Robots Don't Need More Compute. They Need This.
Encord co-founders Eric and Ulrich discuss their $60M Series C, the company's origins before the AI hype, and their focus on building the essential data infrastructure for physical AI and robotics—the next frontier after LLMs.
The Core Problem: Data Infrastructure for Physical AI
A model is only as good as the data it is trained on. Encord is building an AI-native, universal data layer for physical AI and robotics to solve this fundamental problem. The platform helps the world’s leading AI teams—in autonomous driving, robotics, and other embodied AI applications—to create, manage, annotate, and evaluate their data. The goal is to ensure the right data goes into a model while keeping the wrong data out, as even slight errors in a dataset can significantly impair a model’s real-world performance. As AI models scale in complexity, so does the data challenge, making robust data infrastructure critical not just for initial training but for continuous improvement post-deployment.
An Early Bet on AI Before the Hype
Encord was founded in the late 2010s, stemming from the co-founders’ experiences in deep learning research and high-frequency trading. They identified that out of the three core ingredients for AI—models, compute, and data—the data component was the most time-consuming and defensible. They saw the inefficiency in the common practice of sending data to be manually labeled and believed there had to be a better way.
This was before the ChatGPT explosion when AI was not a hot investment category. The founders recalled struggling to raise their seed round, even being rejected by a fund that subsequently invested in an Icelandic dating company, believing its market was larger than AI’s.
The ChatGPT Catalyst: Validating AI for AI Companies
The release of ChatGPT was a pivotal moment. Before it, there was significant skepticism, even from AI companies, about using AI to process and label their own data. Many clients preferred human oversight, not trusting AI for these critical tasks. ChatGPT served as a massive market education tool, demonstrating that AI could be trusted for a wide range of generalizable use cases. This shift in perception was crucial, as it made customers more receptive to Encord’s core value proposition: using specialized “micro models” and other AI-driven techniques to automate the annotation and data curation process.
Following this shift, Encord invested heavily in multimodal AI capabilities—handling not just images and video, but also text and audio together. This positioned them perfectly for the rise of physical AI.
Physical AI: The Next Frontier is a Data Problem
The founders draw a key distinction between digital AI (LLMs) and physical AI (robotics). The success of LLMs was enabled by the vast availability of text data on the internet, which proved that with enough data and compute, models could become incredibly performant. The primary risk was the cost of compute.
In physical AI, the situation is reversed. The compute infrastructure is now readily available, but the bottleneck is the scarcity of high-quality, real-world, embodied data. To achieve the “ChatGPT moment” for robotics and unlock similar scaling laws, the data problem must be solved first. Encord is positioned as the data infrastructure company to enable this breakthrough.
A New R&D Facility for Real-World Data Collection
To address the data collection bottleneck, Encord has opened an R&D facility in the Bay Area. This space allows robotics companies to bring their own hardware and collect training data in a variety of controlled environments. Encord provides the infrastructure to manage this process, effectively tackling the pre-training and data collection phases of the pipeline, which are far more complex for physical AI than simply scraping the internet.
This initiative is part of a larger strategy to manage the entire “data flywheel.” By providing a single, consolidated platform for data collection, indexing, curation, annotation, and model evaluation, Encord helps companies accelerate their development cycles and get products to market faster.
The Indispensable Role of Humans in the Loop
Despite the focus on automation, humans remain critical to the process for two main reasons:
- Pushing the Frontier: For novel tasks where models are still learning—like folding laundry or emptying a dishwasher—human input is essential for creating the initial training data.
- Exception Handling & Safety: The stakes are much higher in physical AI. A bad output from a chatbot is a minor inconvenience; a “hallucination” from a self-driving car or a drone can have catastrophic consequences. The error tolerance is much lower, which increases the quality bar for data and necessitates human oversight for control, supervision, and handling edge cases.
Vision for the Future
With a new $60 million Series C, Encord aims to accelerate its growth and capture the enormous opportunity in physical AI, which represents the vast majority of the world’s economy. Their long-term vision is for every single piece of physical AI data to be processed through their system. The founders advise others in the space to have a clear long-term vision but to be flexible and non-dogmatic about the path to get there, letting the market guide the journey.