What are world models and how do they differ from LLMs?

World models are AI systems trained to simulate and predict how physical environments behave. Where large language models process and generate text, world models learn the physics, geometry, and causality of real or constructed spaces. They can predict how an object will move if you push it, render a realistic 3D scene from a text description, or simulate the consequences of an agent's action before it is taken. Genie 3, the DeepMind model that Emulate's founders co-built, generates photorealistic interactive environments from simple text prompts and was integrated into Waymo's autonomous driving simulation pipeline.

Who are the founders of Emulate AI?

Emulate was founded in August 2026 by Jack Parker-Holder, Matthew McGill, and Philip Ball, all former members of Google DeepMind's world-model team. Parker-Holder was co-lead on Genie 3 and stated his goal as 'building large-scale world models to solve embodied AGI.' McGill and Ball are also credited on the Genie 3 research. All three left DeepMind to start Emulate shortly after Genie 3 shipped.

Why is a $700 million seed round significant?

Pre-product seed rounds above $500 million are rare even by 2026 standards. The round, led by Index Ventures and Lightspeed Venture Partners, would value Emulate at roughly $3.7 billion including the new capital before the company has shipped a commercial product. It is larger than Recursive Superintelligence's seed and close to Ineffable Intelligence's $1.1 billion round, making it one of the three largest UK AI seed rounds this year. The scale signals that institutional investors are betting world models represent a structurally different AI capability from LLMs, not just an incremental improvement.

What are the enterprise applications of world models?

World models enable enterprises to create synthetic training environments for physical AI systems without the cost or risk of real-world collection. Autonomous vehicle companies like Waymo use them to run billions of driving scenarios. Manufacturers can simulate robotic assembly lines, failure modes, and maintenance procedures. Logistics companies can model warehouse layouts and agent paths. Defence and aerospace can test autonomous systems in adversarial simulations. For enterprises deploying AI agents, world models let you test an agent's behavior across thousands of edge cases in simulation before any deployment to a real environment.

Who else is building world models?

The world-model space has three notable players launched from DeepMind in 2026 alone. Ineffable Intelligence, founded by David Silver and backed at a $5.1 billion valuation on $1.1 billion in funding, and Emulate are both UK-based. Yann LeCun's AMI Labs, raised $1.03 billion in March 2026, pursues a related research agenda on Joint Embedding Predictive Architectures. Outside DeepMind lineage, Google itself continues developing Genie internally, and several autonomous vehicle labs run their own proprietary simulation stacks built on transformer-based world models.

Emulate AI: DeepMind Alumni Seek $700M for World Models

Three of the researchers who built Google DeepMind’s Genie 3 have left to found Emulate AI, and investors are ready to put $700 million behind them before the company has shipped a single commercial product. As of September 17, 2026, the UK startup is in advanced talks to close a seed round at a pre-money valuation of around $3 billion, which would make it one of the most richly valued seed-stage companies in history. Bloomberg and the Financial Times reported the deal simultaneously on Thursday morning.

The round is led by Index Ventures and Lightspeed Venture Partners, with Creandum also participating. At $700 million, it exceeds the seed raised by Recursive Superintelligence and trails only Ineffable Intelligence’s $1.1 billion round among this year’s UK AI mega-seeds. Post-money valuation is estimated at $3.7 billion, the FT reported, though Bloomberg cited a $3 billion pre-money figure, suggesting the final number will depend on how much of the $700 million target closes.

What Emulate Builds

Emulate’s focus is world models: AI systems trained to simulate and predict how physical environments behave. Unlike large language models, which process and generate text, world models learn the geometry, physics, and causality of real and constructed spaces. They can predict how objects move, generate photorealistic interactive environments from text prompts, and simulate what happens when an agent takes a specific action, all without any real-world sensors or deployments.

The founders, Jack Parker-Holder, Matthew McGill, and Philip Ball, all worked on Genie 3, DeepMind’s general-purpose world model. Parker-Holder was co-lead of the world-model team. His stated goal in public posts before leaving DeepMind: “building large-scale world models to solve embodied AGI.” The team published the Genie 3 research in early 2026, and all three are listed as contributors on the official model page.

Genie 3 is already deployed in production. Waymo built its World Model on top of it, using the architecture to run hyper-realistic autonomous driving simulations at scale. That real-world validation of the underlying technology is part of what makes Emulate’s fundraise credible at this valuation, even at the seed stage.

