The Gist: Executive Overview

While semiconductor export controls dominate geopolitical attention, memory supply and electrical power are emerging as the binding physical limits on AI deployment. Micron's projection of persistent shortages through 2030 signals that high-bandwidth memory availability now gates model training and inference capacity regardless of GPU access. Simultaneously, Google's willingness to fund complete power generation infrastructure—rather than wait for utility investment—reveals that electricity supply cannot keep pace with hyperscale buildout timelines. NTT's plan to double data centre capacity to 4 gigawatts occurs as operators accept they must act as quasi-utility providers to secure sites.

These infrastructure constraints compound differently across geographies. China's H200 licensing breakthrough matters less if memory shortages limit what can be built with those chips, while the Iran conflict's helium supply disruption threatens South Korean and Taiwanese fabs independent of U.S.-China technology competition. The result is a fracturing of advantage: access to frontier chips no longer guarantees AI capability if supporting infrastructure—memory, power, cooling, rare earths—cannot scale in parallel. This favours actors with integrated control over full technology stacks and supply chains rather than those optimising single components.

The UK's copyright reversal, EU's move to ban non-consensual sexual image generation, and the Department of Defense designating Anthropic an unacceptable supply-chain risk over military use red lines illustrate how AI governance is fragmenting into jurisdiction-specific frameworks with minimal harmonisation. Britain's attempt at a permissive training regime collapsed under creative industry pressure, forcing the government back to square one on a question the EU has already answered differently through AI Act carve-outs. Meanwhile, DoD's Anthropic designation reflects American willingness to accept national security restrictions that European or Chinese frameworks do not impose.

This regulatory divergence creates rising compliance costs for companies attempting global deployments and forces strategic choices about which markets to prioritise. China's state-directed mobilisation of thousands of one-person AI startups—supported by government infrastructure and compute credits—represents a development model incompatible with Western venture frameworks, yet produces distributed capability harder to target with sanctions. The pattern suggests AI development is settling into regional spheres optimised for local regulatory environments rather than converging on global standards, with implications for cross-border data flows, model training practices, and interoperability.

Walmart's restructuring of its OpenAI partnership—abandoning underperforming Instant Checkout agents in favour of embedding Sparky into ChatGPT and Gemini—exemplifies the gap between agentic AI marketing claims and revenue-generating deployments. Tencent's failure to articulate monetisation despite doubling AI investment triggered share price declines, while Meta inadvertently exposed company data through rogue agents and developers report agents destroying entire databases. PwC's directive that partners resisting AI adoption have no place at the firm occurs as enterprises pilot extensively but struggle to demonstrate productivity gains beyond proof-of-concept.

The pattern suggests the current hype cycle may be approaching inflection as buyers demand commercial viability proof. Arena leaderboard's influence on funding and PR cycles—despite being funded by the companies it ranks—illustrates how evaluation mechanisms are entangled with commercial interests. Meanwhile, operational risks from immature governance frameworks create asymmetric outcomes: firms mastering AI safety gain competitive advantage while those deploying without robust controls face material damage exposure. The disconnect between deployment velocity and operational maturity is widening, with few incentives for companies to publicly discuss failures.