Main stage

When should debt enter the capital stack?
William Godfrey (Tangible) · Renji John (Eternal Ag) + more

Equity Alone Built Software. What Funds Physical AI?

Physical AI is brutally capital-intensive, yet founders often fund the asset-heavy part with the most expensive money they'll ever raise. We put a founder, a VC, and a private-credit lender on stage to show how the most capital efficient companies layer debt and equity from day one.

2026 sees growing capital demands for the numerous Physical AI companies scaling. Fleets, factories, deployments that equity isn't designed to fund. Meanwhile, private credit sits on dry powder hunting for exactly this kind of asset-backed yield.

The debate isn't whether to use debt; it's how early you have to think about it. Most of the industry assumes you can bolt debt on later. But the decisions that make a company financeable have to be in place long before the raise, and delaying has high costs. The real tension is between the comfortable default of delay and the discipline of engineering for debt from day one.

If you don't scan the world, how can you model the world?
Chen Feng (NYU, AI4CE Lab)

Chen Feng is an Institute Associate Professor at New York University, Director of the AI4CE Lab, and Founding Co-Director of the NYU Center for Robotics and Embodied Intelligence. He has also been an Amazon Scholar with its Frontier AI & Robotics team since 2026. His research focuses on active and collaborative robot perception and robot learning to address multidisciplinary, use-inspired challenges in construction, manufacturing, and transportation. He is dedicated to developing novel algorithms and systems that enable intelligent agents to understand and interact with dynamic, unstructured environments.

Before NYU, he worked as a research scientist in the Computer Vision Group at Mitsubishi Electric Research Laboratories (MERL) in Cambridge, Massachusetts, where he developed patented algorithms for localization, mapping, and 3D deep learning in autonomous vehicles and robotics. Chen Feng earned his doctoral and master's degrees from the University of Michigan between 2010 and 2015, and his bachelor's degree in 2010 from Wuhan University.

Chen is an active contributor to the AI and robotics communities, such as CVPR, IEEE RA-L, and ICRA, and he has served as an area chair and associate editor. In 2023, he was awarded the NSF CAREER Award. More information about his research can be found at ai4ce.github.io.

Marc Pollefeys (ETH Zurich · Microsoft Spatial AI)

Marc Pollefeys is a full professor in the Dept. of Computer Science of ETH Zurich since 2007 where he leads the Computer Vision and Geometry lab. He is also the director of the Microsoft Spatial AI Lab in Zurich, heading a team of scientists working on spatial perception algorithms for AI assistants and robotics.

He was previously associated with the Dept. of Computer Science of the University of North Carolina at Chapel Hill where he started as an assistant professor in 2002 and became an associate professor in 2005. Before this he was a postdoctoral researcher at the Katholieke Universiteit Leuven in Belgium, where he also received his M.S. and Ph.D. degrees in 1994 and 1999, respectively. His main area of research is computer vision, but he is also active in robotics, machine learning and computer graphics. One of his main research goals is to develop flexible approaches to capture visual representations of real world objects, scenes and events.

Dr. Pollefeys has received several prizes for his research, including a Marr prize, an NSF CAREER award, a Packard Fellowship and a European Research Council Starting Grant. He is the author or co-author of more than 300 peer-reviewed publications. He was the General Chair of ICCV 2019 and ECCV 2014 and Program Co-Chair for CVPR 2009.

Talks

Why can't physical AI exist without eyes?
Ash Cleary (LDV Capital)

Physical AI will not succeed until machines can truly see. To build the next generation of Physical AI – robots that build, heal, protect, and explore – we'll need a revolution not just in computation but in sensing technologies.

The next trillion-dollar companies will be born at the intersection of visual sensing, physical data, haptics and intelligent autonomy. That's why forward-looking capital is already flowing into the technologies that make machines aware: multispectral sensors, vision-first robotics, real-time 3D capture, and synthetic training data. These are not niche subcategories – they are the backbone of the embodied AI revolution.

Humanoid robotics development faces real tradeoffs - adding sensors increases cost; increasing flexibility reduces durability; more motors mean more complexity and energy consumption. This mechanical challenge – coupled with the perceptual and control loops – is why general-purpose humanoids are being held back from becoming a revolution.

Is AI magic or ordinary technology?
Serge Belongie

Is AI magical or ordinary technology? Who is arguing each position, and what can history teach us? This question impacts how technology leaders can most effectively put AI to work for them, viz., as a tool for augmentation or as a replacement for a human workforce.

The eye-watering valuations of tech companies ride on the premise that AI is magic.

Investors, builders, C-suite decision makers, concerned members of the public.

Testimonials of instances when people put AI to work for them and said "Wow, AI is magic!" or "Wow, AI is not magic!"

Roundtables

What does '90% success' mean on the floor?
Mads Paulin (Aitera Robotics) · Mehrdad Farimani (Merphi Consultants)

The roundtable interrogates the gap between the success rates embodied AI companies advertise (often 90%+ in curated demos) and what industrial deployment actually requires — where even small per-step failure rates compound across multi-step tasks and fall far short of the 99%+ reliability, safety certification, and ROI discipline that 30 years of automation has established as the bar.

