Geopolitics & Sovereign Positioning
Top Line
DeepSeek released its open-source V4 model with 1.6 trillion parameters and Huawei pledging 'full support' via new chips, representing the most direct challenge yet to US AI supremacy and a concrete demonstration that export controls have not halted China's frontier model development.
The US House Foreign Affairs Committee advanced 20 new export control measures — described as the 'largest significant upgrade' against China — including the Match Act targeting Chinese chipmakers' access to advanced semiconductor manufacturing equipment, though these remain at committee stage and are not yet enacted law.
A White House memo from Michael Kratsios formally accuses Chinese firms of systematically distilling US AI models, framing model theft as a national security issue and signalling potential new enforcement or retaliatory policy.
Anthropic's Claude Mythos, released only to a controlled consortium due to its autonomous cyberexploitation capabilities, has paradoxically energised China's domestic cybersecurity industry by establishing a new capability benchmark that Chinese firms are now racing to match.
Chinese AI firms are emerging as significant 'token exporters' — four of the top ten models by token consumption on OpenRouter over the past month were Chinese — indicating that software-layer AI influence is expanding globally even as hardware export controls bite.
Key Developments
DeepSeek V4 and Huawei Chip Pledge Expose Limits of US Hardware Controls
DeepSeek released its open-source V4 foundational model in two variants — a 1.6 trillion parameter V4-pro and a 284 billion parameter V4-flash — claiming performance competitive with leading closed-source US models from OpenAI and Google DeepMind. Simultaneously, Huawei publicly pledged full hardware support for DeepSeek's operations via its latest chip generation, a pairing that directly answers the central question US export control architects have been wrestling with: whether restricting NVIDIA access would cap Chinese frontier model development. The evidence from this release suggests the answer is increasingly no, or at least not at the pace policymakers had hoped. South China Morning Post
The Huawei dimension is strategically significant beyond the headline. Huawei's chips remain technically inferior to NVIDIA's H100 and H200 series by standard benchmarks, but DeepSeek's repeated demonstrations of extreme training efficiency — the core innovation that made V2 and V3 globally disruptive — mean that raw compute parity is no longer a prerequisite for frontier capability. The Huawei-DeepSeek axis is now a de facto sovereign AI compute stack: Chinese-designed chips running Chinese-developed models at Chinese-controlled infrastructure. This is the vertical integration the US export control regime was designed to prevent. DeepSeek's concurrent, deliberately small fundraising round — capped at 3% equity dilution — suggests the company is managing its independence carefully, resisting dilutive capital while retaining talent, and does not signal financial distress. South China Morning Post
US Export Control Escalation: Legislative Advance Without Enacted Force
The House Foreign Affairs Committee passed 20 new export control measures targeting China's access to advanced semiconductor manufacturing equipment, with the Match Act — which would bar Chinese chipmakers from accessing US-origin equipment — among the most significant. This package is described by its sponsors as the 'largest significant export control upgrade' aimed at China. It is critical to note the precise legal status: these measures have cleared committee stage only and must pass the full House, the Senate, and receive presidential signature before becoming enforceable law. Foreign policy professionals should not treat this as enacted policy. South China Morning Post
The second-order consequence question — whether controls are slowing China or accelerating domestic substitution — is being answered in real time by the DeepSeek V4 and Huawei developments published the same week. The historical pattern from prior semiconductor restrictions suggests that aggressive controls compress the timeline for domestic alternatives rather than foreclosing capability development. The Match Act's specific focus on manufacturing equipment, rather than finished chips, represents an attempt to attack Chinese chipmakers' production capacity at its foundation, but SMIC and Huawei's semiconductor arms have been operating under restrictions since 2020 and have continued advancing, albeit at a slower pace.
White House Frames AI Model Distillation as Strategic Theft
A White House memo attributed to Michael Kratsios formally characterises the practice of 'distillation' — training smaller models using outputs from larger proprietary US models — as wrongful appropriation by Chinese firms. This represents an important doctrinal shift: elevating what has been treated primarily as a commercial IP dispute into a national security framing. Distillation is a technically standard practice in AI development and is used by firms globally, but the memo's specific focus on Chinese actors signals that the administration is building a policy and potentially legal basis to restrict or penalise this method of capability transfer. BBC
The enforcement challenge is substantial. Distillation occurs through API access and inference calls — transactions that are commercially available and technically difficult to distinguish from legitimate usage at scale. Restricting Chinese access to US model APIs would require either geographic access controls (already partially in place via terms of service at OpenAI and Anthropic) or technical fingerprinting of model outputs to detect distillation-derived training data. Neither approach is watertight. The memo's significance is therefore more in establishing a narrative framework — Chinese AI advancement as theft rather than innovation — than in immediately actionable enforcement mechanisms.
