A warehouse in San Leandro shows what capability building looks like before anyone calls the market obvious
In a warehouse south of Oakland, a young man named Andrew Ceja plays Jenga while wearing a headset. Cameras track his gaze. Sensors along the band read his brain activity as he decides which block to pull and how much force to apply. His job title is pilot. His employer, Encord, calls the work robot training.
The scene reads as novelty. Read the org chart behind the scene instead, and something more instructive appears: a leadership team deciding to build a capability the company was never designed to have.
The question was never "what does the market want next"
Encord was founded to help companies annotate data and evaluate machine-vision models. A tooling business. Useful, defensible, well understood by investors.
Then the customers changed what they were doing. Robotics firms moved toward end-to-end learning for manipulation tasks, and the datasets required for the work were nowhere on earth. As Vineeth Velmurugan, Encord's head of robot learning, puts the problem: "The data simply does not exist."
Most management teams treat a sentence like this as a market timing note. Wait. Watch. Revisit next planning cycle. Encord's leadership drew the opposite conclusion and decided to manufacture the missing input rather than manage someone else's version of the same.
Boards should notice the shape of the decision. The customers did not ask for a new product. They hit a wall. The wall, and not the roadmap, defined the opportunity.
Manufacturing a capability is an organizational act, not a product decision
Extending a product is a familiar motion. Building a capability is a different order of commitment, and the operating requirements land fast.
Encord now runs physical facilities. Leader-follower robotic rigs where a human operator moves one arm and a second arm mimics the motion. Storage racks of plastic vegetables, kitty litter trays, fake flowers, bundled cable. Egocentric video pulled from factories across several countries. A dozen pilots on staff, drawn from the AI annotation workforce, running tasks including pouring coffee and stacking poker chips because every humanoid company has asked for exactly those datasets.
None of this existed in the original operating model. Leadership had to stand up real estate, workflow design, quality standards, and a workforce category with no established labor market. Ceja arrived by way of a waste management company, where his curiosity about technology put him in charge of a robotic trash sorter.
The talent decision underneath the shift deserves particular attention. Encord did not promote a strong internal generalist into the role. Leadership hired Velmurugan, a veteran of OpenAI's robot lab and Berkshire Grey, the warehouse automation firm. Someone who had lived inside the constraint before being asked to solve the constraint commercially.
When a company enters territory the existing team has never operated in, pattern recognition is the scarce asset. Boards tend to underprice how long an organization takes to acquire the pattern on its own.
Disciplined bets look different from expensive faith
The brain wave work with Zander Labs, a German neuroscience startup, is instructive precisely because leadership has not committed to scale.
The stated plan: build an initial brain wave-tagged dataset, run the dataset through customer robotics models, evaluate whether performance actually improves, then decide. Lukas Gehrke, the Zander neuroscientist supervising the effort, thinks measures of mental effort during a task will tell model builders when to deploy their most expensive models. A reasonable hypothesis with a defined test and a decision point attached.
The economics get examined with the same discipline. Densely annotated data, labeled with descriptions like "right hand tightens bolt," is worth roughly 100 times what Velmurugan calls junky ego data, and costs about 20 times more to produce. On paper, a good trade. Leadership tracks both numbers rather than the flattering one.
The comparison to language models breaks here, and executives building AI strategy should absorb why. Frontier labs assembled text corpora by scraping the open web for close to nothing. Physical data has to be manufactured. Velmurugan estimates the field needs something like five times the video volume of YouTube. Manufacturing costs do not fall the way scraping costs did, and any board hearing an AI plan built on the LLM analogy should ask which of the two economies the plan actually lives in.
Position becomes strategy
The most durable asset Encord built may be the vantage point. Sitting between many robotics companies at once, the company sees which data techniques gain traction across the industry before any single customer sees the pattern.
Leadership recognized the position as a strategic asset and made the position part of the pitch. Capability creates information, and information compounds into advantage. Few leadership teams name the effect early enough to invest behind it.
Questions worth putting on the agenda
Where are our customers hitting walls our roadmap does not address?
If we entered that territory, what capability would we have to build rather than buy, and who on this team has built one before?
Which of our current bets has a defined test and a decision point, and which are running on conviction alone?
What do we see, because of where we sit, that no one else in our market sees?
Markets shift faster than org charts. The leadership work is closing the gap deliberately, with the right operator in the seat and a clear view of what the new capability costs.
Christian & Timbers advises boards and CEOs on the leadership required to enter markets before the markets are obvious.