Update, 20 October 2026: a third robot-AI round followed: Eka Robotics closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime, TechCrunch reported. Read our piece.
G42, the Abu Dhabi technology group, and its portfolio company Analog have signed memoranda of understanding with Advanced Machine Intelligence, the world-model company co-founded by Yann LeCun and known as AMI Labs, G42 announced on 20 October 2026. The release calls them non-binding and names no amounts.
The pacts landed in the same two days as two physical-AI funding rounds: Neolix’s $400 million Series E for driverless delivery vans and Dream Labs’ $90 million seed for a robot foundation model. By our arithmetic, the two rounds total $490 million. The other commitments of those two days, from Germany, Nvidia and Hyundai, came without a price.
Business Pill · PHYSICAL AI
A short explainer on physical AI: what changes when AI gets a body. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the only priced items in these two days were two startup rounds, while the agreements involving G42, the German government and Nvidia name no amounts yet.
G42’s release says “The non-binding agreements span frontier AI research, advanced compute infrastructure, sovereign enabled cloud and potential industrial applications.” It describes AMI Labs as developing world models, AI systems “designed to understand physical environments, anticipate outcomes and support reasoning and planning.”
The release says the agreements “provide a framework to exchange information, define shared objectives and shape future cooperation,” and that “Scope, commercial terms and implementation will be determined in separate definitive agreements.” They were signed on the sidelines of the UAE–France Business Dialogue in Abu Dhabi, according to the release.
AMI Labs co-founder and chief executive Alexandre Lebrun is quoted in the release: “World models are how AI will learn to understand and act in the physical world.” G42 says Analog’s founder and chief executive is Alex Kipman, whom it calls the inventor behind Microsoft’s HoloLens and Kinect.
Physical AI on 19 and 20 October 2026: Neolix’s $400 million Series E (Bloomberg) and Dream Labs’ $90 million seed (Business Wire) are the only priced items, $490 million together by our arithmetic. G42 and AMI Labs signed non-binding MoUs and Germany signed a declaration of intent with Nvidia, with no amounts; Hyundai’s 2032 humanoid date is a plan as Kyodo reported it; ARK’s ~$26 trillion is a forecast. Sources: G42 release via Zawya, Bloomberg, Business Wire, Nvidia blog, Kyodo, ARK Invest.
What Was Posted on X
What if AI could experience the world, not just describe it
That’s the frontier we’re exploring with Yann LeCun’s @ylecun Advanced Machine Intelligence @amilabs
G42 and our portfolio company @Analogio_ have signed MoUs with AMI Labs to explore the next generation of AI through world models – systems designed to understand physical environments, anticipate outcomes, and support reasoning and planning.
Signed on the sidelines of UAE-France Business Dualogue in Abu Dhabi, the agreements mark another step in the growing AI collaboration between the UAE and France.
Because the next frontier of AI isn’t just about generating intelligence. It’s about understanding the world we live in.
G42 posted on X at 09:01 UTC on 20 October 2026 that it and its portfolio company Analog have signed MoUs with AMI Labs to explore world models, which it describes as systems designed to understand physical environments, anticipate outcomes and support reasoning and planning. The post, like G42’s release, gives no amounts; the release adds that the agreements are non-binding.
Yann LeCun co-authored a paper that fixes a flaw in AI world models with a rule from the 1830s.
World models are AIs that learn how the world moves by watching video.
Robots can use them to plan by predicting what happens next.
They break down on long predictions.
Each guess becomes the input for the next guess. If every step shrinks the model’s picture of the world a tiny bit, the shrinking stacks up. After enough steps, different situations start to look the same to the model.
Researchers from NYU, Brown and AMI Labs fixed this with a rule physicists have used for nearly 200 years.
In physics, a moving system without friction can stretch and twist, but the total space its possible states fill never shrinks.
The team built that rule into the model’s prediction step. The shrinking is now mathematically impossible at every step, whatever the model learned.
They call it LeHami. They tested it on simulated video of a swinging pendulum against LeWorldModel, a leading model of this kind trained straight from pixels.
> Its speed predictions were 2.8x more accurate over the first 10 steps. > After about 11 steps, it tracked the pendulum’s position better too. > Its picture of the world kept exactly the same size, while the baseline’s shrank at every step. > It matched the baseline’s best accuracy with one-eighth of the training compute. > It has 32% fewer parameters. > They trained it on 10 frames per second, and it predicted just as well at 20 and 40. > When they ran it backward, it returned to its starting point with an error under 0.0001%.
It’s not a clean win. In goal-reaching tests, a baseline that looks at more past frames hit the target more often. And every test here is a pendulum, not a robot in a kitchen.
The team got this result by adding one rule from physics. The model came out smaller than the baseline and cheaper to train.
Should more AI borrow rules from physics instead of scaling up
@alex_verem summarised the LeHami paper on X, describing it as a fix for world models whose picture of the world shrinks over long predictions, and listing its pendulum results along with a caveat that a baseline did better in goal-reaching tests. The figures in this update come from the paper itself.
The Money That Moved
Neolix, a Beijing-based maker of driverless delivery vehicles, raised $400 million in a Series E led by Boyu Capital, Bloomberg reported on 20 October 2026, as we covered that day. The company said its total raised is now more than $1.4 billion, according to Bloomberg, without disclosing a valuation.
