Two events in one week drew the line
On 24 April, US prosecutors charged three people connected to AI server maker Super Micro Computer with conspiring to divert billions of dollars worth of GPU-equipped servers to China in violation of export controls. The FBI alleged the defendants fabricated documents and used shell companies to conceal the operation.
The same day, the US State Department sent a cable to every diplomatic and consular post worldwide warning that Chinese AI companies are stealing American technology through “distillation,” the practice of training smaller models on outputs from larger ones. The cable named DeepSeek, Moonshot AI, and MiniMax.
These are not abstract geopolitical manoeuvres. They are supply-chain signals. For any New Zealand business that has adopted, or is considering, Chinese-developed AI models, the compliance landscape just shifted underneath them.
The distillation problem is bigger than most realise
The State Department cable followed concrete allegations. In February, Anthropic accused DeepSeek, Moonshot and MiniMax of running large-scale extraction campaigns against its Claude model, involving more than 16 million exchanges via roughly 24,000 fraudulent accounts. DeepSeek alone accounted for over 150,000 exchanges. MiniMax ran more than 13 million.
Anthropic warned that models trained this way may strip out safety guardrails designed to prevent misuse in areas like biological weapons development. The company does not offer commercial access in China for national security reasons.
Meanwhile, DeepSeek withheld its flagship V4 model from US chipmakers Nvidia and AMD in February, giving Chinese chipmakers including Huawei a head start. A senior Trump administration official told Reuters the model appeared to have been trained on Nvidia’s Blackwell chip inside mainland China, potentially violating export controls. DeepSeek’s models have been downloaded more than 75 million times on Hugging Face since January 2025.
The appeal is obvious. Chinese developers have closed a five-year AI gap in under eighteen months through algorithmic efficiency. For a cost-conscious Kiwi SME, a free open-source model that benchmarks near GPT-4 looks like a gift. It might be a liability.
Reputational risk is outpacing legal risk
Here is the uncomfortable nuance. Legal experts cited by Mission Media Asia note that AI distillation, while potentially violating terms of service, does not constitute clear IP theft under existing law. There is no international framework to enforce these claims across borders.
But the market is not waiting for courts. AI vendor associations have become the fastest-growing risk disclosure among large public companies, with 38% of S&P 500 firms now flagging AI reputational risk in filings, up from 12% in 2023. For any NZ exporter or firm with US-aligned clients, having DeepSeek in your tech stack is becoming the kind of thing procurement teams ask about.
Chris Newton-Smith, Chief Executive of cybersecurity firm IO, put it plainly in January: “Resilience absolutely depends on an organisation’s ability to integrate cyber security, information security, privacy, and AI governance into a single risk picture.”
Wellington’s silence is not neutrality
New Zealand’s AI governance framework is not built for this. The country was the last OECD nation to publish an AI strategy, releasing it in July 2025 with a focus on enabling adoption. The accompanying responsible AI guidance is voluntary. Enforcement relies on the Privacy Act and Fair Trading Act, neither of which contemplated machine-learning supply chains.
Dr Athar Imtiaz, Applied AI researcher at Massey University, warned in March that the Privacy Act “does not define technical standards for model validation, dataset integrity or audit requirements. Without tailored legislation, accountability becomes fragmented.” In the same piece, Mark Easton, CEO of Nodero, noted that NZ ranked around 40th on the 2025 Oxford Insights Government AI Readiness Index, behind Australia and comparable economies.
Regionally, Taiwan has banned DeepSeek across government agencies. Singapore welcomed it as innovation while simultaneously arresting chip smugglers. New Zealand has taken no public position.
The lock-in problem arrives next
The immediate vendor risk is bad enough. The structural risk is worse. Gartner forecast in January that 35% of countries will be locked into region-specific AI platforms by 2027, up from 5% today. Gaurav Gupta, VP Analyst at Gartner, says decision-makers are “prioritising AI platforms that align with local values, regulatory frameworks, and user expectations over those with the largest training datasets.”
Gartner’s advice for CIOs is to build model-agnostic workflows that allow switching between vendors. For a Kiwi SME that has wired DeepSeek into its customer service pipeline or document processing, that flexibility does not exist yet.
The government’s own strategy estimates generative AI could contribute $76 billion to the NZ economy by 2038. But 68% of NZ SMEs currently have no plans to evaluate or invest in AI at all. When they finally move, they will reach for the cheapest tool available. Right now, that tool may come with compliance baggage that nobody in Wellington is helping them understand.
Sources
- RNZ: US charges 3 people tied to Super Micro Computer with conspiring to divert AI tech to China (2026-04-24)
- Mission Media Asia: US-China AI Distillation – APAC Leaders Face Hard Choices (2026-04-24)
- NZ City News: Anthropic accuses Chinese AI labs of illicit Claude copying (2026-02-26)
- Craig McGregor: The gap is narrowing dangerously between China and Silicon Valley on AI (2026-03-24)
- eCommerce News: NZ faces legal and sovereignty risks as EU AI rules take effect (2026-03-17)
- SecurityBrief: AI, geopolitics & supply chains reshape cyber risk (2026-01-13)
- IT Brief: Gartner warns of rising lock-in to regional AI stacks (2026-01-30)
- NZ Government: AI Strategy – Investing with Confidence (2025-07)
- NZ Government: Responsible AI Guidance for Businesses (2025-07)