Anthropic's Mythos: Controlled Release Creates Asymmetric Cybersecurity Dynamic
Anthropic's Claude Mythos Preview, released to a controlled consortium rather than publicly, has generated an unusual geopolitical dynamic. Its demonstrated ability to autonomously identify and exploit cybersecurity vulnerabilities at enterprise and financial system scale has made it simultaneously a Western offensive cyber asset and a benchmark that defines the new frontier of AI-enabled cyberattack capability. The controlled release means Chinese state and commercial actors do not have direct access, but the public documentation of its capabilities has been sufficient to energise China's domestic cybersecurity industry, which is now working to replicate or surpass the benchmark. South China Morning Post
This illustrates a recurring dynamic in dual-use AI: the announcement of a capability can be as strategically consequential as its deployment. China's cybersecurity sector now has a defined target capability to pursue, and the consortium model — which includes presumably allied nation actors — concentrates Mythos access within a specific geopolitical alignment. The question for strategic stability is whether this model of controlled release to allied consortia becomes the template for future frontier models with military or intelligence applications, effectively creating a two-tier global AI capability landscape aligned along alliance lines. South China Morning Post
China's 'Token Export' Strategy and Tencent's Open-Source Play
Chinese AI models — including DeepSeek, and now Tencent's newly released open-source Hy3 preview — are achieving significant global developer adoption through the software layer, with four of the top ten models by token consumption on OpenRouter being Chinese over the April measurement period. This 'token export' dynamic represents a strategic vector that existing US export controls cannot address: the models run on infrastructure outside US jurisdiction, are frequently open-source, and are accessed by developers globally including in US-allied nations. Tencent's Hy3 is led by former OpenAI researcher Yao Shunyu, illustrating continued talent flow that is itself a form of technology transfer — in reverse direction. South China Morning Post South China Morning Post
The strategic logic of open-source token exports maps directly onto China's historical approach to telecommunications infrastructure: offer capability at low or zero cost to establish ecosystem dependency, accumulate usage data, and position Chinese technical standards as the default. For Global South developers and enterprises that lack the budget for OpenAI or Anthropic API costs, freely available Chinese open-source models are not a compromise — they are the rational choice. This creates a pathway to AI infrastructure dependency on Chinese models that mirrors the Huawei 5G dynamic, and it is advancing without a clear US policy response targeting the software layer.
Signals & Trends
The Hardware Control Thesis Is Losing Empirical Support
US export control strategy since October 2022 has rested on the premise that restricting access to advanced chips — NVIDIA's high-end GPUs in particular — would impose a durable ceiling on Chinese AI development. The DeepSeek V4 release, combined with Huawei's chip support pledge, is the third major data point after V2 and V3 challenging this thesis. The mechanism is DeepSeek's training efficiency innovation, which appears to reduce the compute requirements for frontier-class models significantly enough to operate within the constrained hardware environment. If this efficiency trajectory continues, the strategic relevance of hardware controls diminishes as a ceiling mechanism, and the US needs a policy rethink toward software, data, and talent controls that it has been slower to develop. The pattern to track is whether V4's benchmarks hold up under independent evaluation — if they do, the policy implications for the entire export control architecture are significant.
Alliance-Line AI Access Stratification Is Accelerating
The Anthropic Mythos consortium model, the US export control escalation, and the White House's distillation memo collectively point toward an emerging world where frontier AI access is being stratified along geopolitical alliance lines — not just commercial terms. Mythos goes to a consortium of vetted actors. Advanced chips are denied to Chinese entities. US model APIs face potential China-specific restrictions. On the Chinese side, DeepSeek's open-source strategy and Huawei's chip support create a parallel stack available to non-aligned and sympathetic nations. The implication is that within two to three years, the global AI landscape may feature two distinct capability tiers: a US-aligned frontier with controlled access, and a Chinese-aligned open ecosystem with lower but rapidly improving capability. Nations that have not made a clear alignment choice — much of the Global South — will face increasing pressure to choose, with economic and technical dependency implications either way.
Talent as a Bidirectional Transfer Vector
Tencent's Hy3 model being led by former OpenAI researcher Yao Shunyu is a concrete instance of a pattern that US technology policy has addressed inadequately. Export controls restrict chips and equipment, and emerging policy targets model distillation, but the movement of researchers trained in US frontier labs to Chinese AI companies transfers tacit knowledge — training methodologies, architectural intuitions, research culture — that cannot be captured in a technology control list. The US has historically relied on immigration openness as a talent retention mechanism, but rising visa friction and the deliberate recruitment efforts of well-capitalized Chinese firms are creating effective outflow channels. The intelligence and policy question is whether this talent flow is being tracked systematically and whether it will generate a political response targeting researcher mobility in ways that would significantly alter the global AI talent market.
Explore Other Categories
Read detailed analysis in other strategic domains