Beyond generative AI
World Models:The Next Frontier of Artificial Intelligence.
Exploring the next frontier of artificial intelligence, where systems move beyond generating answers toward understanding, simulating and interacting with the world.
Project ATLAS examines what world models make possible across technology, economics and strategy, and where the evidence still has to be earned.
Interactive briefing · 12 questions · 4 parts
Part I
The Global AI Race
The frontier of artificial intelligence has moved from producing language to simulating how systems behave. World Models are the reason, and the race to build them now sets the pace of everything that follows.
The Global AI Race
Why Are World Models Becoming AI's Next Frontier?
A generative model predicts the next word. A World Model learns how something actually behaves, then runs it forward. That shift turns AI from a system that describes the world into one that can test a decision before taking it, which is why the frontier has moved here.
- From describing the world to simulating how it changes
- A simulation lets a decision be tested before capital moves
- The same capability is what lets an agent plan several steps ahead
The Global AI Race
Who Is Leading the World Model Race in 2026?
No single lab leads on every axis. Google DeepMind holds the interactive frontier, World Labs shipped the first commercial product, and Odyssey, NVIDIA and a cluster of robotics labs each own a different slice. Capital followed: more than $3B went into world-model startups in the first half of 2026.
- Capability is split across labs, not concentrated in one
- Capital is abundant; commercial maturity is still early
- Each player is defined by the environment it simulates
Six different problems, not one market
Interactive simulation
DeepMind · Odyssey · Runway · Decart
Spatial intelligence
World Labs
Physical world models
NVIDIA · AMI Labs
Robotics and embodied AI
Skild AI · Physical Intelligence
Financial world modelling
Generative market simulators · financial foundation models · research to early product
General-purpose platforms
OpenAI · Anthropic · Google
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.
- Compute, data and simulation infrastructure set the entry price
- Small errors compound, and long rollouts drift from reality
- Evaluation and senior research talent are the binding constraints
Part II
Technology and Economics
What a World Model really costs to build, why a general model does not simply substitute for one, and where simulation can honestly improve a financial forecast.
Technology and Economics
What Would It Really Cost to Build One?
Independent frontier development is a research programme, not a purchase order: $100M to $300M over 18 to 30 months on our own planning estimates, concentrated in senior research staff, GPU time, experiments that fail, data and deployment. Most of it is committed before anything works.
- Staffing, compute, experimentation, data and deployment, in that order
- Exceptional researcher packages are exceptions, not typical salaries
- Most of the budget is spent before the first working system
Two planning scenarios sit inside this range. $100M–$150M is a focused build: a narrower applied system, a smaller senior team and fewer training runs. $150M–$300M covers the broader frontier programme, where the money is concentrated in senior research staff, GPU time, data preparation, experiments that fail, evaluation and deployment, almost all committed before anything works. The $200M model is one composition inside that wider range.
The duration follows the work, not a market benchmark: assembling and retaining a senior research team, preparing and licensing data, developing the model, then repeated experimentation, training runs, evaluation and deployment readiness. Eighteen months assumes a focused team and few failed runs; thirty months allows room for the experiments that do not work. It is an illustrative project schedule, not a verified industry requirement.
An exceptional individual package, not an average salary: base compensation, equity or participation in the upside, and retention commitments across a multi-year period, for researchers with a public record of building frontier systems. Published evidence on frontier-lab pay is limited and usually indirect, which is why the figure stays a scenario. It is deliberately not the fully loaded average staffing cost spread across the $200M programme, which covers ordinary senior, engineering and operations roles at market rates.
Planning estimates. The full cost model is open in the development cost research.
Technology and Economics
Why Not Just Use ChatGPT, Claude or Gemini?
General models are excellent, and for many tasks they are the right answer. The comparison is not a subscription against a platform: it is a general model plus the integration, evaluation and operations we build around it, against a licensed platform that already carries financial and blockchain capability.
- A subscription answers requests; a platform carries operational capability
- Integration, evaluation and operations decide the total cost
- No head-to-head result on our tasks exists yet, and we do not claim one
Technology and Economics
Can World Models Improve Financial Forecasting?
Yes, progressively, and the horizon is widening. Research published through 2026 shows generative market simulators reproducing order-book behaviour and stylised market facts, financial foundation models forecasting across instruments from one pretrained model, and deep order-flow models extracting signal over multiple horizons. A World Model adds what a standalone predictor lacks: a state that carries history, dynamics that can be rolled forward, and a forecast whose uncertainty is explicit. Faithful long-horizon rollout and behaviour under regime change remain open, which is exactly what an evaluation contract is for.
