Where is AI unlocking advantages? Explore what is possible at AMA: Software 2026.
Walking into Protein Studios in Shoreditch for Sifted Summit 2026, the name brought to mind one of the defining artificial intelligence breakthroughs of the pre-ChatGPT era.
In November 2020, Google DeepMind announced that AlphaFold had achieved what the organisers of the CASP protein-structure prediction challenge described as a solution to a decades-old scientific problem. Five years after AlphaFold2 was published, its impact is difficult to dispute. The AlphaFold database now provides more than 200 million predicted protein structures, and has been used by millions of researchers around the world.
AlphaFold also shows what AI does not change – AI dramatically reduced a scientific bottleneck. It did not make biology instant.
Knowing, or accurately predicting, the structure of a protein does not remove the need to understand its behaviour in different environments, interactions with other molecules, toxicity, chemistry, manufacturability, experiments, animal studies or clinical trials. Research published in 2026 continues to combine AlphaFold predictions with experimental measurements, while other work highlights limitations in predicting the dynamic ensembles that proteins adopt in the real world. A recent Nature report found that adding more than 20,000 proprietary protein structures held by pharmaceutical companies could improve AI protein models beyond systems trained only on public data.
AlphaFold removed one constraint and exposed the ones behind it. Solving a bottleneck on paper (or silicon) does not remove the physical world.
The point kept resurfacing at the European start-up and investor Summit, where much of the discussion moved beyond chatbots and large language models towards what is increasingly described as “physical AI”: systems expected not merely to generate information, but to perceive, predict and act in the real world.
From producing answers to producing actions
Nick Clegg, the former Meta executive and UK deputy prime minister, now a general partner at Hiro Capital, questioned whether simply extending today’s transformer-based model paradigm indefinitely would be sufficient.
Clegg described generative AI as an oddly contradictory technology: extremely versatile, but also “monstrously inefficient”, requiring extraordinary amounts of capital, energy and data. That concentration, he argued, favours companies able to deploy infrastructure on a scale that Europe is unlikely to match.
But he also described transformer-based AI as “ripe for disruption”, with one particularly important unresolved area being “the interaction between AI and the physical world.”
Clegg’s argument isn’t that large language models suddenly become irrelevant. Instead, he questioned whether scaling the same architectures inevitably produces machines able to understand and navigate dynamic, unpredictable environments.
“I have my real doubts whether that really is true,” he said of the proposition that transformer scaling ultimately delivers a comprehensive understanding of the physical world. Existing generative AI can already help robots perform repetitive tasks, he added, but navigating the physical world in its “wholly unpredictable glory” is another matter.
A later Sifted panel on world models made a similar distinction from the engineers’ perspective.
Rosalyn Moran, CEO of Stanhope AI, described LLMs as fundamentally designed around prediction in language. Her company is pursuing a different type of reasoning system intended to predict, decide and act locally, without continually returning to a server farm. “The edge of physical AI is just a different realm,” she said.
Sarah Gates, VP of Global Affairs at autonomous-driving company Wayve, offered perhaps the more useful taxonomy. “I think the dichotomies should really be about cognitive AI versus physical AI rather than focusing on one particular model type,” she said.
Wayve itself uses several approaches, including language and vision-language-action models. Gates described its in-vehicle system as closer to a “world action model”. More importantly, she warned against assuming that Europe (or anyone else) should bet on winning one particular architecture. AI research will continue changing.
The real transition is from producing answers to producing actions, whatever the architecture underneath. More succinctly, from producing answers to producing actions.
Reality is a harsher benchmark
Operating in the world leaves far less room for error. A language model can hallucinate a paragraph and remain useful. A programmer can reject broken code. A user can ignore bad advice. The consequences become different when an AI system controls a vehicle, robot, medical system or industrial process.
Gates said autonomous-driving software has to be integrated into physical vehicles and meet automotive-grade verification and validation requirements. “It’s a whole different ball game when it comes to deployment.”
Zelda Mariet, co-founder of biological AI company Bioptimus, approached the distinction from science.
With an LLM, she argued, evaluation can involve determining whether an answer is good, useful or plausible. Biological predictions can eventually face much more concrete tests. If a model predicts that a gene is expressed at a specific location, for example, “you can decide whether or not it’s correct”.
Later she put the contrast more colourfully. “I come from the world of training LLMs,” Mariet said, where evaluation could be “very loosey-goosey, for better or for worse”.
“With physical AI, all of a sudden you have this bar… Is it hard? Yes. But on the other hand, as a scientist, someone’s telling you what they want to see. That’s fantastic.” For physical AI, the world itself becomes part of the benchmark. Additive manufacturing knows this problem well.
An AI system might propose a design, process parameter, or build strategy, but that output eventually encounters powder, heat, gravity, and tolerances. A model cannot safely “hallucinate” a material property, laser parameter, unsupported geometry, CT interpretation, or qualification result indefinitely. Eventually, the answer becomes a part.
