The Five Muscles of Physical AI:
What It Takes to Scale Robot Data
Research, hardware, operations, software and deployment, all at once and all at scale. Sequoia's thesis for backing Mecka, unpacked and mapped to what a robotics data farm actually looks like on the floor.
By Serverless Ventures
Published: October 13, 2026 · Read time: ~8 minutes
The short version
Language models were trained on text the internet had already written. Robots have no such inheritance: the record of how bodies move, touch and recover in the physical world has to be captured before it can be learned from.
In its note on backing Mecka, Sequoia Capital names the five capabilities a company needs to do that capturing at scale: research, hardware, operations, software and deployment. Its real claim is the conjunction. Each one is hard alone; the business only works when all five run together.
Why robots have a data problem language models never had
On October 7, 2026, Sequoia partners Anas Biad and James Flynn published Partnering with Mecka: Bringing AI to the Physical World. Mecka, founded by Josh and Jason, describes itself as the data and deployment layer for physical AI: a partner to AI labs, robotics companies and enterprises for training data, reinforcement learning and deployment.
The note opens with an asymmetry worth sitting with. At every stage of the model lifecycle, the digital world gave AI something the physical world does not:
| Stage | Digital AI | Physical AI |
|---|---|---|
| Pretraining | Decades of text already on the internet | No ready-made record of movement, interaction and sensing |
| Reinforcement learning and deployment | Clean, structured environments that scale easily | Noisy, unpredictable, operationally complex real settings |
Sequoia compresses this into one line: physical AI "operates in the field, not in the cloud." The industry has to build both a brain and a body. The brain needs data nobody collected, and the body needs someone doing the work on the ground.
The five muscles
Sequoia calls them muscles, which is the right word: capabilities that only exist if you keep exercising them. For each one, here is what the note asks for and why it is harder than it sounds.
1. Research: see the next wave before it arrives
Identify where robotics is heading, and what that will require, ahead of the labs asking for it.
- Why it is hard: the field is early. The form of the data may change and so may the form of the robots, so a collection programme built for today's model architecture can be obsolete by the time it reaches volume.
- What it buys: lead time. Hardware and operations take quarters to stand up, so the research call has to be made well before demand is visible.
2. Hardware: build the instrument, then build it cheaply
Design the bespoke hardware and wearables that capture the right data, then manufacture them fast, economically and at scale.
- Why it is hard: the capture devices mostly do not exist off the shelf. You are building the infrastructure that creates the dataset before you can create the dataset.
- What it buys: data quality that cannot be added later. Calibration, synchronisation and sensing are fixed at the moment of capture.
3. Operations: every job in every industry
Run large-scale global collection. In Sequoia's words, "volume alone is not enough; the data has to span every job in every industry."
- Why it is hard: diversity is a logistics problem. It means recruiting, training and quality-controlling people across sites, tasks and environments, not adding servers.
- What it buys: coverage. A model generalises across the variation it has seen, and variation only comes from being in many real places.
4. Software and digital AI: raw capture into training-ready data
Turn the raw output into something a model can train on. That output is far more voluminous and complex than text: multi-camera video, body and hand tracking, sensor streams.
- Why it is hard: the streams have to be aligned in time, segmented, labelled and validated. A failed attempt mislabelled as a success teaches the model the wrong thing with full confidence.
- What it buys: unit economics. This is the muscle that turns hours of collection into usable data per dollar.
5. Deployment: robots in front of real customers
Put robots to work with end customers, for reinforcement learning and for production use.
- Why it is hard: it adds field service, safety and customer operations to a company that started out selling data.
- What it buys: the loop. Deployment exposes the failures and edge cases that tell you what to collect next.
The moat is the word AND
The sentence in the note that carries the argument is about the product, not the company: training data today "has to be high volume AND high diversity AND high quality AND delivered fast AND at a good price." The capitals are Sequoia's.
Sequoia's analogy is food delivery. It is easy to deliver a single dish from one restaurant to one apartment block. It is very hard to scale that into a DoorDash. The second half of the analogy matters more for investors: DoorDash built a system where, whether the input is a pizza or a laptop, the customer gets the same experience. Sequoia says plainly that it is not betting on Mecka having picked the right form of data. It is betting on a machine whose output stays what partners want as the input changes.
The cap table reads the same way. Alongside strategics including NVIDIA, Samsung, Microsoft and Qualcomm, the note lists operators such as Tony Xu of DoorDash, Frank Slootman, formerly of ServiceNow and Snowflake, and Milan Kovac, formerly head of Optimus at Tesla, described as people who excel in the disciplines the business requires. Logistics, data platforms, humanoid hardware: it is a roster organised by muscle.
From muscles to a factory floor
An investment note says what has to be true. It does not say what the work looks like. For that, the most useful companion we have found is Leo Su's two-part series on building a robotics data farm, from his newsletter Physical AI & Robotics: Demo to Deployment. It was Su's commentary on the Sequoia note that sent us to it in the first place.
