From Tokyo: in a country where the semiconductor supply chain is treated as a matter of national survival, the latest move by Chinese AI lab DeepSeek reads less like a corporate decision and more like a geopolitical statement. According to sources cited by Reuters, Reuters reported that DeepSeek withheld pre-release access to its upcoming V4 model from both Nvidia and AMD, the two American chip giants that have long underpinned the global AI industry. In their place, domestic Chinese suppliers including Huawei Technologies received an early preview, granting them several weeks to optimise their processors for the new model before the broader market sees it.
AI developers typically share pre-release versions of major models with leading chipmakers such as Nvidia and AMD to ensure their software performs efficiently on widely used hardware. DeepSeek has previously worked closely with Nvidia's technical staff. The abrupt change in approach has raised eyebrows across the semiconductor and AI industries alike.
When an AI lab builds a new model, it typically hands pre-release versions to major chipmakers weeks before launch. Nvidia, AMD, and others use that window to tune their drivers, CUDA kernels, and inference stacks so the model runs well on their hardware from day one. It is routine, unglamorous plumbing work — and it matters enormously for adoption. Instead, Huawei's Ascend chip division and other domestic Chinese suppliers received the V4 preview, giving them weeks to optimise their silicon for what is expected to be a model with a one-million-token context window: architectural features that demand tight hardware-software integration to run efficiently.
Complicating the picture further is a parallel allegation from within the Trump administration itself. The development comes as a senior Trump administration official told Reuters that DeepSeek's latest AI model was trained on Nvidia's most advanced chip, Blackwell, using a cluster in mainland China, in a move that appears to violate US export controls. DeepSeek may seek to remove technical indicators revealing its use of American AI chips, and plans to publicly claim that it used Huawei's chips to train its model, according to the US official. Neither DeepSeek nor Huawei responded to requests for comment, and Reuters was unable to immediately confirm the reasons behind the access decision.
What the Lockout Actually Means
The commercial impact on Nvidia and AMD may be less severe than the optics suggest. Ben Bajarin, CEO of research firm Creative Strategies, told Reuters that "the impact to Nvidia and AMD for general data accelerators is minimal — most enterprises are not running DeepSeek, which serves as a benchmarking model more than anything else." Bajarin added that new AI coding tools are already cutting optimisation time from months to weeks, narrowing the window of any advantage granted to Huawei.
Most enterprise AI deployments outside China still run on Nvidia hardware, and open-source models tend to get community-driven optimisations for CUDA within days of release regardless of whether the lab cooperates. Nvidia's ecosystem of 4.7 million developers, extensive tooling, and battle-tested inference libraries will not collapse because one lab denied pre-release access. The deeper significance is symbolic and strategic, not merely technical.
The move is likely part of a broader strategy by the Chinese government "to try to keep US hardware and models disadvantaged" in China, according to Bajarin. By prioritising Huawei's Ascend chips for optimisation, DeepSeek is accelerating the development of a parallel software ecosystem that reduces long-term dependency on US technology. That longer arc, rather than any single model release, is what Washington's strategists are watching most closely.
The View from the Region
What Australian observers often miss about DeepSeek is how quickly it has reshaped global perceptions of Chinese AI capability without yet winning over Western markets. Chinese open-source AI models have overtaken every other country in downloads on Hugging Face, with DeepSeek alone surpassing 75 million downloads since January 2025. That surge has been concentrated in the developing world, where access to large-scale cloud infrastructure remains limited, rather than in OECD economies where security concerns have tempered enthusiasm.
Australia's posture toward DeepSeek has been unambiguous. DeepSeek has been banned from all federal government systems and devices in Australia, following warnings that the data it collects could be available to the Chinese government. Australian Home Affairs Minister Tony Burke said the government "will not hesitate to act when our agencies identify a national security risk." Federal corporations including NBN Co, the ABC, and Australia Post have banned DeepSeek from their internal systems despite being exempt from the commonwealth ban.
The concern is not simply about one chatbot application. As the Australian Strategic Policy Institute has noted, the deeper risk lies in DeepSeek's model being embedded in third-party products and services that touch critical infrastructure, from energy grids to industrial control systems, where the stakes of any compromise are far higher than a leaked email. For technology leaders in Australia, the strategic implications extend beyond whether DeepSeek itself should be deployed. Analysis from ACI Research argues that Australia has much to learn from China's emphasis on energy-efficient, low-power AI architectures, particularly as Australia's data centre electricity consumption is projected to triple, reaching around six per cent of national grid demand by 2030.
A Race That Cannot Be Paused
The legitimate counterargument to a purely hawkish reading of events is worth taking seriously. Analysts who study Chinese technology development point out that DeepSeek spun out of a commercial hedge fund, not a state-owned enterprise, and that the lab's engineers have demonstrated genuine technical ingenuity in building competitive models under constrained chip access. DeepSeek claimed to use fewer chips and spend much less money to train and develop its models, creating a global perception that the barriers to entry for developing advanced AI are lower than previously thought. For middle powers, including Australia, that perception opens possibilities as well as risks.
Australia is a major supplier of the critical minerals integral to the compute used to train and run AI models. Some speculated that DeepSeek's computational efficiency gains would depress demand for those minerals, but Jevon's paradox suggests the opposite: as AI becomes cheaper, its use becomes more widespread, which raises overall demand. The United States Studies Centre at the University of Sydney has argued that Australia should invest in capturing a bigger stake in the global AI supply chain regardless of which lab wins the model benchmarks.
The V4 episode does not resolve the tension between those competing imperatives; it sharpens it. Canberra must hold two things in mind simultaneously: the genuine security concerns that justify its existing restrictions on DeepSeek, and the risk that a reflexively closed posture forfeits insight into innovations that could inform Australia's own sovereign AI ambitions. China's focus on efficient, low-power AI architectures may offer strategic lessons for Australia's sovereign AI ambitions, particularly given local energy and environmental constraints. That is not an argument for lowering the guard; it is an argument for thinking more precisely about what the guard is protecting and why.
The Department of Home Affairs has framed Australia's approach as risk-based and technology-agnostic, which is the right instinct. The harder work, as the DeepSeek story keeps demonstrating, is applying that framework with enough sophistication to distinguish between a chatbot's data risks and the broader structural contest for AI's next hardware and software standards. Reasonable people can disagree about precisely where those lines fall. The mistake would be to stop drawing them at all.