Dream Labs, founded by former leads of Nvidia’s GR00T robotics research, announced a $90 million seed round on 19 October 2026, led by B Capital along with Kindred Ventures, according to its Business Wire release. Its X post gives its Dream-1 robot model a starting price of $3 an hour.
The Commitments Without a Price
In Berlin on 19 October, Nvidia’s blog says, “Chancellor Merz, the Federal Ministry for Economic Affairs and Energy and the Federal Ministry for Research, Technology and Space signed a Joint Declaration of Intent with NVIDIA.” The same post says “Huang announced new planned investments in Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace.” It gives no amount for any of them.
Hyundai Motor Group is stepping up its push into robotics and physical AI, Kyodo reported on 20 October 2026. Kyodo says the group aims to develop more advanced autonomous manufacturing systems at its innovation centre in Singapore by 2030, before introducing humanoid robots and wider automation at future factories from 2032.
Kyodo also reported that Hyundai Motor’s shares, which reached a record closing high of 755,000 won in June, closed at 312,000 won on Friday. It quoted iM Securities analyst Lee Sang-soo: “Shares of Hyundai Motor have fallen sharply amid weakness in its core automotive business.”
The Forecast
ARK Invest’s 2031 Teslamodel, published on 19 October 2026, says humanoid robots “could scale to a ~$26 trillion total addressable market, roughly twice that of the robotaxis’ $10-15 trillion.” ARK writes that in its current model, external humanoid robot sales scale to about 1 million in 2031. That is a forecast, not money committed.
As we read it, the commitments sit at different layers: G42 offers compute, cloud and deployment to a world-model lab, Germany offers a policy framework, and Neolix and Dream Labs raised money for products already on sale.
As we read it, Hyundai’s 2032 date and ARK’s 2031 model are the two figures that look furthest ahead, and both are plans or forecasts rather than spending.
As we read it, the G42 release ties the AMI Labs MoUs to its existing presence in France, its Grenoble data centre and its partnership with Mistral AI, which it says was formalised in May 2025.
Alexandre Lebrun, AMI Labs, in G42’s release
“World models are how AI will learn to understand and act in the physical world”
Three Implications
TWO ROUNDS, $490 MILLION Neolix’s $400 million and Dream Labs’ $90 million are the only priced items, by our arithmetic.
TERMS STILL TO COME G42 says scope, commercial terms and implementation will be set in separate definitive agreements.
DATES, NOT DOLLARS Hyundai’s 2030 and 2032 targets, as Kyodo reported them, and ARK’s 2031 model are plans and forecasts.
The framework lists, among four moats: “The Container Moat — built by closed data loops inside a customer’s environment, nascent but the deepest position in the stack.”
As we read it, the deals here are attempts to get inside such environments: factories for Hyundai, delivery fleets for Neolix, and industrial applications for G42 and AMI Labs. The framework was published in July 2026 and does not discuss any of these companies.
Update, 20 October: The World-Model Research Behind AMI Labs
AMI Labs, the startup in the G42 MoUs, is also listed on a research paper posted to arXiv on 11 October, “LeHami: Symplectic Latent Dynamics for World Models”. Its authors are Romina Yalovetzky of New York University, Randall Balestriero of Brown University and Yann LeCun, listed with New York University and AMI Labs. The paper replaces the prediction step of a JEPA world model with a Hamiltonian flow that, in its words, “preserves phase-space volume at every rollout step”, so the model’s picture of the world cannot shrink over long predictions.
On simulated video of a pendulum, the paper reports that LeHami “leads by 2.8× over the first ten steps” on velocity prediction against LeWorldModel, matches that baseline “with at most one-eighth of the training compute”, and has “about 32% fewer parameters (3.03M against 4.44M)”. It does not win everywhere: in the paper’s goal-reaching tests, its success rate on the position goal was 0.480, against 0.800 for LeWorldModel.
As we read it, this is a research result on a pendulum, not a product. We read the paper itself after a summary by @alex_verem on X pointed to it.
What Is Not Established
G42’s release names no investment, compute volume or timeline for AMI Labs, and says terms will be set in separate definitive agreements. Nvidia’s blog gives no amounts for its four planned German investments, and Hyundai’s 2030 and 2032 dates are plans as Kyodo reported them. ARK’s figures are its own model. The $490 million total is our arithmetic and counts only the two rounds.
Business Engineer Framework
The Four Intelligence Moats
A Business Engineer framework on where intelligence accumulates in each AI paradigm, and who can capture it.
Over two days, physical AI drew two priced rounds totalling $490 million by our arithmetic, plus a set of agreements, a declaration and a factory roadmap that name no amounts. The G42 and AMI Labs MoUs are the newest of these, and G42’s own release says their scope and terms are still to be agreed.
A note on sourcing. This piece is built on G42’s release of 20 October 2026 (via Zawya), Kyodo’s report on Hyundai, Nvidia’s 19 October blog and ARK Invest’s 19 October model, each read in full, plus our own earlier pieces on Neolix (Bloomberg) and Dream Labs (Business Wire). The $490 million total is our arithmetic and counts only the two funding rounds. The 20 October update rests on the LeHami paper on arXiv (2610.15267), read in full. Nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