- One market state: prices, order flow, liquidity, on-chain activity and macro together
- Simulate before committing: alternative paths and regime shifts rolled forward
- Probabilistic ranges with stated assumptions, scored out of sample
Part III
Financial Applications
Liquidity, security, operational load and machine-to-machine commerce: the places where the technology meets a blockchain financial business that already operates.
Financial Applications
How Could AI Transform Blockchain Finance?
Across four measurable surfaces, all of which a market-state model can serve: liquidity intelligence that reads venues, pools and chains as one book, execution and capital efficiency inside a fragmented market, continuous review of smart-contract risk, and token utility that follows paid usage rather than preceding it. Each surface has an observable state, a forecast that can be scored and a baseline already running today.
- Liquidity, depth and fragmentation as one observable state across venues and chains
- Volatility transmission and liquidity stress signalled ahead of the move
- Execution, reconciliation and capital allocation scored against today's baseline
- Smart-contract security as a continuous check, never the only signature
Financial Applications
What Happens When AI Agents Work Together?
One generalist tries to do everything and drifts. A team of specialists, each watching one part of the environment with shared memory, produces a single view that can be checked and corrected. The pattern that works keeps verification outside the agent that produced the work.
- Monitoring, reasoning and verification split across specialists
- Shared memory keeps conclusions consistent and auditable
- Independent verification before publication; human ownership at the end
Financial Applications
How Will AI Agents Buy and Sell Intelligence?
x402 makes it concrete: a request comes back priced instead of refused, the agent pays from an allowance a person approved, and the data is released in the same exchange. Settlement can run in stablecoin, in a native token where the rail supports it, or in a hybrid the treasury manages.
- A request priced, paid and delivered without a person in the loop
- Settlement design chosen per rail: stablecoin, native token or hybrid
- Any token utility follows paying usage, not the reverse
x402 payment journey: an external agent requests intelligence, pays, and receives the data.
Part IV
The Strategic Opportunity
Why licensing an existing capability beats building one, what the three-year proposal covers, and how the advantage could compound over a decade.
The Strategic Opportunity
Why License Existing Technology Instead of Building?
Because the alternative is a multi-year frontier programme with an uncertain outcome. A licence converts that programme into an operating capability in months, at roughly a tenth of the capital, and carries the technical risk with the provider rather than with us.
- Months to capability instead of years of research
- About a tenth of the capital, with the technical risk carried by the provider
- The option to build stays open, and exit terms protect it
The Strategic Opportunity
What Exactly Would the License Deliver?
Access rather than ownership: three years of a working platform shaped around our use cases, with a custom user experience, integration support and validation on our own data. Source code, model weights and ownership stay with the provider.
- Functionality that already runs, without a development programme
- A customized layer we own operationally, on a platform we do not own
- Usage rights for three years, not an asset on the balance sheet
The opportunity in figures
Planning estimates, subject to technical and commercial validation. Each figure is an internal assumption, not a confirmed supplier quotation.
The Strategic Opportunity
Why Could This Become a Major Strategic Advantage?
Because the case compounds rather than saves once. Better liquidity and services, sharper economic intelligence, earlier risk control, lower avoidable operating cost and new AI services, with a defined share of what is saved reinvested in capability that stays relevant as the technology moves.
- Better liquidity, services and economic intelligence
- Smarter risk management and lower avoidable operating cost
- Reinvestment in capability keeps the organization technically current
Board recommendation
The opportunity is not to join the race from zero.It is to accelerate with technology that already exists.
Advanced World Models are becoming a foundational intelligence layer for financial infrastructure, automated services and risk management. Existing advanced AI technology, specialized financial applications, established blockchain infrastructure, autonomous AI services, x402 payment capability and disciplined financial controls together form a credible foundation for long-term technological growth, without opening a frontier development programme from zero.
Before final execution
- Capabilities and measured performance
- Economic value against the alternatives
- Three-year total cost of ownership
- Commercialization rights
- Data ownership and portability
- Service continuity and exit terms
- Security and model-change control
- Renewal and pricing conditions
- Regulatory compatibility by jurisdiction
Project ATLAS is proposed as a validation-first path: technical proof on our own data, a financial case against the real alternatives, and a commercial structure that can be exited.
Advancing the licensing opportunity through technical validation, commercial negotiation and defined investment approval milestones.