Our Investment in Mundo

The data supply chain for perceptual intelligence

We’re proud to partner with Jason Liao, Kenneth Wu, Garreth Lee, Naj Anwaer, Ahnaf Muqset, and the Mundo team as they build the data layer for the next generation of AI.

Mundo creates proprietary data supply chains from real-world organizations: all the “boring” places where the internet stops and actual expertise begins.

The simple idea: the world’s best data is not sitting on the internet. It is locked inside real-world organizations that do not yet know how to monetize it.

The Data Is Not on the Web Anymore

The first era of AI was built on the open internet. Scrape enough text, add enough compute, and models get dramatically better.

That era is changing.

The next frontier is not just more tokens. It is a richer human experience: how people speak, pause, interrupt, gesture, perceive, reason, and interact with the physical world. Voice agents, video models, robotics, avatars, and multimodal systems all need data that captures the messiness of the real world.

But that data often does not exist in a clean, labeled, model-ready form. It is fragmented across old-school organizations: companies that have high-fidelity equipment, specific verbiage and dialects, expert reasoning, and specialized workflows that are impossible to scrape from the web.

Most data companies try to solve this with freelancers. Mundo is building software to turn businesses into data suppliers.

From Freelancers to Supply Chains

Traditional data collection is fragile. A freelancer records audio on an iPhone in a closet. A gig worker labels a task without domain context. A vendor scrambles to assemble supply for each new request.

Mundo’s approach is different. They can source expert-grade, multilingual audio data tailored to the specific needs of frontier research teams.

That matters because frontier labs do not just need “more data.” They need highly specific and high-quality data. Can a model handle interruptions? Can it speak Hindi with a Mumbai accent? Can it produce Arabic speech with the right local cadence? Can it understand jargon-heavy medical or legal audio?

Mundo’s platform can assemble the right supply chain for the job, then iterate as model needs change.

Most tools deliver datasets. Mundo delivers the factory that makes them.

Why This Moment Belongs to Mundo

Three shifts make this urgent now.

First, AI has moved from the pre-training era to the post-training era. Models are no longer differentiated only by scale. They are differentiated by the quality, specificity, and freshness of the data used to fine-tune, evaluate, and improve them.

Second, AI is becoming perceptual. The next generation of models will need to understand not just text but speech, video, motion, domain expertise, and real-world behavior.

Third, customer pull is already evident. Mundo is working with customers including OpenAI, Microsoft, Thinking Machines, and other LLMs and frontier labs.

This is not a market that needs evangelizing. Labs already know data is the bottleneck. Mundo is giving them a new way to source it.

The Team Moves Like a Lab and Operates Like a Marketplace

We love companies where the team’s intensity matches the shape of the problem. Mundo is one of those companies.

Jason Liao, CEO, is a former quantitative researcher and AI researcher. The original insight came from his time at Tsinghua, where he saw that researchers were often training in English simply because the data infrastructure for non-English languages did not exist. That became the seed for a much broader thesis: the AI data layer is underdeveloped for anything beyond scraped English web text.

Kenneth Wu, CTO, brings the engineering and systems background needed to turn messy real-world operations into software. 

Garreth Lee, CPO, adds deep AI data credibility from Hugging Face and Cohere, where he worked on major pre-training datasets.

Naj and Ahnaf are anchor operations, which are core to Mundo’s advantage. This is not a pure software product pretending that the physical world is simple. It is a coordination problem across customers, supply nodes, domain experts, quality control, timelines, and rights.

The company was a lean team of five when we invested, which makes the traction even more remarkable. The Series A is not about proving that customers want this. It is about giving Mundo the engineering, product, and operations capacity to say yes to demand without breaking quality.

The Moat Is in the Nodes

The best marketplaces do not just aggregate supply. They make supply more valuable by connecting it to a network.

Mundo is doing that for data.

Every new node increases what the platform can create. An individual company or enterprise supplying data is not enough. But stitched together through Mundo’s software, they become a programmable data supply chain.

That network compounds. Mundo signs exclusivity agreements with partners, maintains exclusive relationships with the majority of its base, and has seen partners renovate facilities and train staff on Mundo’s workflows. That is the kind of moat we like: not just software, not just services, but a living supply chain that gets harder to copy every time it works.

The Supply Chain for AI’s Next Frontier

Mundo starts with a painful wedge: frontier labs need better audio data, especially for multilingual and speech models.

But the bigger vision is more important.

The same platform that can create bespoke audio nodes can turn physical-world organizations into the training infrastructure for AI systems that need to perceive and act in the world.

As AI becomes more capable, the bottleneck moves closer to the real world. The winning models will need data that is fresher, more specific, more human, more physical, and harder to replicate.

That is what Mundo is building.

We’re excited to partner with Jason, Kenneth, Garreth, Naj, Ahnaf, and the Mundo team as they build the data supply chain for perceptual intelligence and turn the world’s most underutilized organizations into the infrastructure layer for frontier AI.

‍