia and Emerald AI are partnering with power companies on flexible facilities that can modulate consumption. OpenAI is negotiating to purchase fusion power output directly from Helion. These moves indicate that AI infrastructure strategy now prioritises securing power before securing chips — a fundamental reordering of deployment dependencies.

Simultaneously, thermal constraints are becoming the primary limit on semiconductor scaling. Nvidia's next-generation racks will consume 160kW and require full liquid cooling, matching power density that makes air cooling obsolete. Heat flux exceeding 1,000 W/cm² in advanced chips creates metrology gaps that constrain performance validation. The combined effect is that AI clusters face physical density limits independent of Moore's Law, forcing operators to build more facilities rather than pack more compute into existing footprints.

Huawei's Atlas 350 claiming to outperform Nvidia's export-controlled H20 demonstrates that semiconductor restrictions are accelerating China's development of competitive alternatives rather than preventing it. Alibaba's new inference chip, China's dominance in robotics supply chains for motors and rare earth materials, and the country's positioning around low-cost energy and manufacturing integration all reveal structural advantages built through industrial policy. Even Nvidia's CEO acknowledges China as 'formidable' in physical AI due to supply chain control.

The deployment of over $2 billion in Chinese AI surveillance systems across Africa extends this advantage beyond hardware to installed base and operational dependencies. Once African governments adopt Chinese infrastructure, maintenance contracts, software updates, and data integration create lock-in that Western competitors cannot easily displace. China's open-source AI strategy compounds the pressure by flooding global markets with capable free models that force US firms to continuously justify premium pricing.

The rapid restructuring of Google's browser agent team as the industry pivots toward coding agents demonstrates that even leading labs lack conviction about which agent modalities will matter most commercially. OpenAI is consolidating its entire research organisation around building an autonomous AI researcher while merging consumer products into a single superapp. Gemini's task automation reaching users shows genuine app control but is slow and clunky in practice, revealing the gap between lab capabilities and production-ready consumer experiences.

This volatility creates execution risk for enterprises planning around specific agent types and suggests the path from proof-of-concept to reliable product is longer than anticipated. The convergence of multiple labs on autonomous systems as the next capability frontier represents a significant strategic bet that scaling current architectures with agentic scaffolding will yield transformative capabilities — but if the approach hits fundamental barriers, it could create a capability plateau across the industry simultaneously.