The forge of intelligence: exploring the rise of physical AI

Image generated by AI

As with any investment, your capital is at risk.

Robots are not a new idea. In fact, they feature in some of the world’s oldest literature. Homer’s Iliad (8th century BC) describes the workshop of Hephaestus, the Greek god of fire and metallurgy. In this workshop, Hephaestus makes weapons for other gods and for heroes like Achilles.

Some of Homer’s conjured images are familiar to the imaginations of modern readers and relatively mundane, such as automatic gates for the heavens. Others are more exciting: automatic bellows for Hephaestus’ forge and wheeled tripods that deliver ambrosia to the gods.

Appropriately for the current moment, Homer even depicts the first robots with artificial intelligence: Hephasteus’s handmaids “all cast in gold but a match for living, breathing girls. Intelligence fills their hearts [and] voice”.

These visions of ancient automata are inextricably linked to the idea of saving labour and making it more productive. Hephaestus needed help in his forge because he had a physical disability. It would’ve been impossible for him to do the work the other gods required of him without technology.

Today, as the world faces profound labour shortages, we suspect that demand for labour-saving machines will increase. We believe a secular step-change in demand presents exciting investment opportunities for us to back modern companies in their pursuit of the ancient goal of embodied intelligence.

From cognitive to physical intelligence

Since the launch of ChatGPT in 2022, investors have been preoccupied with AI’s cognitive power: reasoning, writing, coding, and problem-solving. Most market analysis has focused on its impact on office and knowledge workers who, until now, have been largely insulated from automation. Yet this narrow focus risks missing a more profound transformation: the impact of intelligent machines on the physical world.

The rapid progression of automative capabilities could reshape industries that still rely heavily on manual labour. And with labour accounting for more than half of global GDP, the potential economic and social consequences are vast.

In a recent discussion with academics at the Massachusetts Institute of Technology, we explored how far-reaching the consequences might be. Advances in generative models may enable machines not only to think but also to perceive and act. NVIDIA’s founder, Jensen Huang, captured the shift neatly when he said, “The ChatGPT moment for general robotics is just around the corner.”

SymBot moves autonomously and dynamically in response to its environment

Flexible, general-purpose robots designed to perform multiple tasks have long been the holy grail for roboticists, and they are now moving closer to commercial reality. Some will take human form, others will not. But either way, they are beginning to exhibit the human-like abilities of understanding unstructured environments and adapting to change.

‘Humanoid’ remains the popular label, but I prefer the term physical AI. While much of the world has been designed by people for people, and it makes sense that robots will initially resemble the human form, this is an unnecessary distraction from the more crucial point.

Whether the robots resemble human beings or not, they will increasingly think and behave much more like humans. Physical AI is broader, more accurate, and less emotive. It encompasses all systems where intelligence meets motion, controlling machines in the real world — from industrial robots to drones and autonomous vehicles.

A brief history of false dawns

I should also disclose something else. While I am a long way from being a science-fiction nerd, I have long been fascinated by the potential for robots to eventually move beyond their protective cages on factory production lines and be able to perform more than a single pre-programmed task.

This interest in robotics dates back to my early career, investing in Japanese equities, where robots were already an integral part of manufacturing. Yet despite steady progress, the field has experienced repeated disappointments.

So, you have been warned. There is a real risk that I am just a stale bull on this topic. However, I believe recent advances are overcoming many of the historic hurdles.

Traditional industrial robots are powerful but rigid. They sit inside safety cages, performing pre-programmed motions in tightly controlled settings. Any deviation from their expected instruction can halt production. Over the decades, progress has unfolded through a series of gradual iterations: faster arms, better precision, but little flexibility.

Many companies have tried to break that pattern. Honda’s ASIMO drew headlines in the 2000s but struggled with balance and agility. Google, after acquiring Boston Dynamics in 2013, attempted to accelerate progress through massive data collection but hit a wall: there simply wasn’t enough relevant training data. SoftBank later tried to integrate several robotics acquisitions yet still fell short of commercial viability.

By 2020, most investors had written off general-purpose robotics as a dead end. Even OpenAI (the creator of ChatGPT) closed its robotics division. The task of teaching a machine to perceive, reason, and move through the real world seemed too complex to solve through narrow, isolated models.

The foundation model breakthrough

The arrival of foundation models has changed that. These large, general AI systems integrate multiple capabilities – perception, planning and control – within a single architecture. Machines can now learn from multimodal data: text, images, video, and sensor inputs. Just as language models can generate prose or code, these physical models can generate motion. They can interpret their environment, plan actions, and adapt to tasks they have never explicitly been trained for.

Equally important is the rise of large-scale simulation. Platforms such as NVIDIA’s Isaac Sim enable robots to perform millions of virtual trials before operating in the real world. This addresses the historical shortage of training data and, when combined with generative learning techniques, dramatically accelerates progress.

