AI Arms Race Deepens: Cyber Thresholds Crossed, Supply Chains Exposed

The White House's pivot toward closer Anthropic cooperation following Mythos's demonstrated offensive cyber capabilities is the clearest illustration yet of governance being outpaced by deployment. The earlier Pentagon dispute over military use terms was resolved not through regulatory pressure but through Washington's recognition that it could not afford to be without the capability — a pattern that will repeat each time a new capability threshold is crossed. Trump's public endorsement of an Anthropic-DoD deal, alongside Anthropic's simultaneous advance into European banking via Mythos, reveals a company that has understood this dynamic and is deliberately targeting sectors where regulatory complexity advantages an early-moving, compliant-by-design incumbent.

The commercial dimension reinforces the governance problem. Amazon's deepened infrastructure commitment to Anthropic, OpenAI's consultancy channel and PE joint venture, and Adobe's agentic platform pivot all create switching costs and distribution dependencies that make future restriction or displacement progressively more difficult. By the time governments are ready to regulate these architectures, they will already be embedded in defense procurement, financial services infrastructure, and the enterprise software stack of hundreds of PE portfolio companies. The window for structural intervention is narrowing faster than regulatory processes are moving.

China's RISC-V ecosystem declaration, coordinated open-source model releases from its largest platforms, and the gallium export ban collectively represent three layers of the same strategy: build a sovereign stack, weaponise critical dependencies in the transition period, and export the result to non-aligned markets before U.S. architecture lock-in becomes irreversible. The CSET procurement evidence confirms the military dimension is not separable from the commercial one — the PLA was acquiring U.S. chips precisely because they were the world standard, and the RISC-V pivot is an acknowledgement that that standard can no longer be reliably accessed. Jensen Huang's framing of Huawei-optimised DeepSeek as a structural defeat is accurate: if globally diffused models are trained and benchmarked on Chinese hardware, the APIs and architectural assumptions of global AI infrastructure shift accordingly.

The U.S. side of this competition is fragmented in ways the Chinese side is not. Hyperscaler-model partnerships are consolidating compute into Amazon-Anthropic, Microsoft-OpenAI, and Google-DeepMind verticals, but sovereign compute initiatives — the UK's £500 million fund, the EU AI continent programme — remain at early procurement stages. Meanwhile, middle powers are explicitly refusing to choose sides, extracting technology transfers and infrastructure deals from both blocs. The Gulf states hosting Cerebras's primary revenue base while also engaging with Chinese infrastructure players exemplify a swing-state dynamic that reduces both Washington's and Beijing's ability to use AI partnership exclusivity as a diplomatic instrument.

Three separate signals this week converge on the same vulnerability. Helium supply disruptions linked to Qatar's geopolitical exposure, the U.S. gallium stockpile gap exposed by China's export ban, and ASMPT's record results confirming advanced packaging equipment remains a multi-quarter bottleneck all represent categories of infrastructure risk that post-2022 supply chain frameworks systematically underweighted. The 2021-22 chip shortage focused attention on wafer capacity and lithography equipment — the visible, discussed bottlenecks. The next disruption cycle is more likely to originate in the lower-visibility input layer, and the geographic concentration profile of these inputs — Qatar for helium, China for gallium, a small number of equipment vendors for advanced packaging — means moderate disruption propagates rapidly to fab-level constraints with limited buffer.

The packaging constraint has the most direct near-term implication for hyperscaler buildout commitments. TSMC's CoWoS capacity expansion depends on equipment with 12-18 month lead times, meaning the gap between the $300 billion-plus in announced hyperscaler infrastructure spending and actual compute coming online will be determined by packaging yield as much as wafer starts. If packaging capacity underperforms against the current expansion trajectory, announced data centre timelines will slip — a risk that is not priced into current infrastructure investment theses.