VENDORiQ: Will China be the Key to Enabling Australian Sovereign AI?

Why it’s Important

When AI is Set Free, Digital Sovereignty Becomes Possible

The release of Alibaba’s Qwen3 model represents another compelling example of challenging the notion that a few large, well-funded US laboratories hold an insurmountable lead on generative AI – the so-called ‘AI moat’.

This concept gained wider attention following a leaked internal Google memo from May 2023, titled ‘We Have No Moat And Neither Does OpenAI’. The memo argued that the rapid advancements occurring within global open source AI communities posed a greater competitive threat than other large commercial labs. It contended that open source models were quickly approaching or matching the performance of proprietary models, often doing so with significantly less computational resources and at a fraction of the cost, thereby lowering the barrier to entry for cutting-edge AI development.

Qwen3 appears to embody several aspects of the ‘no moat’ argument. 

By releasing the Qwen3 models as open source under the permissive Apache 2.0 licence, Alibaba made these models immediately accessible to a global community of researchers, developers, and enterprises. This fosters rapid experimentation, fine-tuning, and deployment, potentially accelerating innovation outside the control of any single vendor. The availability on platforms like Hugging Face and GitHub further facilitates this diffusion.

While US technology export controls aim to limit China’s access to the most advanced semiconductor chips (like top-tier NVIDIA GPUs), Chinese firms have reportedly been innovating in algorithm design, data optimisation, and leveraging domestic or slightly less advanced hardware more efficiently. 

This drive for efficiency, coupled with substantial investment and a large talent pool, is enabling the development of increasingly powerful models, potentially at a lower per-parameter training cost than some Western counterparts. 

While global access to cutting-edge Chinese-made AI chips that outperform NVIDIA in terms of cost-performance (though they are less powerful, they are a lot cheaper) is currently limited, their existence and use in training large models domestically underscore a growing capability that could influence the global AI hardware and model landscape over time. 

The above conforms to IBRS predictions in early 2023, that no single group of US vendors will secure a lasting monopoly in AI.

Rethinking AI Vendor Risk

For Australian organisations, this shifting landscape necessitates a re-evaluation of AI sourcing strategies. An overreliance on models from a limited number of vendors, predominantly based in one country, introduces significant vendor risk, including potential exposure to geopolitical tensions, changes in export controls, or shifts in vendor priorities. 

The availability of high-performing, open source alternatives from diverse origins, such as Qwen3 from China and various European open models, broadens the options available for building robust and resilient AI capabilities. Looking to vendors beyond the traditional sphere is becoming a valid and arguably necessary activity for strategic technology planning.

The potential for models such as Qwen3 to serve as a basis for a local, sovereign AI platform is real. Its open source licence permits unrestricted use, modification, and distribution, addressing a fundamental requirement for sovereign capability – the ability to control and understand the core technology. The range of model sizes supports various deployment scenarios, from smaller models runnable on departmental or edge hardware, to larger variants requiring enterprise-grade infrastructure, potentially within national data centres. Support from widely used open-source tools simplifies deployment. Multilingual capabilities are also crucial for applications serving diverse populations.

Sovereign AI is Not Without Challenges

However, implementing Qwen3 or similar open models as a sovereign platform requires significant internal expertise in model deployment, fine-tuning, and ongoing management. 

Organisations must assess the actual performance of these models on their specific tasks, beyond headline benchmarks which can sometimes be subject to ‘benchmaxxing’ or not reflect real-world performance or specific national contexts (as suggested by some independent commentary). 

For example, reported benchmark figures for Qwen3 show strong results in certain areas (e.g., leading on AIME25 and Codeforces among open models) but can be less clear or competitive in others (e.g., the lower GPQA score potentially without the hybrid reasoning fully engaged). Organisations must conduct their own rigorous evaluations. 

Furthermore, while the models are open source, the availability of base models for the largest variants is often incomplete (Both DeepSeek and Grok fall into this category), which could impact fine-tuning efforts for highly specialised sovereign applications. 

Considerations around data privacy, security, and potential biases inherent in any model trained on vast datasets, regardless of origin, must also be thoroughly addressed.

In Summary, There Really is No Mote

In summary, Qwen3 is more than just a new model release; it is symptomatic of deeper trends in the AI landscape – the accelerating pace of open source development, the increasing capabilities of non-US AI powerhouses, and the potential for technological innovation to circumvent traditional control points like advanced chip access. 

These factors are contributing to a more diverse and competitive global AI ecosystem, presenting both opportunities and challenges for organisations seeking to build secure, performant, and strategically sound AI capabilities.

Who’s Impacted?

• Chief Information Officers (CIOs): Consider models outside of the major US vendors in strategic planning for AI adoption. This includes assessing vendor risk associated with single-origin dependencies, evaluating the cost-effectiveness of open source models versus commercial APIs, and planning the necessary infrastructure (compute, storage) for potential on-premises or private cloud deployments to support sovereign or data-sensitive workloads.
• AI Team Leads/Data Scientists: Need to technically evaluate non-US AI capabilities (reasoning, coding, multilingual support, tool use) against specific use cases. Its open source nature allows for hands-on experimentation, fine-tuning, and integration into existing AI workflows. Understanding the hybrid reasoning architecture and optimising its use will be key. They must also assess hardware requirements and explore deployment options using tools like Ollama and llama.cpp.
• Technology Policy Makers: Should analyse China’s (and other nation’s) open source AI releases and performance in the context of national AI strategies and the pursuit of sovereign AI capabilities. This involves considering how such models can support domestic innovation, reduce reliance on foreign technology, and inform policy regarding AI development, standards, ethics, and cross-border data flows. The emergence of competitive models from non-traditional sources also has implications for international technology policy and trade discussions.

Recommendations

• Evaluate AI Models Directly: Download and test open source AI products such as Qwen3 models, relevant to your organisation’s anticipated workloads. 
• Conduct Independent Benchmarking: Do not rely solely on vendor-provided benchmarks. Establish internal evaluation processes to assess Qwen3’s performance, accuracy, hallucination rates, and biases on datasets and tasks representative of your specific needs.
• Assess Deployment Options: For potential sovereign or data-sensitive applications, investigate the hardware requirements and technical feasibility of deploying a broader set of AI models, such as Qwen, on your own infrastructure or within trusted national cloud environments. Evaluate the practicalities of using supporting tools for local inference.
• Diversify AI Sourcing: Actively explore and evaluate AI models from a variety of vendors and origins, including open source options from non-US sources. Develop a multi-vendor AI strategy to mitigate risk and leverage diverse capabilities.
• Develop Internal Expertise: Build or acquire the technical skills necessary to work with open source models, including fine-tuning, deployment, and managing the associated infrastructure.
• Inform Policy: Technology policy makers should engage with technical experts to understand the implications of open source models like Qwen3 for national AI capability, supply chain resilience, and the regulatory environment. Consider how policies can support the development and secure use of diverse AI technologies.