The debate asks whether embodied AI companies understand what it actually takes to earn a spot on the factory floor, and what they'd need to do differently — on reliability, business model, and deployment readiness — to become a genuinely attractive automation solution rather than a "silver bullet" pitch. 2026 is the year embodied AI moves from pilots to real procurement decisions, backed by billions in capital. Sell demo-day success rates as production-ready, and the industry risks repeating GM's "lights-out factory" mistakes of the 1980s or the reliability frustrations of early day cobots resulting in stalled lines and years of industry skepticism. This debate sets an honest bar for "deployment-ready" before more capital and credibility are spent finding out the hard way.

Embodied AI providers' definition of success compounds to well under 60% on real multi-step tasks — far short of the 99%+ reliability, safety certification, and ROI discipline industry has built over 30 years. Vendors bet rapid model improvement closes that gap; skeptics see GM's "lights-out factory" and early cobots repeating: capital and confidence outrunning what the technology can actually deliver on a live floor. The open question isn't whether it works in a demo, but whether the AI industry understands what "good enough" means to the person who has to sign off on it.

Is it a terrible idea for robotics companies in Europe not to go full-stack?
Jonathan Karl (Angel Invest) · Francesco Ricciuti (Visionaries Tomorrow)

We want to explore where durable value will accrue in robotics as the stack matures. The session will examine whether the winners will be verticalized, data-rich deployment players or full-stack platform companies that absorb more of the value chain over time.

Our co-hosts have slightly different views. One is that robotics startups can build strong businesses around site-specific data, deployment know-how, and productized integration, especially in mission-critical or highly contextual environments. The opposing view is that this advantage is temporary, and that hardware platforms, especially from China, will keep moving up the stack until software and integration players are squeezed into low-margin roles.

Investors, robotics founders, industrial operators, and researchers thinking about commercialization and deployment. It should be especially useful for anyone trying to understand where defensibility really sits in the robotics value chain.

The session is designed as a moderated roundtable rather than a presentation. We'll open with a clear disagreement between the two of us, use that tension to frame the discussion, and invite participants to challenge both sides with examples from their own companies, research, and deployment experience.

Can you underwrite a robot?
Giulio Brugnaro · Manas Gosavi (Production Capital)

Capital doesn't flow to what it can't trust, and nobody yet agrees how to measure, verify, or insure what a robot does. This roundtable explores the missing trust infrastructure that turns physical AI into a financeable asset class.

Physical AI is moving from pilots to real deployments, and the bottleneck is shifting from what robots can do to whether operators, lenders, and insurers can price their risk. Decisions made today around telemetry, standards, reporting and guarantees will determine which robots scale commercially and which remain venture-funded experiments.

Capital demands performance history. Performance history requires deployment at scale. Deployment at scale requires capital. Breaking that cycle requires more than capable robots. It requires agreement on who bears the loss when they fail, and what infrastructure would make someone confident enough to own that risk.

This roundtable aims to bring together the people shaping how robotics scales: investors, founders, insurers, operators, and others building the data, standards, and risk models the industry still lacks.

What's Europe's unique wedge to win with Physical AI?
Mala Valroy (Thursday VC) · Vasileios Balntas (Staer)

The easy answer treats "Physical AI" as one problem with one obvious path: build a VLA, feed it a massive sim-to-real pipeline, scale. But that bundles three separate bets into one: Sim-to-real is a data/training methodology: sim is just one way to get data, teleop and on-robot RL are others. VLA is an architecture hypothesis about mapping perception and language to action: you can do sim-to-real without a VLA, and train a VLA on nothing but real data. Physical AI is the outcome, agnostic to both.

The unspoken equation is usually: Physical AI = big VLA + massive sim-to-real pipeline. That's a current research bet, not a definition. And, it's precisely the strategy that rewards whoever has the most compute and the most deployed robots.

This roundtable will pull the bundle apart: where sim-to-real actually breaks (perception, dynamics, or contact), whether VLAs are learning physics or just pattern-matching until they aren't, and what's left to bet on once you stop assuming compute and fleet size are the only paths to Physical AI. Finally, it discusses whether there is a sub-vertical only Europe could win — and what would €1B of European capital need to stop funding to go after it?

Europe: too few workers, too much robot fear?
Hampus Jakobsson (Pale Blue Dot) · Skander Garroum (PROTOTYPE)

China ships 90% of the world's humanoids and calls them a demographic fix, Europe calls them a job threat while 570,000 German industrial jobs sit unfilled, and the US calls them a race. This session pulls together deployment data & stories, unit economics, and the current (conflicting) evidence from the US, Germany, and China to ask what actually happens to labor, demand, wages, and the countries still waiting to industrialize.

At €48.30 an hour for German manufacturing labor, Europe has some of the fastest robot payback math anywhere and the strongest union veto over using it. Whether European industry deploys, delays, or watches China ship 90% of the machines is a capital allocation question being answered in the next years and we will try to answer it.

The empirical record currently splits by geography: industrial robots cut employment and wages in the US and China, while Germany lost no net jobs because unions and a growing service sector absorbed the shock, at the price of stagnant wages and a falling labor share. Whether general-purpose humanoids repeat the German pattern or produce something without precedent, with knowledge-work AI landing in the same decade is the main question.

Operators and investors first, the people actually allocating capital to this in the next years. Founders building humanoids and policy people second, because the debate will decide what they can sell and how it gets regulated.

The session opens and closes with the same vote: will humanoids destroy more European jobs than they create by 2035? And then we collect data/anecdotes from the audience and guide them across our talking points.