A point of view on physical AI

Two ways to predict the world. Choosing is the strategy.

Machines predict the physical world by learning it from data (a world model) or computing it from equations (a physics engine). Today, brain work is augmented by AI. Tomorrow, body work will be. This briefing explains where each approach belongs and what to ask before you commit.

Read the series →

The numbers behind the shift

Builders are no longer experimenting at the margins. They are shipping.

US$1B

Raised by both Yann LeCun's new world-model lab and Fei-Fei Li's World Labs

The largest seed round on record for a European startup underscores the acceleration of funding for foundational AI.

Crunchbase, 2026

US$18.8B

Raised by robotics startups worldwide so far in 2026

Already ahead of 2025's full-year total of US$15 billion and past the previous 2021 peak, with half the year still to run.

Crunchbase, 2026

Downloads of NVIDIA's Cosmos world foundation models

NVIDIA, the chip company whose hardware trains most large AI models, reports adoption at scale.

NVIDIA, Jan 2026

What we believe

Three reasons this belongs on a CEO's radar now, not later.

CONVICTION I

Capital is already moving.

As outlined in the statistics above, investment capital is flowing to the first movers and core technology providers in this space. The underlying models are being adopted at scale, not merely trialled, and competitors and suppliers are placing their bets now. The cost of understanding this field is far lower than the cost of being surprised by it.

CONVICTION II

A demo is not a product.

The costliest error in physical AI is mistaking a convincing demonstration for a working product. The gap between "works in simulation" and "works on our floor" is the difference between a budgeted automation project and a write-off. Asking the right question before the money is committed is a leadership habit, not a technical one.

CONVICTION III

The winning move is judgement, not spend.

The advantage does not go to whoever buys the largest model. It goes to whoever matches the right approach to the actual problem and plans realistically around what it costs to run. That is strategy, and strategy is the chief executive's job.

Before you commit

Five questions every leader should ask

You do not need the mathematics. You need enough fluency to challenge a vendor claim, scope a pilot honestly, and tell a genuine capability apart from a confident label. Five questions carry most of the value.

Match method to requirement

“Which approach fits this problem?”

Visual variety and synthetic data favour world models. Numerical accuracy and physical contact favour physics engines. A vendor offering one as a cure-all for both is a vendor to question. Match the method to the requirement rather than to the headline.

Test on the unseen

“How does it perform on unseen, contact-heavy tasks?”

This is the one question that tells a leader most of what they need to know. A confident demonstration says little about real-world readiness. A genuine answer covers unfamiliar objects, real-world contact, and how the system was tested. A vague answer is itself a useful signal.

Plan around the cost of use

“What will it cost to adopt, adapt, and run?”

Running a model to make a prediction is undemanding. Fine-tuning the best models needs server-class compute. Budget for adapting accessible pre-trained models rather than training frontier ones from scratch, which is rarely the right call for most businesses.

Apply the strict definition

“Is it really a world model?”

The label is being stretched, and several products marketed as world models are something else on close inspection. A true world model is action-conditioned and predicts the next state of an environment you can interact with. Holding to that simple test separates real capability from confident branding.

Combine both, with clear eyes

“Which tool is carrying which load?”

World models and physics engines attack the sim-to-real gap from opposite directions, and neither fully closes it today. A practical programme uses both. Be explicit about which tool is doing which job, and evaluate each against the task it actually carries.

First-hand evidence

What we tested, and the lesson it points to

The simulation

Read more about our simulation in our blog series.

1The test part 1We trained pick-and-place policies entirely in simulation and tested them in simulation, without deploying on a real robot. On a single task with enough demonstrations, the policy succeeded reliably.
2The test part 2Asked to transfer to an object it had not been trained on, performance dropped sharply. The failures were instructive rather than random: incorrect object localisation, the gripper opening too early, and configurations the model had never seen.
3Lessons learnedThe tentative lesson is that broad competence does not come for free. It appears to need relevant data and careful evaluation. Read more about our simulation in our blog series.

Three predictions

What we expect to be obvious by 2027, and uncomfortable to admit today.

Forward calls invite disagreement. We make these on the record, with our reasoning written down, so the next volume of this briefing can hold us to them.

Within 12 months

The "world model" label gets policed

Buyers start applying the strict test before they sign: is it action-conditioned, and can you interact with it? Confident labelling stops passing procurement, and vendors reposition products that were never world models to begin with.

Within 18 months

Contact handling decides what goes commercial

Progress in differentiable and learned physics over the next eighteen months shapes which physical-AI applications become commercially viable. The organisations that evaluate against unseen, contact-heavy conditions set realistic expectations and avoid expensive disappointments.

Within 24 months

Physical AI approaches its ChatGPT moment

Physical AI today is where generative AI was a few years ago: clearly capable, still inconsistent. Much of the progress in language models came from the engineering built around them. A similar trajectory here points towards an inflexion point comparable to the 2023 ChatGPT launch. It would make robots and autonomous systems viable in settings out of reach today.

The two approaches

Learned or computed. Where each one belongs.

World model

Learned

from data and experience

Predicts plausible futures the way a batter reads a fastball
Strong at visual variety and diverse synthetic training data
Valuable wherever real-world trials are slow, costly, or dangerous
Less reliable where exact numbers decide the outcome
Example: simulating traffic, weather, and pedestrians for autonomous driving

→

Physics engine

Computed

from equations

Predicts numerically grounded futures from physical law
Strong at accuracy and physical contact, where force and geometry decide success
Simulates rigid bodies, soft bodies, fluids, and particle systems
The hardest shared problem is contact, which is mathematically discontinuous
Example: dexterous manipulation, part insertion, and locomotion that must transfer to hardware

Two ways to predict the world. Choosing is the strategy.

Builders are no longer experimenting at the margins. They are shipping.