The UK Government is shifting its attention beyond the software models towards hardware and real-world adoption, explained Kanishka Narayan, the UK’s Minister for Artificial Intelligence. From his perspective, physical AI takes on another meaning: “More and more we need to have a collective conversation about what the geography of build is, and whether that makes sense for all of us globally as well.” He added that where countries beyond the obvious leaders in AI (USA and China) can win is in their viability as trusted partners, highlighting the techno-democracies of Europe and India.
The internet gave LLMs their data. What trains a robot?
Physical AI creates another problem: where does the training data come from? The rise of generative AI was partly enabled by the existence of vast quantities of digitised human output. Text, images, software code and video had already accumulated online at enormous scale.
Many of the datasets required for physical intelligence do not exist in equivalent form.
Nazo Moosa, managing director at Paladin Capital Group, gave the example of sound. Humans recognise events such as glass breaking in another room. A physical machine may need the same ability. But there is no conveniently labelled internet-scale dataset containing every sound a robot could encounter in every environment. Generating that data physically is expensive. That makes simulation platforms increasingly valuable, Moosa argued.
Biology has a similar constraint.
Bioptimus works with high-resolution images of biopsies and molecular measurements. Mariet said access to hospitals and laboratories with sophisticated measurement equipment is therefore fundamental to the company’s ability to train its models.
“We have to be in locations where this is even possible to build a technology before we can even sell it,” she said. That suggests another difference between the two AI eras. The internet gave generative AI its training material. Physical AI may have to manufacture its own.
Manufacturing AI may eventually require some combination of build chamber metrology, failed builds, maintenance records, material batches, inspection data, and post-processing outcomes.
For additive manufacturing, the question is who owns the manufacturing training set. The machine manufacturer? The MES provider? The service bureau? The materials company? The metrology supplier? Or the aerospace and medical companies running the parts?
Those datasets may become far more valuable as AI moves closer to production.
The factory as AI infrastructure
China is already treating real-world environments as part of the infrastructure required to train physical AI.
Moosa pointed during the panel to large training environments available to Chinese robotics companies. That claim is supported by a broader national push. In June, China’s Ministry of Industry and Information Technology and the state-assets regulator launched a programme explicitly calling for real-world training spaces across manufacturing, logistics, maintenance, healthcare and other settings, with the aim of accumulating high-quality physical-machine data. Beijing has already developed dedicated humanoid-robot training facilities recreating industrial workshops, shops, homes and healthcare environments. China’s most recent five-year-plan gives additional insight into the direction of travel.
In August, I visited Huafast’s Shenzhen 3D printing farm, where 5,600 Bambu Lab machines form part of a wider production network of around 15,000 desktop printers. What was striking was not simply the number of printers. It was everything surrounding them: labour, materials, maintenance, utilisation, order routing and the problem of keeping thousands of relatively inexpensive machines productive.
One printer I saw had logged approximately 7,800 printing hours. Now imagine the information generated not by one printer, but by thousands operating continuously through different materials, geometries, failures, maintenance interventions and production jobs.
The valuable dataset is what happens across millions of machine-hours when additive manufacturing becomes production. It is what happens across millions of machine-hours when additive manufacturing becomes production. That is not to suggest Huafast is currently training such a model. The point is that manufacturing scale itself creates an informational asset.
Factories may become part of the infrastructure on which physical AI is trained.
The bottleneck moves again
This also changes where value may accrue in AI.
Zeynep Yavuz-Willson of General Catalyst argued during an earlier Sifted panel that attention has focused heavily on the technology itself when the larger economic challenge is diffusion. “This is not AI as a technological challenge,” she said, “but this is AI as a change management challenge.” Physical AI takes that problem further.
A successful industrial AI company may need algorithms, hardware, specialised datasets, simulation, manufacturing partnerships, certification, safety validation, distribution and customers willing to let the system interact with real processes. That may create opportunities outside Silicon Valley. Industrial ecosystems, laboratories, automotive supply chains, hospitals and manufacturing expertise matter more when AI has to interact with matter rather than pixels.
But those same requirements also make deployment harder. AI may now be approaching the same transition more broadly. Its first great advantage was that software could move faster than reality. Its next challenge is that reality gets the final vote.
Register now for AMA: Software 2026. Join 3D Printing Industry on 22 October for expert presentations, panel discussions and live audience Q&A exploring the software shaping additive manufacturing. Register here.
To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on LinkedIn.
Explore the full Future of 3D Printing and Executive Survey series from 3D Printing Industry, featuring perspectives from CEOs, engineers, and industry leaders on the industrialization of additive manufacturing, 3D printing industry trends 2026, qualification, supply chains, and additive manufacturing industry analysis.
