Can you start by giving us an introduction to Nfinite and digital twin technology, and explain how it connects to retail?
Nfinite is what we call a physical data platform. Our mission is to transform incomplete 2D information into digital twins—realistic, high-quality 3D data. From there, we enable unlimited visual experiences: images, videos, animations, augmented and virtual reality, even in-store applications. The company was founded seven years ago, after four years of R&D to develop the capability to build these high-fidelity 3D models. Today, this has become the largest platform of its kind to convert 2D content into structured 3D representations.
We launched our platform in 2021 with a focus on retail. By converting product catalogs into digital twins, we could power e-commerce experiences at scale. This started with online product pages but quickly expanded into media campaigns, retail media, in-store experiences, and even printed catalogs. We work with both 3D representations of products and 3D environments, such as homes, warehouses, and stores, offering retailers consistent, realistic visuals across multiple channels.
Can you explain how high-fidelity 3D environments are used to train AI models, and how this shift has impacted your business?
When we founded Nfinite, the vision was clear: digitize the world in 3D.
At the time, much of the visual content online was produced by photographers or videographers. It seemed inevitable that technology would replace this process—both in creating and personalizing visual experiences. Retail was a natural entry point because e-commerce relies heavily on product visuals as the main link between consumer and product.
As AI evolved, especially from 2022 onwards, we saw large visual models emerging alongside language models. But these models depended almost entirely on human-generated data. That struck us as inefficient. Our belief was that synthetic 3D data—data generated from high-quality digital twins—could train models with more accuracy, especially when spatial understanding was required. By 2024, this became more widely recognized as data scarcity grew, and synthetic data became essential for scaling AI training.
How does 3D synthetic data help advance AI, especially for robotics?
Synthetic data can come in many forms, but our focus is on 3D synthetic data generated from digital twins. This allows us to produce unlimited visuals to train models in areas like computer vision, image and video generation, and even robotics. Unlike purely AI-generated synthetic data, which can degrade itself over time, 3D-based synthetic data is grounded in spatial accuracy and physical realism.
This shift became even more significant when industry leaders like Jensen Huang emphasized the importance of "Physical AI"—intelligence that interacts with space. Training a robot requires an understanding of dimensions, geometry, and spatial relationships, something 2D data cannot provide. By generating high-quality 3D datasets, we’ve improved models’ detection accuracy tenfold and enabled more advanced simulation environments for robotics and autonomous systems.
How far behind is spatial AI compared to language or image-based AI, and when will it reach mainstream adoption?
AI has evolved in clear stages: text models, then image and video models, and now Physical AI that interacts with space. Each step demands exponentially more data. While text data is abundant, image and video data is less accessible but still available in massive quantities online. For Physical AI, however, the required 3D high-quality data simply doesn’t exist at scale. This lack of data is the biggest bottleneck today.
That said, the scientific community increasingly agrees that even for image and video models, spatial accuracy is key to consistency—meaning 3D data is becoming essential across the board. As for adoption, language models are already embedded in daily life, and image/video models are rapidly entering creative workflows. Physical AI is still in its infancy, with fewer than 50 companies worldwide working on it. One prominent AI scientist described its stage today as equivalent to where language models were in 2016. Mainstream adoption is likely several years away, but the trajectory is clear.
How are major retailers using your digital twin technology to manage thousands of SKUs and improve shopper experiences?
Retailers today are pursuing five key strategic objectives: building omnichannel shopper experiences, advancing retail automation, deploying AI merchandising, expanding retail media and data monetization, and streamlining supply chains. All of these objectives rely on digital twins. Without accurate 3D product data, it’s impossible to deliver consistent, customized, immersive content across multiple platforms.
For example, AI merchandising requires the ability to simulate in-store layouts, detect when items are out of stock, and optimize traffic flow in a store. Supply chain automation, meanwhile, depends on AI’s ability to recognize and handle products, which requires accurate 3D models for training. By generating digital twins of entire catalogs and environments, retailers can move faster toward these strategic goals and deliver seamless experiences across every channel.
Your work involves processing large amounts of 3D data. What infrastructure challenges do you face, especially in Europe and the U.S.?
There’s no question that the compute infrastructure required for the next generation of AI is far beyond what exists today. As Eric Schmidt recently pointed out, we may only have 1–10% of the necessary capacity to support these workloads. Robotics, in particular, is extremely compute-intensive, which is why investment in edge computing is accelerating, moving more processing onto devices rather than centralized data centers.
That said, for Nfinite, infrastructure hasn’t yet been a bottleneck. We don’t train foundational models ourselves, which are the most energy-intensive processes. Our compute needs, though significant, are manageable and benefiting from falling costs and rising performance. Access to the latest GPUs like H100 or H200 can sometimes be challenging, but overall, we’ve been able to deliver to clients without limitations. The real infrastructure crunch lies ahead as AI scales further, particularly in robotics.
Why are open standards so important in the field of Physical AI?
For me, this is a philosophical belief: innovation accelerates in open ecosystems. When technology is open, it democratizes access and allows large developer communities to experiment, iterate, and build on each other’s work. Closed systems, while sometimes commercially attractive, limit the pace and scale of innovation.
We’ve seen this play out globally. China, for example, releases many of its models openly, which means developers worldwide—even in the West—end up building on them. That level of openness accelerates adoption and progress. For Physical AI, where the challenges are complex and the need for collaboration is high, open standards are not just beneficial; they’re essential.
Finally, what is your perspective on the U.S.–China race in AI, and do you see risks in advancing too quickly?
From conversations with leaders across the AI field, it’s clear that both the U.S. and China view artificial intelligence as the next decisive factor in global power—economically and militarily. That’s why enormous investments are being made by governments and companies alike. The AI race is real, and the stakes are high, as artificial intelligence is seen as the next "superpower weapon" in this global competition.
At the same time, the pace of progress is dizzying. Just a few years ago, breakthroughs came every few years; now, new scientific papers appear weekly, overturning assumptions we thought were immovable. This makes it far riskier for entrepreneurs to place bets, as the landscape shifts so quickly. While I firmly believe AI will be a force for good, it must be guided by boundaries, supervision, and responsibility. Otherwise, as with any powerful tool, the potential for harm exists alongside the immense potential for progress.