Meanwhile, hardware has also improved across the board. Depth cameras and tactile sensors are increasingly affordable and yet more capable. Force sensors can now detect resistance and adjust grip, making collaboration with humans safer. Actuators have shifted from hydraulic to electric power, again lowering costs and improving reliability.

And, after decades of fragmentation, the industry is finally moving towards standardisation, helped by open-source models and government-led coordination. As I will go on to explore, this is particularly true in China, resulting in a step-change in both capability and economics.

The cost of advanced robots has fallen from around $250,000 in 2022 to roughly $100,000 in 2024, with credible projections towards $20,000 in the foreseeable future. When a robot that never sleeps approaches the cost of a human worker, adoption becomes inevitable.

The emerging adoption curve

We are still in the early stages of the rollout of physical AI, but the momentum is clear. The first wave of applications is most likely going to be in semi-structured environments such as warehouses and factories, where flexibility is valuable but conditions are still predictable.

Gradually, as robots become more advanced, we should expect to see them out in the real world. Hazardous materials disposal, nuclear maintenance, disaster recovery and mining seem the obvious first industrial use cases before robots appear in hospitals and care homes.

Market projections vary, but the direction is unmistakable. The global robotics market, valued at $100bn today, is growing by more than 10 per cent annually. If the United States alone were to fill its projected shortfall of 2.1 million manufacturing workers by 2030 with embodied AI at a cost of roughly $30,000 per unit, that would represent a $60bn market — roughly two-thirds of the entire sector today. Extending the logic globally could result in an eventual addressable market of $1tn or more.

The global contest: brain versus body

And while I write about a burgeoning robotics ecosystem, we may be facing a bifurcation in the road, with two distinct markets emerging. Because, as with other strategic technologies, physical AI is becoming a geopolitical race.

China has natural advantages when it comes to the robotic body. It has the world’s largest manufacturing base, with strong expertise in batteries, electric vehicles and automation, where there is a reasonable degree of overlap in the components required in robotics.

Government policy has also channelled capital and talent into the field, from Made in China 2025 to a 2023 directive from the Ministry of Industry and Information Technology that names humanoid robots a key growth driver.

Lastly, the Belt and Road Initiative (BRI), China’s state-backed infrastructure and trade programme launched in 2013, has positioned China to dominate many of the crucial raw materials needed.

China’s advantage within the robotic brain, however, has been less clear. Despite being home to more than half of the world’s AI engineers, China has achieved much of its progress in AI using leading-edge hardware provided by Western companies, such as NVIDIA, and has leveraged open-source AI models, including those from OpenAI and Meta.

This outsourcing is a key vulnerability for China’s robotic ambitions. But recent progress by companies like DeepSeek could be a signal that China is inching closer to a fully independent robot ecosystem.

We recently spent time in China meeting with several generative AI foundation model companies. Two things stood out as increasing their odds in leading globally on robotics. First, a focus on multimodal models that builds on their existing strength in video. Second, a clear steer from the government to adopt an open-source strategy to increase compatibility across the hardware ecosystem.

China’s approach is familiar: scale quickly, drive costs down, and export aggressively once domestic dominance is established. This mirrors the country’s success in other targeted technologies, such as drones, where DJI captured the global market by undercutting Western competitors.

There are now more than 60 active humanoid companies in China, led by firms such as Unitree, which already sells a humanoid robot for about $16,000. The winners will emerge battle-hardened, ready to sell globally at competitive prices.

The United States, by contrast, leads in software and semiconductors — the brain of physical AI. Companies like NVIDIA, Google, and Tesla are key players in this ecosystem. But the US lacks a strong domestic supply chain for key mechanical components, having spent decades prioritising capital-light innovation over production. Only about 10 per cent of American factories have a single robot, placing the country tenth globally in robot density.

Yet, there is a clear and ever-expanding need for more advanced automation in the US, particularly if the dream of onshoring manufacturing is ever to come to fruition.

Productivity in US manufacturing is heading in the wrong direction, with approximately 2,000 manufacturing workers retiring every day, often without passing on their lifetime of knowledge and expertise.

The National Security Commission on AI in the US has acknowledged the risk of falling behind China in the field of robotics. However, the central government has yet to develop a long-term strategic plan to support the industry, as policies shift from one administration to the next.

Were a US company able to develop proper vertical integration in robotics, owning more of the supply chain, from components to assembly, the prize could be huge, given the lack of domestic competition and likely apprehension toward importing Chinese “humanoid” robots. But the current domestic supply chain presents extremely challenging hurdles for current leaders like Tesla or Figure AI.

The outcome may be two distinct ecosystems: Chinese embodied AI systems dominating hardware and deployment, and US firms leading in intelligence and chips. Given current geopolitical tensions, cross-border integration looks unlikely. The strategic question is whether China can develop self-sufficient “brains” before the US rebuilds its “body.” The first to succeed will hold a critical advantage.

Investment opportunities

For the Long Term Global Growth portfolio, this emerging landscape already touches several holdings.

