To train AI models that generate lifelike video and control robots that mimic human movement, tech companies need tens of thousands of hours of very specific footage. Saint John, N.B.-based Versos AI hopes to build a big business by running the pipes that deliver all that data.

The firm helps clients “build a trainable dataset out of raw video files at scale,” CEO Chris Keevill told The Logic. Independent studios, streaming services, sports leagues and other content producers can use Versos’s technology to process, label and categorize their footage. Tech firms and brokers then search all those libraries at once to find and license what they need. 

Silicon Valley giants are spending huge sums on training fodder for AI models, which in turn is driving huge valuations for firms with very large or very specialized sets of data. In June 2025, Meta reportedly paid US$14.3 billion to take a 49 per cent stake in Scale AI, a data-labelling firm, and recruit its chief executive to run its new superintelligence unit. In November, Salesforce acquired Informatica, a data management firm previously backed by CPP Investments, in a US$8-billion deal. 

Model makers are also paying professionals to annotate data and train models on fields like banking, law or visual effects, often recruiting them via startups like Micro1, Mercor and Surge AI, which have attracted lots of funding and some blowback.

Versos is a minnow by comparison, but differentiates itself from those firms in two ways, according to Keevill. It’s focused on video rather than the text data on which the frontier models that power ChatGPT and Claude have historically been trained. And Versos is using AI rather than humans to annotate the data. 

To keep improving frontier models, tech firms are increasingly training them on video, which holds information about how things appear and interact in the real world. “Knowing what the AI is looking at is part of the secret sauce,” said Neha Khera, managing partner of the IRV Fund at Montreal-based investor Innovobot. “I’m watching a video—is that a ball, a tree, a child?” Versos’s technology provides the context to make the data useful, said Khera, who has backed the firm.

Keevill and chief science officer Cezar Grzelak co-founded Versos in November 2022, after seeing the image-generation model Midjourney and realizing AI would lead to a proliferation of new visual content. They set out to build a system that could identify individual scenes and objects within video and 3D files. 

In the early days, Versos was “a science project,” Keevill said. In its first year, the startup sought investment from RiSC Capital, but the Toronto-based venture firm couldn’t yet see a market for the product, recalled partner Colin Webster. 

Then, last spring, tech firms approached two studios with which Versos was working, seeking to license their content to train frontier AI models. That opened up a new market and new funding. In December, Versos announced it had raised a $1.85-million seed round led by Khera’s Innovobot, with participation from Webster’s RiSC Capital.  

Versos now has 10 staff, split between Saint John and Halifax. It’s “building traction with the frontier labs,” Keevill claimed, although he would not name its model-maker clients. He said tech firms are using Versos because they can easily find the footage they need, and avoid legal issues by licensing it from rights holders.  

As developers try to fill in gaps in their models’ knowledge, they’re making “very specific, very bespoke asks for types of videos to train on,” according to Keevill. For example, a tech firm might want to buy 10,000 hours of whale scenes. Not even the most dedicated ocean documentarian has that much, but Versos can pull it together from multiple studios that sell data on its platform. 

One client recently asked Versos for footage that appeared to have been shot from an iPhone or GoPro of internal spaces like homes and offices, plus depth information and 3D “meshes” of the objects captured. Such data could be used to train humanoid and other robots that could one day replace or operate around people. 

Keevill said the standard rate to license video for AI training is about a dollar a minute, so individual libraries with hundreds of thousands of hours of footage could be worth millions each time they’re sold. Specialized scenes can go for up to $30 per 30-second clip. “It is quite lucrative for the rights holders,” he said.

Versos charges content owners to process and label their data, and takes a cut of the licensing fee that they get from model-makers. It also plans to get into the data-generation business by filling gaps where the footage that tech firms want either doesn’t exist or isn’t plentiful enough. To that end, Versos has built a smartphone app that lets people shoot and upload video to add to its platform. 

Data marketplaces could help address some of the copyright issues AI currently faces. AI firms are already fighting scores of potentially costly legal battles over the data they used to train the last generation of their models. “They didn’t pay for a lot of the text that they got, and now they know that was a faux pas,” said Webster, adding that AI developers can avoid more problems by licensing video from rights holders via platforms such as the one developed by Versos.

While a small number of tech giants currently make up most of the demand for AI training data, Keevill sees a big market ahead for Versos. Builders of frontier models will keep needing more and different kinds of footage, plus data from sensors and other sources. And a popular new class of AI system called world models, which could someday control autonomous vehicles and robots, also trains on visual data. 

Versos plans to start raising a new round in the fall. “The dream is that we become the central utility for video training data, for the world, forever,” Keevill said.