Part 1 starts with the unit of account. For robots, he argues, it is not a video clip but a trajectory: "A video clip can show what happened. A robot trajectory records how the robot made it happen." A complete trajectory carries seven things:
- Observation: RGB, depth, point clouds, tactile, force and torque.
- Action: joint positions and velocities, end-effector pose, gripper state, base movement.
- Robot state: arm, end effector, base and sensor status.
- Task context: the instruction, the goal, the scene, the objects, the success criteria.
- Timing: timestamps and synchronisation across every stream.
- Outcome and labels: success or failure, why it failed, skill segments, recoveries.
- Calibration and metadata: camera parameters, robot and sensor configuration, quality report.
Around that unit he builds a five-layer pipeline (define the data unit, capture at the source, structure the task, validate before training, close the feedback loop). Part 2 adds the five building blocks of the farm itself: a production plan, a physical world of scenes and objects, a collection system, a production workflow, and day-to-day operations. Lay the two frameworks next to Sequoia's and they line up:
| Sequoia's muscle | Su's equivalent | What it looks like in practice |
|---|---|---|
| Research | Begin with a plan; define the data unit | Which tasks, which industries, what quality bar, how many trajectories a day |
| Hardware | Capture at the source | Rigidity, sensor calibration, multimodal sync, low frame drop, force and state feedback |
| Operations | Physical world; collection system; operations | Scene and object libraries, a robot fleet, trained operators, standard procedures, audits |
| Software and digital AI | Structure the task; validate; production workflow | Collect, process, annotate, validate, deliver, the same way for every trajectory |
| Deployment | Close the feedback loop | Model failures in the field become the next round of targeted collection tasks |
The mapping is ours, not either author's. One difference to keep in mind: Su writes about real-robot data, largely teleoperated, while Sequoia's mention of wearables and body and hand tracking points to human-worn capture as well. Su treats the sources (internet-scale human data, simulation, egocentric capture, real-robot data) as "complementary, not interchangeable." The production discipline carries across all of them.
Three lines from the floor worth keeping
- On hardware: "You cannot fully clean your way out of bad robot data." Downstream software can detect a misaligned or incomplete capture. It cannot recreate the physical truth that was never recorded. This is why muscle 2 cannot be outsourced to muscle 4.
- On operations: "The goal is not to build a robot lab. The goal is to build a repeatable production system for real-robot data." As a farm grows, he adds, consistency matters as much as throughput.
- On software: "Without a production workflow, a data farm produces recordings. With one, it produces reliable datasets."
Deployment is where the loop closes
The fifth muscle looks like a separate business bolted on to a data company. It is better read as the part that makes the other four compound. A deployed robot fails in specific, informative ways: poor lighting, a transparent object, something slipping from the gripper, a person walking through the scene. Each failure is a specification for the next collection task.
That is the strategic logic of bundling data and deployment. The supplier who watches robots fail at customer sites knows what data is worth collecting before the customer can articulate it, which feeds straight back into muscle 1.
Our read
Three things we take from this, and one caution. These are our views, not Sequoia's or Su's.
Three of the five muscles are atoms, people and logistics. Expect capital intensity, headcount and margins that look closer to a marketplace or a manufacturer than to a model lab.
Which capture method wins is still open. The durable asset is the ability to retool collection quickly when the answer changes, so ask how fast a company has switched before.
A small, software-first team cannot run global collection. It can build the layer that decides whether a trajectory is trustworthy, which every collector needs.
That last point echoes what we argued for digital agents in Build the Environments, Not the Workforce: when the deliverable is software with a working verifier, engineering is the moat. The physical version is harder, because the verifier has to reason about contact and timing rather than a rendered screen, but the shape of the opportunity is the same.
The caution: a funding announcement is advocacy, not analysis, and the note gives no revenue, volume or customer figures to test the thesis against. The open questions are real. Labs and robot makers with enough capital may bring collection in-house. Simulation and synthetic data may absorb more of the demand than expected. And a company holding all five muscles carries all five cost lines while the market is still forming. "All at once and all at scale" describes the moat and the burn in the same breath.
Eight questions for any robotics data company
- What is your data unit, and does it record action, state and outcome, or only video?
- What capture hardware do you own, and what does it cost per unit at your current volume?
- How are streams synchronised and calibrated, and what share of raw capture do you reject?
- How many distinct tasks, sites and industries does the dataset span, not just how many hours?
- How do you train and certify operators, and how do you measure consistency between them?
- Who confirms that a trajectory labelled as a success actually succeeded?
- When the preferred data format last changed, how long did it take you to retool?
- Do deployment failures reach the collection team as tasks, and how quickly?
Sources: Partnering with Mecka: Bringing AI to the Physical World by Anas Biad and James Flynn, Sequoia Capital, October 7, 2026; Leo Su, How to Build a Data Farm for Robotics, Part 1: The Real-Robot Data Pipeline and Part 2: Building the Data Farm. Sequoia's note does not state the size of the round; Su's commentary on it describes a $60M financing led by Sequoia, which we have not independently confirmed.