On the brain side, NVIDIA is central. Its chips, simulation tools, and foundation models make it indispensable. In China, Horizon Robotics’ twin-chip and software platform positions it well for domestic opportunities. As ever, TSMC remains well-positioned to serve both the US and Chinese AI ecosystems.

On the body side, opportunities are forming further up the value chain. Battery leader CATL has begun developing specialised power systems for robots, reflecting how electric-vehicle know-how is migrating into embodied AI.

In the US, Symbotic’s warehouse automation platform shows what near-term deployment looks like. Amazon has also now deployed over a million robots across its network, the largest industrial mobile-robot fleet on earth, while Meituan and Tencent are experimenting with service robots.

As for potential new ideas, component specialists often capture sustained value in automation cycles. In Japan, companies such as Harmonic Drive and Nabtesco have historically built durable franchises this way.

In the current shift, sensors appear poised to take on that role as they already represent around 40 per cent of Tesla’s Optimus robot cost base. Chinese firms such as Sanhua, Leaderdrive and Shenzhen Inovance, which I am familiar with from my time living in Shanghai, could follow the same path as earlier Japanese leaders.

The lesson from past industrial shifts is that the highest sustained returns often accrue not to those assembling finished products, but to those supplying critical inputs and enabling technologies. Physical AI may prove similar.

Remaining bottlenecks

Despite recent progress, several obstacles remain. Robots still lack fine motor dexterity, and their ability to operate safely in highly unstructured environments is unproven. Even with simulation, data diversity remains a limiting factor.

Beyond these technical issues, substantial ethical and governance challenges also exist. Questions of liability, privacy, and bias must be addressed before large-scale deployment. The potential for military use adds further complexity. These issues will require new regulatory frameworks and sustained international coordination.

From an investment perspective, the capital intensity of hardware manufacturing means the industry’s economics will differ from those of software. Company margins may compress quickly once competition scales. Patience and selectivity will be essential.

None of this diminishes the scale of what is unfolding. Generative models, simulation and hardware innovation are converging to bring intelligent machines into the real world. We are likely five to ten years away from a tipping point — when robots can learn from each other, collectively build experience, and operate with a degree of autonomy that transforms both production and services.

If the first industrial revolution extended human muscle, and the digital revolution extended our minds, the rise of physical AI will merge the two. The implications for labour, productivity and geopolitics are profound.

The craftsman’s legacy

In the Iliad, Hephaestus’s workshop is more than myth; it is a meditation on invention. His automata blur the line between the living and the made, suggesting that intelligence, once embodied, can reshape what is possible. Three thousand years later, we are approaching that threshold in earnest.

The modern forge is no longer fuelled by fire but by data and code. The artisans are engineers and model-builders, not gods. Yet the ambition is recognisable: to extend human capability, to fill the gaps left by scarcity, and to create tools that can work alongside us.

As investors, we should see in Hephaestus’s forge not just myth but precedent. Moments when technology changes what labour means have always created both disruption and opportunity. The challenge now, as then, is to recognise which creations will endure and to back the builders shaping this new age of intelligent craft.

Risk factors and important information

The views expressed should not be considered as advice or a recommendation to buy, sell or hold a particular investment. They reflect opinion and should not be taken as statements of fact nor should any reliance be placed on them when making investment decisions.

This communication was produced and approved in January 2026 and has not been updated subsequently. It represents views held at the time of writing and may not reflect current thinking.

Potential for Profit and Loss

All investment strategies have the potential for profit and loss, your or your clients’ capital may be at risk. Past performance is not a guide to future returns.

This communication contains information on investments which does not constitute independent research. Accordingly, it is not subject to the protections afforded to independent research, but is classified as advertising under Art 68 of the Financial Services Act (‘FinSA’) and Baillie Gifford and its staff may have dealt in the investments concerned.

All information is sourced from Baillie Gifford & Co and is current unless otherwise stated.

The images used in this communication are for illustrative purposes only.

Important Information

Baillie Gifford & Co and Baillie Gifford & Co Limited are authorised and regulated by the Financial Conduct Authority (FCA). Baillie Gifford & Co Limited is an Authorised Corporate Director of OEICs.

Baillie Gifford Overseas Limited provides investment management and advisory services to non-UK Professional/Institutional clients only. Baillie Gifford Overseas Limited is wholly owned by Baillie Gifford & Co. Baillie Gifford & Co and Baillie Gifford Overseas Limited are authorised and regulated by the FCA in the UK.

Persons resident or domiciled outside the UK should consult with their professional advisers as to whether they require any governmental or other consents in order to enable them to invest, and with their tax advisers for advice relevant to their own particular circumstances.

Financial Intermediaries

This communication is suitable for use of financial intermediaries. Financial intermediaries are solely responsible for any further distribution and Baillie Gifford takes no responsibility for the reliance on this document by any other person who did not receive this document directly from Baillie Gifford.

10059558