A Coordinated US Campaign: From White House Memo to Global Warnings
The latest US China AI clash has moved into a new, coordinated messaging phase. A White House memo by science adviser Michael Kratsios alleges that foreign entities “principally based in China” are running an “industrial-scale campaign” to distil US frontier AI systems. Officials say Chinese actors use tens of thousands of proxy accounts and jailbreaking techniques to systematically extract capabilities from American AI models, stripping out safeguards and exploiting US innovation. In parallel, the US State Department has circulated a diplomatic cable to embassies worldwide, instructing diplomats to warn governments about alleged AI IP theft by Chinese firms such as DeepSeek, Moonshot AI and MiniMax. This goes beyond earlier, generic accusations of technology theft by directly targeting AI models, methods and vendors, turning model security and AI export controls into core themes of Washington’s broader tech diplomacy push.
How Distillation Became an IP Battleground – and Why DeepSeek Stands Out
At the heart of the dispute is knowledge distillation, a legitimate machine learning technique where a smaller “student” model is trained to mimic a larger, more capable “teacher” model. US officials argue that Chinese labs are abusing this approach by using unauthorized access, proxy networks and prompt jailbreaking to harvest proprietary behaviors and safety-tuned responses from American systems, then embedding them in rival models. OpenAI and Anthropic executives have told US lawmakers they detected adversarial distillation activity originating from China on their platforms. DeepSeek’s 2025 R1 model, widely believed by experts to rely heavily on distillation, delivered near frontier performance at dramatically lower cost, instantly becoming a symbol of China’s AI ascent. The memo warns that such campaigns can generate models that benchmark competitively while omitting security protocols and neutrality safeguards that US developers painstakingly build into their own products.
DeepSeek V4 on Huawei Chips: A New Flashpoint for US China AI Tensions
DeepSeek has now escalated its technological bet with the preview release of its V4 large language model, a 1.6 trillion parameter system boasting a 1 million token context window. Crucially, V4 is the first major frontier DeepSeek AI model optimized for Huawei’s Ascend AI processors instead of Nvidia GPUs. Earlier, the company’s V3 reportedly relied on 2,048 Nvidia H800 chips and drew scrutiny over whether it bypassed US AI export controls via intermediaries. By shifting to Huawei AI chips, V4 sidesteps restricted US hardware and strengthens China’s domestic AI stack. DeepSeek offers two variants: V4-Pro and V4-Flash, with the latter a 284 billion parameter model. The company says V4 trails OpenAI and Anthropic’s latest systems by only a few months while outperforming other open-source rivals in coding and reasoning. For US policymakers, the pairing of alleged AI IP theft with Huawei-based training makes DeepSeek a focal point for both model and semiconductor concerns.
Export Controls, Cloud Access and a Fragmenting AI Supply Chain
Washington’s latest accusations are likely to feed directly into tougher AI export controls on chips, cloud and model access. US agencies have already tightened rules on advanced Nvidia hardware and are probing how Chinese companies maintain access through overseas resellers and intermediaries. The White House memo highlights “sophisticated access infrastructure” that spreads traffic across thousands of fraudulent accounts to evade platform restrictions. Future measures could target high-end cloud computing capacity, API access to US frontier models, and even secondary markets where Chinese firms source restricted components. For global AI supply chains, this raises the risk of bifurcation: a US-centric stack built on Nvidia and proprietary frontier models, and a China-centric stack anchored by Huawei, open-source models like DeepSeek V4 and local cloud providers. Companies in Southeast Asia that depend on both ecosystems may find integration harder as compliance, licensing and due diligence obligations multiply.
What This Means for Malaysia: Risks, Compliance and Staying Stack-Agnostic
For Malaysian startups, universities and enterprises, the DeepSeek controversy is more than a distant geopolitical story. Using Chinese AI models or Huawei-based infrastructure now carries heightened sanctions, compliance and reputational risks, particularly for organisations also partnering with US cloud or AI providers. Institutions integrating DeepSeek AI models or hosting workloads on Huawei AI chips should map all data flows, contractual links and funding sources, and screen counterparties against evolving US and UN lists. To remain flexible in a potential tech cold war, Malaysian players can adopt a stack-agnostic strategy: containerised deployments, model-agnostic orchestration layers and hybrid multi-cloud architectures that can switch between US and China providers as rules change. Transparent governance, clear AI safety policies and robust documentation of how models are sourced and secured will be essential to reassure regulators, investors and international partners on both sides of an increasingly fragmented AI ecosystem.