Why World Models Matter for Enterprise AI

The enterprise case for world models is not primarily about chatbots or coding assistants. It centers on training and testing AI agents that operate in the physical world or in complex multi-step workflows.

The core problem: deploying an AI agent into a real environment, whether a factory floor, a hospital network, or a financial trading system, requires extensive testing across rare and dangerous edge cases. Gathering that data in the real world is expensive, slow, and sometimes impossible to do safely. A capable world model provides a synthetic environment where agents can run billions of iterations, fail safely, and be evaluated on corner cases that might never occur in a data collection campaign.

Waymo’s integration shows the pattern. Rather than driving test vehicles millions more miles to cover rare scenarios, Waymo generates those scenarios in simulation using a world model built on Genie 3. The agent trains on the synthetic experience and transfers the learned behavior to the real car. The same logic applies to industrial robots, logistics agents, autonomous drones, and any AI system that must reason about physical causality before acting.

For enterprise software agents, the implication extends beyond physical environments. A world model that understands document workflows, UI interactions, or ERP system state could allow organizations to simulate what an AI worker will do before granting it live access. That simulation-first deployment model is increasingly relevant as enterprises govern multi-agent systems operating across sensitive business applications.

The DeepMind Spinout Wave

Emulate is the third major AI lab to spin out of DeepMind’s London office in 2026 alone.

Company	Founders	Round	Valuation	Focus
Ineffable Intelligence	David Silver (AlphaGo lead)	$1.1B seed	$5.1B	World models, planning
Emulate AI	Parker-Holder, McGill, Ball (Genie 3)	$700M seed	$3.7B	Physical world simulation
Recursive Superintelligence	Richard Socher et al.	$600M+ seed	Undisclosed	Self-improving AI

The pattern is consistent: researchers who led a specific breakthrough at a major lab leave to commercialize the capability independently, and they raise capital at valuations that would have been inconceivable for pre-product companies three years ago. Investors are not betting on a product roadmap; they are betting on the team’s ability to repeat and extend what they proved in a closed research setting.

Yann LeCun’s AMI Labs fits a parallel track. LeCun’s Joint Embedding Predictive Architecture research, a competing approach to world models, attracted $1.03 billion in March 2026. The field is consolidating around two or three technical approaches, all with significant institutional backing.

The competitive dynamic for enterprises is one they have seen before in LLMs: a wave of well-funded labs racing to establish a capability benchmark, followed by commoditization of the base technology and competition at the product and integration layer. Enterprise buyers who ignored the early LLM wave until 2024 lost two years. World models are arguably at an earlier and more defensible stage.

Enterprise Implications

Enterprises thinking about AI strategy through 2027 should register a few signals from this round.

Physical AI is no longer a side bet. Three major DeepMind spinouts, Waymo’s production integration of Genie 3, and the depth of capital flowing into simulation infrastructure all point to physical AI becoming a core enterprise category. Companies with significant physical operations, logistics, manufacturing, facilities management, should be evaluating world-model vendors now rather than waiting for the technology to mature further.

Simulation-first agent deployment is becoming standard. The playbook that Waymo demonstrated, validate in simulation before shipping to production, is the template that the next generation of AI agent platforms will follow. Enterprises that deploy AI agents into live systems without a simulation-testing layer face disproportionate governance risk as agents acquire more autonomy.

Talent concentration matters. The Emulate round is also a signal that the people most capable of building the next generation of physical AI have left the large labs. Enterprises looking to partner with or acquire world-model capability should expect to engage with independent labs, not just platform vendors.

Emulate has not announced a product timeline or a commercial API. The round is still closing. But given the Genie 3 heritage and the pattern set by the other DeepMind spinouts, the company is likely to be a significant supplier to enterprises building physical AI and simulation-grounded agent systems within 12 to 18 months. If you are planning your AI infrastructure roadmap for 2027, Emulate AI is worth watching closely.

For enterprises already building agentic platforms that must navigate the growing complexity of multi-model physical AI, the question is not whether world models will become part of the stack. It is whether to build on top of the emerging open standards or wait for a platform vendor to abstract the complexity away. Enera’s perspective on that choice is that by the time the abstraction layer arrives, the organizations that understand the underlying capability will have an irreversible advantage.