How will artificial intelligence change the world of manufacturing? Sam Altman, CEO of OpenAI, has no doubt that the effect will be transformational [1]:
“If we have to make the first million humanoid robots the old-fashioned way, but then they can operate the entire supply chain—digging and refining minerals, driving trucks, running factories, etc.—to build more robots, which can build more chip fabrication facilities, data centers, etc, then the rate of progress will obviously be quite different.”
It’s difficult to know what to make of this vision. Taking it at face value, it seems to represent a profoundly unimaginative view of the future, in which there is a straight replacement of workers in factories by humanoid robots. Factory automation has developed hugely since my brief period as a production line worker in 1980, but this hasn’t occurred by a one-for-one replacement of people by robots.
Most people have seen pictures of modern car factories, with robot arms carrying out repeated operations like welding with great precision. But, as Tim Minshall explains in his excellent book on manufacturing [2], robots are just one example of the many devices that can carry out physical operations in an automated factory. If you are automating a chemical factory, you don’t do it by getting a humanoid robot to open the valves and stir the tanks. The most sophisticated factories that currently exist – the chip fabrication facilities that produce the GPUs that underpin AI, as well as the CPUs in our phones and computers – are almost entirely automated. In the fab, a silicon wafer goes through hundreds of complex process steps without being handled by a human – but the robots that move the wafers from tool to tool run on wheels, not legs.
So is Altman just saying that automation makes capital goods cheaper, and that leads to a self-reinforcing process of increasing productivity? That’s certainly true, but it’s a process that is neither new, nor having much to do with large language models or generative AI.
This isn’t to say that AI and machine learning aren’t important for manufacturing, or that this importance won’t grow in the future. There’s huge scope to optimise the operation of each individual tool in a manufacturing process – whether that’s a milling machine or an ion implanter – especially when coupled with rapid feedback from analytical instruments and quality control monitoring. Here a combination of machine learning and physics-based simulation allows the operation of each tool to be optimised both in terms of efficiency and outcome. Putting together simulations of each tool with a digital depiction of the overall process flow gets one to what is fashionably called a “digital twin”, a virtual model of the entire production process, which allows optimisation of manufacturing at the system level.
Large language models will have their place too; effective knowledge management is a crucial function for big organisations, difficult to get right. It’s important to understand where the knowledge has to come from, though. The knowledge you need to build an advanced factory isn’t available on the internet, so it’s not something that you can ask a commercial LLM for. It’s tacit knowledge, much of it carried around in the heads of the engineers and technologists who have built and run a factory before. Detailed process knowledge will be produced in the operations of the machines, and learning to capture this data and use it to inform AI and machine learning models will be important for the success of the manufacturing firms of the future.
What will a manufacturing firm look like in the age of AI? A good place to start would be to look at what’s probably the most important, and most technologically sophisticated, pure-play manufacturer in the world, TSMC [3]. Its highly automated fabs manufacture the world’s most powerful chips – the material base that underpins the entire AI industry, and its communications stress how much it is using AI in its operations as well.
Despite its high level of automation, TSMC still employs a lot of people – a global workforce of 84,000. But these aren’t unskilled production line operatives – more than 80% of them are graduates, more than half educated to Masters level.
The scale of the company is indicated by its $90 billion revenue – a million dollars per employee. This produces $37 billion net income. It’s.a hugely capital intensive company, spending more than $23 billion annually on new production facilities and equipment. The advanced tools that this money buys are of course themselves the products of highly sophisticated manufacturers. The Dutch firm ASML, that produces extreme UV lithography tools costing hundreds of millions of dollars each, is just the most famous of these suppliers – and ASML in turn relies on firms like Trumpf (for lasers) and Zeiss (for reflecting optics). This is the epitome of a high-tech supply chain.
Unsurprisingly, TSMC does a great deal of research and development to improve its processes and develop new ones. Its $6 billion annual R&D expenditure is greater than that of many medium size countries. This produces a lot of patents – but interestingly, TSMC choses to protect much process knowledge as trade secrets, and maintains a formal and systematic way of tracking and maintaining these, as well as rewarding the employees who develop them.
So, we have two contrasting visions of AI in manufacturing. In Sam Altman’s version, armies of humanoid robots, powered by advanced general intelligence, build supply chains to make even more robots, leading to runaway growth. But looking at where we’re at now, what we’re seeing is AI driving process and system optimisation in factories that build on current trends in automation. It’s the contrast between these different visions that is in part driving the different approaches to AI in China and the USA. We’ll see which one is closer to reality.
[1] From Altman’s June 2025 blogpost: “The gentle singularity“.
[2] Tim Minshall, “Your life is manufactured“.
[3] All figures here taken from TSMC’s 2024 Annual Report.