Geopolitics & Sovereign Positioning

Anthropic's Claude Mythos Preview, released April 7, has self-taught capabilities to penetrate software infrastructure systems previously considered among the most hardened targets available. According to analysis in War on the Rocks, Mythos effectively erases the barrier between nation-state-level offensive cyber capability and non-state actors — a threshold that U.S. planners had long relied upon as a structural buffer. The implication is not merely that cyberattacks become more frequent, but that the asymmetric advantage previously held by well-resourced state actors like China's PLA or NSA-tier operators is now potentially accessible to a much wider field of adversaries, proxies, and criminal networks acting on behalf of states.

The White House held what was described as a 'productive' meeting with Anthropic CEO Dario Amodei, according to BBC News and South China Morning Post. Critically, this meeting occurred despite an earlier dispute between the Pentagon and Anthropic over the terms of military use of its models. The fact that the Trump administration moved to rebuild the relationship rather than restrict or sanction Mythos deployment suggests Washington has concluded that the offensive utility and economic significance of Anthropic's capabilities outweighs the risk of distancing — a calculation with direct parallels to the early nuclear period. The CFR frames Mythos as an inflection point, noting defenders face a race against time across a vast and under-secured attack surface.

CSET's new analysis of Chinese military procurement documents provides the most direct open-source evidence to date that the PLA has been systematically targeting advanced U.S. compute — including Nvidia GPUs — for AI-enabled military applications. As reported by CSET, the procurement records function as a revealed-preference dataset: rather than inferring Chinese military AI ambitions from capability assessments or strategic doctrine, they show direct purchasing intent. This substantially strengthens the policy case for controls that critics have dismissed as economically costly without clear security benefit.

The analysis lands alongside Jensen Huang's public warning that DeepSeek optimising on Huawei chips would be 'a horrible outcome' for the U.S., as quoted by South China Morning Post. Huang's argument is structural: if the models that diffuse globally are trained and optimised on a Chinese hardware stack, the standards, APIs, and architectural assumptions baked into global AI infrastructure will reflect Chinese rather than American technological choices. This is a longer-horizon argument than the current export control debate, which focuses on preventing near-term military acquisition, and suggests the U.S. is managing two distinct timelines simultaneously.

China has declared a complete RISC-V ecosystem established domestically, with Chinese Academy of Sciences researchers presenting the Xiangshan processor series at the Zhongguancun Forum. As South China Morning Post reports, this is not merely a chip development story but a deliberate strategic choice to adopt an open-source instruction set architecture that falls outside U.S. export control jurisdiction and cannot be revoked under technology licensing agreements. The contrast with ARM — which is subject to U.S. pressure points — is deliberate. A functional RISC-V ecosystem means China can develop, manufacture, and export chips for AI workloads without licensing exposure to American IP.

This architecture bet intersects directly with China's open-source LLM strategy. Moonshot AI's release of Kimi K2.6 as open source, alongside a stated consensus among Alibaba, ByteDance, and Tencent to promote open-source models per South China Morning Post, is not purely a commercial decision. Open-source models optimised for RISC-V hardware would create a self-reinforcing Chinese AI stack that is exportable to the Global South without triggering U.S. licensing or export control provisions. The strategy mirrors how China used open-source telecom standards in 5G to establish de facto infrastructure dominance in developing markets.

China's December 2024 ban on gallium exports to the United States landed against a U.S. National Defense Stockpile holding zero gallium reserves, as detailed by War on the Rocks. China controls 99 percent of global primary gallium production and has now converted that chokehold into an active policy instrument. Gallium nitride is critical for next-generation semiconductors used in high-frequency radar, electronic warfare systems, and the power electronics that underpin AI inference at the edge — precisely the applications where military AI capability is being built. The zero-stockpile position is not a supply chain management failure; it reflects a decade-long institutional assumption that adversarial resource weaponisation was a theoretical rather than operational risk.

The War on the Rocks analysis explicitly frames this as a repeat of the silicon manufacturing mistake — ceding domestic production capacity over years of cost optimisation, then discovering that dependency has become a strategic liability only after the leverage has been applied. The remediation timeline for building alternative gallium supply chains or synthetic substitutes is measured in years, not months. In the interim, U.S. defense AI hardware programs face either cost increases from alternative sourcing or program delays — both of which benefit China's relative position in military AI deployment timelines.