The Gist: Executive Overview

While chip shortages persist—PC manufacturers face six-month CPU lead times, up from two weeks—the binding constraint is moving to physical infrastructure that operates on longer timescales. Panasonic reports data centre backup batteries sold out years ahead, while Micron's Singapore fab expansion requires 400-500 power transformers, more than double typical demand and exceeding any single manufacturer's annual output. Proposed federal legislation would impose a construction moratorium on new data centres entirely.

This creates a structural ceiling on expansion regardless of chip availability. Hyperscalers like Prologis are securing multi-year power capacity commitments, but heavy electrical gear—transformers, switchgear, battery systems—cannot be scaled as rapidly as server hardware. The political dimension adds unpredictability: Virginia representatives are split between economic development interests and grid concerns, illustrating how infrastructure bottlenecks are becoming political bottlenecks.

Chinese chipmaker CXMT doubled revenue to $8 billion and is positioned to challenge Western firms in high-bandwidth memory for AI accelerators ahead of a major IPO. China is projected to control 42% of global mainstream chip fabrication capacity by 2028, with Shenzhen publishing action plans for AI server supply chain production and Alibaba advancing RISC-V chip architectures as alternatives to restricted ARM and x86 designs. Meanwhile, Chinese AI models from DeepSeek and MiniMax have overtaken US rivals in token consumption.

The simultaneous progress in indigenous chips, open-source architectures, and commercial model deployment creates a closed-loop ecosystem increasingly insulated from Western restrictions. While cutting-edge lithography remains a constraint, China is succeeding in mature nodes and advanced packaging—sufficient for inference and agent workloads where performance gaps with frontier chips matter less. Tencent's integration of AI agents into WeChat's billion-user base demonstrates deployment at population scale ahead of Western democracies.

Anthropic launched autonomous mode allowing Claude to control computers without supervision, while both OpenAI and Google integrated direct purchasing into their chat interfaces—OpenAI with an Agentic Commerce Protocol, Google through Gap partnership. Arm produced its first proprietary AI inference CPU with Meta, OpenAI, Cerebras and Cloudflare as initial customers, signalling inference workloads are becoming distinct from training. Harvey and Granola commanded premium valuations for vertical agent applications over horizontal infrastructure plays.

This represents convergence across previously separate trends: autonomous agents need specialised inference silicon, creating opportunities for new chip entrants; they require different economic models than one-shot completions, driving labs toward transaction fees and enterprise subscriptions; and they demand sector-specific implementations rather than general platforms, favouring vertical integrators. The architecture shifts extend to processor design—Arm forecasts agent workloads will drive CPU demand over GPUs as orchestration becomes the bottleneck.