ProjectAtlas
Back to the presentation

The Global AI Race

Why Are World Models So Difficult to Build?

World Models represent a major technological advancement, with emerging research demonstrating their ability to simulate complex environments and evolving system dynamics. Developing frontier-level capabilities from scratch, however, requires extraordinary computational resources, high-quality data, specialized scientific expertise and substantial long-term investment. This complexity makes access to an existing operational World Model particularly valuable: it can significantly accelerate technological deployment, reduce the need to reproduce foundational development and create a powerful platform for further innovation. Reliability, scalability and application-specific performance remain important areas for continuous validation and improvement.

Main finding

Difficulty is scientific before it is financial. Compute and data set the entry price, a causal and stable representation of dynamics is still unresolved, errors compound over a rollout, evaluation is fragmented across incompatible benchmarks, and the senior researchers who can lead the work are few. Each barrier is a reason a licensed platform may be worth more than a build.

What supports it

  • Compounding error is structural: autoregressive rollouts feed the model its own predictions, so one-step inaccuracies drift the imagined state away from anything seen in training, superlinearly in horizon (arXiv 2607.06401, July 2026).
  • Evaluation is fragmented: visual quality, physical plausibility, intervention response and downstream decision quality are measured by different benchmarks, and an open-loop score is routinely over-read as control capability (arXiv 2607.06401, July 2026).
  • Training compute for frontier systems has risen by orders of magnitude across model generations, which is why the entry price is measured in tens to hundreds of millions of dollars (Epoch AI; Stanford AI Index).
  • A learned dynamics model is valid only inside the distribution it was trained on, which makes representation and regime change the central engineering problem rather than an edge case (Sutton & Barto 2018).
  • Talent is the binding constraint: a small number of researchers have shipped world-model systems at scale, and their compensation is not representative of the market.

What this does not claim

  • Vendor demonstrations are not measurements. Reported interaction lengths and realism are vendor claims unless independently reproduced.
  • Difficulty also protects incumbents: a barrier that stops us stops most competitors, which is an argument for licensing rather than a proof of value.

Planning assumptions used

  • Barriers are assessed for a financial application, not for a general-purpose frontier model.
  • Any evaluation we commission uses the same held-out data for every candidate system.

Open before signature

  • Which barrier is decisive for our use case: data, evaluation, or horizon?
  • What would it cost to reproduce one published result on our own data?
  • Which of these barriers disappears if we license instead of build?

Sources

Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.