Geopolitics & Sovereign Positioning

A research team including Huawei Technologies has completed post-training of the DeepSeek-V4-Pro model using Ascend 910C chips, according to South China Morning Post. This is strategically significant because it represents a leap beyond AI inference — running finished models — into post-training, a computationally intensive process that requires sustained high-bandwidth memory and chip-to-chip interconnect performance closer to what Nvidia's H100 and H200 deliver. The export control thesis has long held that denying advanced training chips would delay Chinese frontier model development by years. This development does not invalidate that thesis entirely, but it compresses the assumed gap.

The pairing of Huawei's domestic silicon with DeepSeek's efficiency-optimised architecture is not accidental — it reflects a deliberate Chinese strategy of co-developing hardware and model architectures to work around the performance ceiling imposed by denied access to Nvidia's top-tier chips. CAS Star's decade-long bet on photonics, highlighted separately by South China Morning Post, signals that Chinese state-linked capital has been systematically building alternative compute pathways — photonic chips, domestic interconnects — that could further reduce dependence on US semiconductor architecture. The strategic implication: export controls are buying time, not halting China's trajectory, and the window during which that time advantage is militarily meaningful is narrowing.

The launch of the Beijing Space Intelligent Computing Research Institute, a state-backed entity, marks China's formal institutional commitment to space-based AI compute as a sovereign infrastructure priority, per South China Morning Post. The timing is not coincidental: SpaceX is preparing for what may be a $75 billion IPO that would fund significant expansion of its own orbital AI compute capacity through Starlink's low-earth orbit constellation. China's move represents a classic competitive dynamic — matching a US commercial capability with a state-directed programme to avoid dependence on foreign orbital infrastructure for AI workloads.

The strategic logic of space-based AI computing is that it offers latency advantages for certain real-time applications and provides a layer of redundancy that is architecturally separate from terrestrial infrastructure, which is vulnerable to targeted strikes, sanctions, and physical interdiction. For China, which faces the prospect of undersea cable severance or terrestrial data centre disruption in a Taiwan contingency, sovereign orbital compute represents both a warfighting resilience measure and a long-term hedge against the possibility that terrestrial AI infrastructure becomes a target in great-power competition. This is a confirmed institutional launch, not merely a policy announcement.

Trump's new AI executive order has drawn analysis from both CFR and the Atlantic Council as a meaningful but insufficient shift. CFR's assessment is that the order signals a departure from the administration's previous posture of near-total deregulatory permissiveness, but falls well short of creating the cybersecurity architecture needed for a country operating AI-dependent critical infrastructure at scale. The Atlantic Council's expert panel identifies language in the order that could be read as laying groundwork for federal AI procurement standards and interoperability requirements — but these remain directives to agencies rather than binding legal frameworks with enforcement mechanisms.

The geopolitical consequence of US domestic AI governance ambiguity is underappreciated. US allies making decisions about whether to adopt American AI infrastructure, standards, or export control frameworks are watching Washington's internal coherence closely. An administration that simultaneously promotes AI exports aggressively — reflected in Trump's planned meetings with AI company CEOs to discuss US investment, per BBC — while lacking a credible domestic governance framework creates a credibility problem in multilateral settings. It weakens the US position in forums like the UN's Global Dialogue on AI Governance, where Washington's ability to lead on standards depends partly on demonstrating that it governs its own AI sector.

War on the Rocks publishes a significant analytical piece arguing that the US Department of Defense's pivot to an AI-first warfighting posture contains a self-undermining logic: the publicly released frontier AI models that underpin Pentagon systems — from Project Maven's intelligence fusion to Anduril's Lattice — can be harvested by adversaries without any system breach, simply by studying the publicly available model weights and outputs, per War on the Rocks. This is a qualitatively different threat model from traditional espionage. The concern is not that adversaries will steal classified training data, but that the logic, behaviours, and decision architectures of US military AI systems are increasingly readable from commercial releases.

This matters because the US military's competitive AI advantage has been built on the assumption that model supremacy — having the most capable frontier models — translates into durable operational advantage. If adversaries can rapidly replicate that logic through distillation, the lead time between US capability development and adversary near-parity collapses. The Atlantic Council separately identifies agentic AI as enabling a new form of irregular warfare targeting financial system vulnerabilities, per Atlantic Council, which compounds the strategic risk: the same model architectures that enable US military AI also enable adversary financial and infrastructure attacks at machine speed.

CFR's analysis of the UN Global Dialogue on AI Governance frames it explicitly as a stress test for the multilateral system: whether institutions designed for an earlier era of technology governance can establish legitimate and inclusive AI oversight before national and regional frameworks calcify into incompatible regulatory blocs, per CFR. The diagnosis is that failure is not merely a missed opportunity — it actively accelerates fragmentation by signalling to swing states and Global South nations that they must choose between US-aligned or China-aligned AI governance tracks rather than participating in a genuinely multilateral alternative.

The structural challenge for the UN process is that its two most powerful members have divergent interests in the outcome. The US under the current administration has shown limited appetite for binding multilateral AI commitments that could constrain domestic commercial actors. China, conversely, has strategically engaged multilateral AI governance forums as a standard-setting opportunity while maintaining domestic AI governance that would not pass the criteria it advocates internationally. Both dynamics undermine the UN process's legitimacy with exactly the middle-power and developing-economy participants it needs to be credible.