Robots Don't Need More Compute. They Need This.
TL;DR
Encord’s bet is that physical AI is bottlenecked by data, not compute — Ulrich argues robotics has the opposite problem from LLMs: we already have massive compute infrastructure, but real-world multimodal data is scarce and hard to curate.
The company went from an “unsexy” YC idea to a 300-customer business with $110 million raised — Encord says it now works with more than 300 AI teams including Toyota, has 150 employees across London and San Francisco, and just announced a $60 million Series C led by Wellington Management.
ChatGPT didn’t just create demand — it taught even AI companies to trust AI on their own data — before that moment, Encord’s customers still preferred human labelers over automation, despite Encord already having “micro models” that could pre-label from just two or three examples.
Encord is expanding upstream into real-world data collection for robots — the company opened a Bay Area R&D facility where robotics companies can bring robots into controlled environments to generate pre-training data Encord then manages, annotates, and evaluates.
Humans still matter most at the frontier and in exception handling — in physical AI, mistakes are far costlier than a bad chatbot response, so people remain essential for supervising edge cases like laundry-folding robots, self-driving cars, or drones that cannot afford to hallucinate.
Ulrich’s clearest founder advice: indecision compounds like debt — his takeaway from building Encord is that even wrong decisions are usually cheaper than waiting, because “you’re constantly paying interest on decisions that you don’t make.”
The Breakdown
From data plumbing to the physical AI boom
Nico opens by framing Encord as “AI-native data infrastructure,” and Eric and Ulrich make the pitch simple: if models are only as good as their data, someone has to make sure the right data gets in and the wrong data stays out. Their platform is a universal data layer for physical AI — create, manage, annotate, and evaluate multimodal data for robotics and other real-world systems.
The early conviction looked niche — even inside YC
Ulrich traces the idea back to the late 2010s, when he was doing deep learning research at Imperial and Eric was shipping models in high-frequency trading. They believed data, not models or compute, was the most painful and defensible layer — even while investors were chasing fintech, crypto, and remote work. One fund rejected them saying AI wouldn’t be a big enough market, then invested in an Icelandic dating company instead.
The first product: automate labeling, especially for computer vision
The original Encord product was built to replace the clunky workflow of sending data “to the Philippines” for labeling and waiting for it to come back. They had a clever system of “micro models” — tiny specialized models trained on just two or three examples that could automate annotation across slices of the data distribution. The twist: customers often didn’t trust AI to touch their data, even if they were AI companies themselves.
Why ChatGPT changed Encord’s trajectory
ChatGPT was less a product pivot than a market education event. Once people saw a general-purpose AI system work, they became more willing to trust AI for annotation and workflow automation. Encord then leaned harder into multimodal AI — not just images or video, but image plus text plus audio — which naturally pushed them deeper into robotics and physical AI.
“For robots, the internet doesn’t exist”
Ulrich’s core thesis lands here: LLMs advanced fast because the internet gave them cheap pre-training data, so throwing more compute at the problem worked. Physical AI is the reverse — the compute exists, but the embodied real-world data has to be collected the hard way. That’s why Encord is now building Bay Area facilities where robotics companies can bring robots into environments designed for data generation and training.
The full flywheel: pre-training, deployment, and what breaks in the real world
Encord isn’t stopping at labeling; it wants to own the whole data flywheel from indexing and curation through pre-labeling, post-training, deployment, observability, and exception handling. That matters because physical AI has almost no tolerance for mistakes: a bad ChatGPT answer gets a thumbs-down, but a hallucinating self-driving car or falling drone can do real damage. Humans stay in the loop both at the frontier and as supervisors for edge cases.
Customers, ambition, and the Stripe analogy
Today Encord says it serves more than 300 AI teams across robotics, self-driving, and autonomous systems, and highlights a YC robotics startup building a laundry-folding robot as one example. With its new $60 million Series C, the company wants to expand fast because, as Ulrich puts it, 80% of the world’s economy is still about moving or doing things in the physical world. Their ambition is blunt: become for physical AI data what Stripe is for payments.
Hiring humans — and agents — plus the founder lesson that stuck
Asked about growth, they say they’re hiring in both London and San Francisco, but also hiring “agents” for solutions, marketing, sales, and engineering, including one already living in Slack with a name and face. Ulrich ends on two founder lessons: make decisions faster because indecision accrues “interest,” and don’t be dogmatic about your route. His metaphor is memorable: a startup is a rowboat in a stormy ocean — you know the island you’re trying to reach, but you can’t just draw a straight line and ignore the waves.