"Companies with less than $10 million in revenue don't need to pay us at all"

About two months after Lightricks announced its intention to split into two independent companies, the company is launching LTX, a new video model for its AI operations. This time, the company seeks to present the video not only as a final product for creating films, animation, or marketing content, but as an infrastructure that can also be used for training robots, operating avatars in real-time, and building simulations.

The split in question is expected to be completed by the end of the year, subject to tax authority approval, and will effectively separate Lightricks' profitable app business, led by Facetune, from LTX. Each company will have its own management, board of directors, and investor recruitment plan. The move was accompanied by the layoff of 75 employees, alongside an intention to open about 25 new positions in AI fields.

Now, LTX is trying to establish itself not just as a company that develops tools for video creation, but as a provider of a "world model" — a system that receives a given state, predicts how it is expected to develop, and can help a machine plan its next step in the physical world.

"If you have a system that receives a frame and knows how to guess what the next frames will look like, it can also be used in robotics," says Dr. Zeev Farbman, co-founder of Lightricks and CEO of the new company, in an interview with Globes. "We want to build the expensive part, so that startups in the field have a base they can work on."

$50 million for the model

"This year, a breakthrough occurred following a paper called DreamZero," explains Farbman. "The researchers showed that a system that receives a frame and knows how to predict what the next frames will look like can also be used in robotics. A robot operating in the world is required to predict what a certain action will look like and what its results will be, and this is the same ability that lies at the base of a video model." However, unlike a video clip, a robot receives information not only from cameras, but also from touch, pressure, and temperature sensors.

LTX

Field of activity: Developer of open "world models" (Open World Models) — AI technology that creates video, audio, and simulates the physical world.
History: LTX is the AI division of Lightricks, which in a few months will become a company in its own right.
Employees: 250 in Jerusalem, New York, Los Angeles, and London.
Something else: The company's models are customizable on edge devices; according to the company, they are the most common in the world in the field, with over 35 million downloads.

According to Farbman, these data can be converted into a visual representation and fed into the model alongside the video. "Sensor readings become an image where each pixel represents, for example, the intensity of pressure or temperature. We train the base model, and companies adapt it to their data and uses."

Today, training a base model requires huge amounts of data and computing power. According to Farbman, this is the part of the process that most startups cannot fund themselves. He explains that "the most expensive process is training on all the data on huge amounts of video from the internet. We spend about $50 million a year on this. There are companies, particularly Chinese ones, that spend more."

"When you hear a startup say it's building a world model, you have to understand that it doesn't have a $50 million budget; we want to give these companies the expensive part: training the model. They can do the adaptation to their data and product themselves."

According to Farbman, one of LTX's technological bets is on building a model that needs a smaller number of tokens — the units of information the model processes — and thus reduce the amount of computing power required to run it. "Small companies operate with more limited computing budgets; it needs to be a model that a company can afford to run."

Implementing use in companies

The fact that LTX allows downloading the model, running it on private infrastructure, and adapting it to user needs raises a central question: how does the company intend to profit from it? According to Farbman, LTX does not want to rely only on selling access to the model via API, but to develop a licensing model that varies according to the industry and product. "As long as a company has less than $10 million in revenue, it doesn't need to talk to us at all," he says. "It can use the model for any use. Once it crosses the $10 million revenue threshold, it needs to talk to us about licensing."

The license structure, he says, will not be uniform. In the case of an online video creation service, LTX may receive a percentage of the revenue generated using the model. In the case of an animation company running the model on its servers, the payment may be derived from the volume of productions it creates.

This method allows LTX to first expand the use of the model among startups and researchers, without requiring upfront payment from them, and to start charging money only if the use of the technology develops into a business with significant revenue.

At the same time, Farbman divides the possible uses of the model into three areas. The first is media and entertainment, including the creation of animation clips and marketing content; the second is real-time video, for example, virtual avatars that can have a live conversation. "Instead of getting an answer in text, it will be possible to talk to a virtual lawyer who will appear on video or even as a hologram. For this, the model needs to generate video quickly and in real-time," he explains.

The third area is Physical AI, which includes robotics and simulations and also arouses interest among defense-tech companies. LTX does not intend to build robots itself, but to provide companies with the base model on which they can develop and train their systems.

Working with everyone

In the field of world models, LTX competes with companies with much larger resources, led by Nvidia, which developed the Cosmos platform for Physical AI applications. However, Farbman argues that it is precisely the interests of hardware manufacturers that create an opportunity for LTX. "Most giants choose a closed path, Nvidia is an exception." According to him, LTX seeks to operate in a way that is not tied to one chip manufacturer.

"It is very important to us that the model works on Apple hardware, on AMD, and on Nvidia. This combination is part of our differentiation." Alongside this, Farbman emphasizes the decision to publish the model openly. "We want to build an open infrastructure that everyone will want to use. Our belief is that safety comes from the fact that many people can look at the problem and see what is happening. Models that everyone can examine are safer."

The new model is being launched during a transition period of splitting for Lightricks. After its completion, Lightricks will include Facetune, the rest of the mobile products, and the influencer platform connecting brands to creators.

According to Farbman, fewer than 100 employees currently maintain the mobile activity, which generates revenue of more than $100 million. LTX, on the other hand, has about 250 employees, most of them in research and development roles, and continues to hire employees in AI-focused fields. According to sources, once the separation of the companies is completed, it will be possible to look at both financial partners for the separate activity and new potential investors for LTX.