The next defensible layer in AI may not be the model, but the infrastructure that turns physical objects into reality-grade 3D training data.
General-purpose AI models are becoming less differentiated. For founders and investors, the implication is immediate: the next defensible AI businesses will not win by wrapping the same model. They will win by controlling scarce, domain-specific data and the infrastructure that turns it into reliable outputs.
Physical AI makes that shift impossible to ignore. Robots, simulations and world models cannot learn the physical world from 2D internet data alone. They need high-fidelity 3D understanding. The useful test is to identify what data a system needs, why it is difficult to produce and whether controlling it creates an advantage that compounds.
In this EUVC episode, we speak with Michael Brehm (General Partner and Founder at Redstone), Ben Scheidt (Partner at Redstone) and Franz Tschimben (Co-Founder of ALLSIDES).
One model cannot serve every physical problem
The race to build the largest general-purpose model is giving way to a harder question: what must an AI system know to perform reliably in a specific environment?
Michael argues that the answer changes with the use case. Coding, cybersecurity, robotics and industrial augmented reality place different demands on compute, hardware, labels and data. In his words, “one size fits all doesn’t work”.
That changes where founders should look for defensibility. A generic model may be rented. A scarce data source, specialised production system or deeply integrated workflow is much harder to reproduce.
Founder and investor test: Name the requirement that changes by vertical, then ask whether the company controls it. If the answer is only prompting, orchestration or access to the same foundation model as everyone else, the moat is probably thin.
Physical AI cannot learn the world from the internet
Ben frames the shift as a move from bits to atoms. AI has scaled logic, language and other internet-native tasks, but economic impact in the physical world requires systems that can perceive shape, material, distance, collision and movement.
A robot cannot safely manipulate an object because it has seen a flat product image. It needs a structured representation of the object and its physical properties.
“For these robots to work, they basically use 3D data as their native language.”
This is the distinction between more training data and the right training data. Scraped text, images and video were abundant enough to accelerate the first wave of generative AI.
Reality-grade 3D data is scarce and costly to create. The opportunity lies in closing that gap for robotics, simulation, world models and other systems that must act rather than only generate.
Michael identifies the control point directly: “The core bottleneck today is not more parameters, it’s better reality-grade data.”
Portfolio review question: What information does the application need to operate reliably that cannot be downloaded, scraped or cheaply synthesised? That unavailable input may be a better guide to value than the visible application layer.
Quality at scale turns infrastructure into a moat
ALLSIDES approaches this constraint as a production problem. Its automated scanners capture physical objects and generate relightable 3D digital twins with mesh and physically based rendering data.
Franz describes a process that reduces production from roughly two days of manual work to about five minutes per object in the system discussed in the episode.
The point is not speed alone. The system is designed to preserve the exact measurements, detailed geometry and physical properties needed for downstream applications.
This matters because quality and quantity usually trade off. A manually produced 3D asset may work for one campaign but cannot supply a large training set. A cheap reconstruction may scale but fail the reliability threshold for robotics or simulation.
Infrastructure becomes defensible when it raises both ceilings at once: more assets, produced faster, without surrendering the properties the downstream model needs.
ALLSIDES has worked with clients such as Meta, Amazon, Nike, adidas, Zalando, NUREG, GORE-TEX, La Sportiva and Inditex Group brands including Zara, Massimo Dutti and Bershka.
Franz also describes a deep integration with NVIDIA, which is shaping the 3D ecosystem through its tooling and SimReady standard.
The customer mix shows how the same data layer can serve different needs. AI labs need training data. Fashion and commerce companies need scalable digital assets. Robotics and simulation teams need representations that behave reliably in virtual environments.
A data factory can become a compounding platform
The immediate product is a scanner and data-production system. The larger ambition is a reusable data layer on which customers can train models, generate new assets and build applications.
“We’re now setting out to build the largest ever created dataset and data infrastructure.”
The bigger opportunity is the compounding sequence behind it. Hardware creates proprietary 3D data. As that dataset grows, it becomes reusable across customers, while deep integrations reveal which tools and vertical applications should sit above it. The result is a path from scanner to data platform to application layer.
Ben compares this position to supplying the shovels in a gold rush. Yet the valuable shovel is not generic infrastructure. It combines hardware, software, scientific knowledge and a dataset that compounds.
Michael extends the thesis further: once a company understands and controls a large data layer, it can add tools to manipulate, combine and deploy that data. It becomes more than a supplier of raw inputs.
Founder takeaway: Map the sequence before calling the business a platform. What proprietary asset does the first product create? How does each deployment improve it? Which adjacent workflow becomes possible only after the data reaches sufficient depth and scale?
European deep tech must be global from the beginning
ALLSIDES was built from South Tyrol, but Franz says the company was global from the outset. At the time of recording, it had close to 45 people across 16 nationalities, a New York subsidiary and dedicated attention to Asian markets.
His advice is to “think globally from the beginning”, but not to treat internationalisation as a remote hiring exercise. Founders need to spend time in the market, build the first local team around trusted people and protect the company’s culture through those early hires.
Michael adds a useful distinction: Europe’s disadvantage is not necessarily effort. US markets often move faster because contracts, decisions and transactions turn more quickly.
Taken together, Franz’s internationalisation experience and Michael’s point about faster US transaction speeds suggest a practical rule: build market proximity wherever the critical customer, talent or standards-setting ecosystem sits.
For a physical AI infrastructure company, technical development, data operations and commercial learning may require different geographic footprints.
Five questions to pressure-test a physical AI opportunity
Is the data genuinely scarce? Identify what the system needs that public or synthetic sources cannot reliably provide.
Does quality survive scale? Compare throughput and cost without losing the fidelity required by the downstream application.
Does each deployment deepen the moat? Look for compounding data, workflow knowledge, integrations or standards alignment.
Is there a credible adoption ladder? Give smaller customers a way to validate the data, then define how they graduate into higher-volume or embedded use.
Is the company close to the market shaping the category? Global ambition requires founder presence, trusted early hires and proximity to customers, talent and technical standards.
The next wave of AI will not be won by larger models alone. In physical AI, value moves towards the companies that can capture reality, structure it and make it usable at scale.
For founders, that means building around the hard input rather than the visible interface. For investors, it means tracing defensibility back to the bottleneck and testing whether every customer makes that advantage stronger.
Listen to the full episode to hear how Redstone and ALLSIDES assess the 3D data bottleneck, the platform opportunity and the work required to scale Physical AI infrastructure globally.
