Physical AI is next – Australia can be the proving ground

The AI boom has so far been measured in models, chips and data centres. The next phase will have arms, legs and embodied intelligence.

China’s Unitree Robotics provides one of the clearest tests yet of how much investors will pay for that future. The Hangzhou-based company has priced its Shanghai IPO at a valuation (Opens in new window) of about US$9 billion, raising around US$900 million. But the extraordinary number is its price-to-earnings ratio: 219 times its 2025 earnings. The retail offering was more than 8,000 times (Opens in new window) oversubscribed.

Unitree is profitable and growing rapidly. Humanoids have yet to be deployed economically at the scale implied by its valuation. Investors are therefore not principally buying current cash flows but an option on “Physical AI” embodied in these robots becoming the next great computing platform.

That is why Chinese AI company DeepSeek’s US$20.8 million (Opens in new window) strategic investment in Unitree is as interesting as the IPO valuation. The intention is to combine DeepSeek’s model capabilities with Unitree’s expertise in robotics, motion control and embodied intelligence.

It represents the convergence of two technological trajectories: increasingly capable AI “brains” and affordable robotic “bodies”. Connecting AI to sensors, actuators and motors and it gains the capacity to manipulate the physical world, meaning robots will increasingly not simply optimise manufacturing, logistics or mining processes but perform parts of them autonomously.

China has emerged as the global leader in humanoid robotics manufacturing, benefiting from its access to hardware and engineering know-how at scale. More than 140 Chinese companies (Opens in new window) are now developing humanoid robots, having launched over 330 models in 2025. In the first half of 2026, Chinese firms shipped more than 97% of humanoids globally (Opens in new window). The industry is increasingly dominated by two national champions: Unitree Robotics and AgiBot, which together now command roughly three-quarters of global shipments (Opens in new window).

America has different strengths: frontier AI models, computing platforms and deep pools of capital. Physical AI looks, for the next several years, like an asymmetric two-lane race. China leads in manufacturing scale, hardware and supply chains; America in frontier models, computing and capital.

Each is racing towards the other’s strength.

But other countries do not need to compete across the entire technology stack. Japan has strengths in robotics and special components, Germany in industrial automation, and Singapore in logistics and port automation.

Australia needs to find its own position.

Its comparative advantage may lie in specialised technologies, systems integration, testing and safety assurance.

Mining is an obvious candidate.

Australian mines are large, remote, hazardous and expensive to staff. They are also relatively controlled environments where the economic value of autonomy can be measured. Australia already has extensive experience (Opens in new window) deploying autonomous mining equipment in such conditions.

That makes the resources sector not merely a customer for robotics but potentially a global test bed for Physical AI.

This matters because deployment itself generates technological capability. Putting increasingly intelligent machines underground generates operational data and exposes problems that cannot be discovered in laboratories. That knowledge feeds back into hardware, software and AI models.

This would allow Australia to position itself between technology development and mass deployment, providing demanding real-world environments in which Physical AI systems are tested, validated, adapted and improved.

The same logic could extend to agriculture, logistics, infrastructure inspection, construction, healthcare and aged care.

For a middle power, this offers an alternative to technological self-sufficiency. Australia does not need to own the entire stack to capture value from it. But it does need to ensure that deployment creates knowledge and capability that remain in Australia, while securing a special position in the global value chain.

There is another reason Australia could occupy this niche.

Recent incidents involving frontier AI agents – for instance, models attempting to evade shutdown or copy themselves during safety testing have shown frontier models taking unauthorised actions (Opens in new window) during cybersecurity evaluations. It is tempting to describe this as AI “going rogue”. A more accurate interpretation may be human hubris.

The deeper problem is that AI capabilities can advance faster than the safety and governance infrastructure surrounding them.

Physical AI raises the stakes. A software agent crossing a digital boundary can compromise digital systems. An autonomous machine operating in a mine, hospital, warehouse or home can cause real-world harm.

Safety cannot be treated simply as a brake on innovation. For autonomous systems, safety itself becomes a technological capability required for deployment.

Australia has an opportunity here. Its experience regulating mining, healthcare, aviation, workplace safety and autonomous operations could support specialised capabilities in testing, certification, human-machine interaction and safety assurance.

Government does not need to decide how every robot should work. But it can establish performance-based standards, liability rules, testing infrastructure and regulatory sandboxes for safe experimentation and deployment. These capabilities could serve not only the domestic market but also become exportable services and expertise.

Rather than simply subsidising domestic production, government can help create the testing infrastructure, standards, regulatory institutions and industry-research connections through which deployment generates capability.

As AI acquires a body, Australia should aim to become one of the places where robots learn how to work in the real world.