Emulate was incorporated in August. It has no product. It has no website. It has not disclosed how many people work there. As far as the outside world is concerned, it is three names and a registration document.
It is in advanced talks to raise as much as $700 million, at a valuation of roughly $3.7 billion including the new money, in a round reported by the Financial Times and led by Index Ventures and Lightspeed Venture Partners.
Divide that valuation by the number of days the company has legally existed and you get a figure somewhere north of $80 million per day. That is not a typo, and it is not really a joke either. It is a fairly precise measurement of what a very specific kind of expertise is worth in the autumn of 2026.
The short version
- Emulate is a London based AI company incorporated in August 2026, still in stealth
- Founded by Jack Parker-Holder, Matthew McGill and Philip Ball, all previously on DeepMind’s world model work
- In talks for a seed round of up to $700 million at about $3.7 billion post money, led by Index Ventures and Lightspeed
- The team came from Project Genie, DeepMind’s system for generating interactive 3D environments
- World models are not chatbots. They simulate how things behave, which is what robots and self driving systems need
- Terms are not finalized and could still change
- Two other ex-DeepMind spinouts have raised billion dollar rounds on similar terms in recent months
What a world model actually is
Almost everything the public knows about AI comes from large language models. You type, it answers, and the thing underneath is predicting text. A world model is a different object with a different job.
Nvidia, which has its own line of these, describes them as systems that “understand the dynamics of the real world, including physics and spatial properties,” and that can predict what happens next when something changes. The practical version: a glass leaves your hand, and the model knows it falls, knows roughly how it breaks, and knows what the pieces do afterward. A robot arm approaches a shelf, and the model knows which paths end in a collision before the arm moves.
| Large language model | World model | |
|---|---|---|
| Predicts | The next token of text | The next state of a scene |
| Trained on | Text, code, images | Video, sensor data, physics simulation |
| Output | An answer you read | An environment you can act inside |
| Good at | Language, reasoning about language | Consequences, space, cause and effect |
| Bought by | Anyone with a keyboard | Robotics, autonomy, simulation, games |
The strategic case for world models goes like this. A language model can describe the physical world because people have written about it. It has never been in it. If you want machines that operate in physical space, you need something that has learned consequences rather than descriptions, and you cannot get there by scraping more text.
Nvidia’s framing is that a pretrained foundation model “handles the heavy lifting” while targeted training on proprietary data handles the rest, “cutting development from years to months.” That is the pitch investors are buying: a base layer that every robotics company rents instead of building.
Why these three names are the entire asset
Emulate has nothing to demo, so the round is not about a product. It is about provenance.
Parker-Holder, McGill and Ball worked on DeepMind’s Genie effort, the line of research that produced systems capable of generating playable, interactive 3D environments from a prompt. Genie is one of the few world model programs that has visibly worked in public, and the people who built it are a very short list.
Venture investing at this end of the market has quietly become a talent market. You are not underwriting a business plan. You are paying for the option on a team that has already done the hard version of the thing once, on the assumption that they can do it again outside a corporate structure. It is closer to a transfer fee than a traditional seed round, and the sums have started to look like transfer fees too.
That logic is not new this year, it is just getting more expensive. Regulators have started noticing when very large amounts of money move in exchange for people rather than assets, which is why the Justice Department took an interest when Nvidia spent around $20 billion without technically buying a company.
The number, in context
Emulate is not an outlier. It is the third data point in a pattern that has formed over a handful of months.
David Silver, another DeepMind alumnus, raised roughly $1 billion for Ineffable Intelligence at a $5 billion valuation. A group of former OpenAI and DeepMind staff built Recursive Superintelligence to a $4 billion valuation on about $600 million. Set next to those, Emulate’s terms look less like an anomaly and more like the going rate.
Europe is part of this story too, and not by accident. The continent spent a decade watching its best researchers leave for American labs, and the money now arriving is partly an attempt to keep the next generation of them at home. That was the explicit argument when Mistral raised Europe’s largest technology round on a sovereignty pitch rather than a benchmark one.
The part the games industry noticed first
There is a detail in this story that sounds like trivia and is not.
When Genie was released in January, the companies whose share prices moved were not AI companies. Take-Two, Roblox and Unity all saw valuations slide. Investors looked at a system that generates interactive environments from a prompt and drew a conclusion about who builds environments for a living.
That reaction was almost certainly premature. Generating a playable space is a very long way from generating a game that anyone wants to finish, and the gap between those two things is where the entire craft of the medium lives. We wrote about this in detail when the first wave of these demos landed, and the short version is that AI cannot make good video game worlds yet, and the reasons are structural rather than a matter of scale.
But the market reaction tells you what the technology is understood to threaten, and it explains part of why a company building this gets priced the way Emulate is being priced. Games are the visible application. Robotics and autonomy are the one the money is actually chasing.
What would have to go right
It is worth spelling out the bet, because “no product” is doing a lot of work in the coverage and it cuts both ways.
| The bull case | The bear case |
|---|---|
| World models become the base layer for every robot, vehicle and simulator, rented rather than rebuilt | The base layer gets commoditized by Nvidia, DeepMind and a Chinese lab within two years |
| The Genie team is one of very few groups that has shipped a working version of this | Research talent does not always survive the transition to running a company |
| $700 million buys enough compute to skip the years of scrapping for GPUs | That much capital at seed sets a bar the company has to clear at every future round |
| Physical AI is the one large market that language models cannot serve | Real world data is far harder to collect than text, and nobody has solved that cheaply |
The compute point is the underrated one. A $700 million seed is not really seed capital, it is an infrastructure budget. It exists because training this class of model requires hardware access that a normal startup would spend three years and several dilutive rounds trying to assemble. Investors are effectively buying the team a running start against labs that already own their own data centers.
What to watch
- Whether the round closes on these terms. It is reported as advanced talks, not a signed deal, and terms at this size move
- What the first demo actually shows. A generated environment is a screenshot. A robot policy trained inside one and working in the real world is a business
- Who else leaves DeepMind. Three spinouts in a few months is a pattern, and the fourth will tell you whether it is still accelerating
- Whether anyone publishes benchmarks. World models have no equivalent of the leaderboards that made language model progress legible, which is convenient for everyone selling one
- How fast Nvidia moves. It is already shipping world foundation models and has a strong incentive to make the base layer free
The honest read on Emulate is that nobody outside the round knows whether it is worth $3.7 billion, including the people writing the checks. What they know is that the number of teams who have built a working world model is tiny, that the window to back one is short, and that being wrong about the price hurts less than not being in the deal at all.
That is not irrationality. It is a market where scarcity has moved from technology to people, and the invoices have started to reflect it.
