# Mundo raises $20M to build training data for AI beyond text > Source: > Published: 2026-08-24 18:04:16+00:00 # Mundo raises $20M to build training data for AI beyond text **GreatPoint Ventures led the Series A as the Vancouver startup expands from multilingual datasets into audio, video and evaluations.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [Mundo AI](https://mundoai.world/research/perception-is-the-next-frontier) ## Why it matters Mundo is betting that frontier AI's next constraint is proprietary sensory data: conversations, video and human context that cannot be recovered from web text or reduced to transcripts. [Mundo](https://mundoai.world/), the Vancouver AI data business founded by Jason Liao, [Garreth Lee (@garrxth)](https://x.com/garrxth), [Naijide Anwaer (@naijide_a)](https://x.com/naijide_a) and Kenneth Wu, [announced on August 24th](https://mundoai.world/research/perception-is-the-next-frontier) that it raised a $20 million Series A led by [GreatPoint Ventures](https://www.gpv.com/). Y Combinator, [Next Frontier Ventures](https://nextfrontiercapital.com/) and [E12 Ventures](https://www.e12.ventures/) also participated. Mundo said the financing follows a previously unannounced $4 million seed round, bringing its total funding to $24 million. Mundo did not attach a valuation to the Series A. Liao's path to Mundo started with the data problem the founders are now selling investors on. According to [Mundo's Y Combinator profile](https://www.ycombinator.com/companies/mundo-ai), he encountered a shortage of usable non-English training data while conducting machine-learning research abroad, after working as a quantitative researcher at a Canadian hedge fund. Lee previously worked on pretraining data at Cohere and tokenization at Hugging Face, while Anwaer was a platform product manager at Binance.US. Wu previously held engineering and quantitative roles, including at Amazon Web Services and the Ontario Teachers' Pension Plan. The four founders took Mundo through Y Combinator's Winter 2025 batch with a narrower proposition: supplying AI developers with human-created datasets in languages poorly represented in the material available online. Mundo worked directly with native speakers and ran collection, annotation and quality-control operations in the countries where those languages were spoken. The Series A marks a broader product and research bet. Mundo now describes itself as a data layer for "perceptual intelligence," its term for models' ability to interpret speech, video, gestures, timing and environmental context. The work includes datasets, model evaluations and applied research across audio, video and other emerging data formats. That repositioning follows the direction of frontier-model development. A transcript preserves the words in a conversation while discarding interruptions, tone, hesitation and body language. Video models face similar gaps around sequence, intent and incomplete views of a scene. Mundo is betting that labs will pay for data designed around those failures, along with evaluations that measure whether a new training run actually fixed them. Mundo says its data and evaluations are already used by leading AI labs, without supplying customer metrics in the funding announcement. An [Augment Ventures portfolio profile](https://www.augmentventures.com/portfolio/mundo/) previously described Mundo as infrastructure for non-English model training, while a July 2025 investor update said Mundo was sourcing and structuring audio, image and video data across more than 150 languages. Those claims came from Augment, an earlier Mundo backer. The progression from multilingual data to broader sensory data is commercially important. Language collection gave Mundo an operating network for sourcing human input across regions and cultures. Audio and video can make that network more valuable because model developers need controlled environments, clear provenance and task-specific labels that cannot be assembled by scraping another tranche of websites. Mundo's approach also carries an execution constraint. Custom data collection, evaluation design and applied research can become labor-intensive as each model lab requests different modalities and failure cases. Mundo has to turn those projects into reusable software and repeatable operations if revenue is to grow faster than headcount. Y Combinator currently lists Mundo with 30 employees, up from the four founders at launch. Mundo said it will use the Series A to expand research, engineering and operations, and is directing applicants to its [careers page](https://mundoai.world/careers). The funding gives Mundo room to test whether AI laboratories will treat multimodal data and evaluations as recurring infrastructure rather than one-off collection contracts. GreatPoint's $20 million bet says the next bottleneck sits outside the model architecture, in the expensive work of teaching machines what humans notice without being told.