The Gist: Executive Overview

Meta's potential 20% workforce reduction to fund infrastructure buildout, chip material costs doubling from supply chain disruptions, and the US withdrawal of global export permit rules all point to the same underlying dynamic: the capital intensity of AI development is forcing zero-sum resource allocation decisions. Companies are financing AI capabilities not through incremental investment but by cannibalising existing operations—cutting headcount, accepting margin compression, or abandoning regulatory frameworks deemed too restrictive. This is not a funding constraint problem but a prioritisation problem, revealing that AI transformation cannot coexist with business-as-usual operations.

The pattern extends to nation-states. The UK is weaponising public procurement to build domestic capabilities, while China's capital markets are pouring billions into indigenous alternatives despite export controls. Russia's growing digital isolation constrains its access to global AI networks, creating divergent development pathways. Meanwhile, chipmaking material shortages threaten to delay capacity expansions regardless of available capital. The result is an increasingly fragmented global AI ecosystem where geographic, regulatory, and economic boundaries determine access to capabilities—a fundamental shift from the previous decade's assumption of technology diffusion across borders.

The Anthropic-Pentagon confrontation, Meta's reported safety team vulnerability in layoffs, and documented AI-induced psychosis all demonstrate that corporate AI safety mechanisms are inadequate when confronted with commercial incentives or state-level pressure. Anthropic has removed core safety documentation as it negotiates military contracts, Meta is potentially cutting safety functions while scaling capabilities, and the industry only discovered chatbots could trigger psychosis after deployment to billions of users. No major lab currently conducts systematic psychological harm testing on vulnerable populations before launch, and safety evaluation frameworks focus on isolated outputs rather than interaction dynamics or emergent harms.

Defence procurement is accelerating this erosion. The Pentagon's $20 billion Anduril contract and broader military AI spending create lucrative markets where safety constraints are reframed as strategic liabilities. National security framing positions any reluctance to deploy as weakness in competition with China, creating institutional pressure that voluntary commitments cannot withstand. The North Korean exploitation of AI for sanctions evasion adds another dimension—adversaries are already weaponising openly available capabilities, creating pressure to deploy first and govern later.

AWS partnering with Cerebras for inference chips, subsea cable delays from Middle East conflict, and the UK pushing sovereign procurement all reflect a shift in competitive focus from algorithmic breakthroughs to infrastructure control. Hyperscalers are curating portfolios of specialised silicon to reduce Nvidia dependency and lock in workloads, while governments are using procurement mandates to shape domestic capabilities. Physical infrastructure—data centres, subsea cables, chip manufacturing capacity—is becoming the constraining resource rather than algorithmic innovation or engineering talent.

This infrastructure-first dynamic is creating new chokepoints and vulnerabilities. Meta's 2Africa cable delays compound bandwidth constraints precisely as distributed training requires massive inter-facility data transfers. Material cost inflation threatens chip production economics regardless of fab capacity. Export controls on AI accelerators create regulatory fragmentation that forces parallel infrastructure stacks. The result is that access to compute and connectivity increasingly determines who can participate in frontier AI development, shifting competitive advantage toward actors who control physical infrastructure rather than those with superior algorithms or talent.