Buying GPUs does not buy the skill, the kitchen or the rights. Figure’s compute partnership is the training budget. Index is still the data side.
The deal that moved this month is easy to screenshot. Figure and Nscale announced a multi-year compute partnership: up to around 100,000 NVIDIA GPUs, multi-billion dollars of training capacity, Helix as the workload. Nscale becomes a preferred compute provider and a shareholder. The cluster talk is Vera Rubin, a Texas site, a number large enough to look like a foundation-model story wearing a humanoid. Brett Adcock’s own sentence is the one that matters for anyone who is not buying the GPUs. To put humanoids in homes, Figure is constrained by data and compute. Helix gets better the way every learned system gets better: more of both. Index is how they are trying to buy the first half. The partnership is how they are trying to buy the second.
That is the market post. Not a funding mash-up. Not a glossary of pretraining. A lab can light up a warehouse of accelerators and still have nothing for the policy to see. Compute scales a dataset. It does not invent a completed task, a chest-camera view of a grasp, a reject code, or a yes from the people in the room. If the files are lifestyle footage, you will train lifestyle footage faster. If the files are uncleared homes, you will train a diligence problem at industrial batch size.
Two constraints, only one of them is for sale at a cloud shop
Adcock did not say Figure was constrained by compute. He said data and compute. The industry prefers the second word because it fits a press release. GPUs have a vendor, a delivery date, a dollar figure, a photograph of a hall. Data has a kitchen, a person, a skill that was or was not finished, a face that should not have been in frame. Figure has spent the autumn showing both sides on purpose. Index came out of stealth in August: crowdsourced first-person video of real tasks, creators paid by the minute, devices in the field, bookings into homes and businesses, a pipeline they built after vendor zips failed the bar. The Nscale announcement is the other column on the same whiteboard.
Helix 2.5 already told you which column moved the robot. Figure pretrained on Index, then walked a humanoid into thirty unseen Bay Area homes. Holding task-specific data, architecture, training and evaluation fixed, Index pretraining alone moved zero-shot success from 9 percent to 56 percent on tidying, towels and beds. That is not a compute result. That is a data result with a compute bill attached. More GPUs would have trained the 9 percent policy faster. They would not have turned it into the 56 percent policy. The hours of aligned human video were the prior. The cluster was the furnace.
Teleoperation, hall video, simulation and found POV will all be poured into that furnace if someone does not label them. They are not the same object. A person driving a robot is useful. A camera in the corner of a demo hall is useful for checking whether the task finished. A simulator is cheap variation. A YouTube vlog is first-person-shaped. None of them becomes a specified, consented skill because you reserved another rack. Mix them if you must. Label them. When the mug fails, you need to know which object to collect more of – and you need the rights to collect it again. A hundred thousand GPUs will not write that sentence.
The kitchen does not ship with the cluster
Imagine the skill the compute is being asked to scale. A person unloading a dishwasher, or picking a mug off a cluttered counter, or opening a drawer that sticks. A third-person camera shows a body moving through a room. You can label the scene. You still do not see the grasp, the occlusion when the hand covers the mug, or the moment they change their mind and use a different cup. A chest-mounted or glasses-mounted camera sees the cupboard handle, the mug, the other objects in the way. For a policy that has to reach and place, that difference is the dataset. Buying more steps on the wrong viewpoint is how you get a very expensive prior on hall video.
The next kitchen has a different mug, a reflective hob, a person walking through. Simulation does not give you the damp sponge. Found video does not give you a chain that covers commercial robot training. Europe is tightening the web-copyright side this autumn, and homes and faces were never covered by “it was online.” Index is Figure’s answer to that gap: pay for the room, record the task, filter the fraud, keep going. Copying the GPU announcement without copying the capture loop is how a smaller lab spends a quarter on capacity and still has lifestyle footage on disk.
Rights travel with the camera, not with the purchase order for the cluster. A pilot in a lab is one consent conversation. A wearable in a real home is another. A booked creator in someone else’s cafe is a third. If faces are not required for the skill, keep them out. First-person cameras see children, post on the fridge, laptops on the counter. Blurring later is a processing choice. It is not retroactive consent, and it often destroys the hands you needed. If marketing stills or a launch video might appear later, say so now. You will not get a clean expansion of rights after the clips are famous inside a weight that took a multi-billion-dollar furnace to bake. Internal capture does not make the legal problem disappear. It makes it your problem at the scale of the cluster.
Quality is where the factory actually starts, and compute will happily scale the absence of it. Someone has to throw away the take where the task was not completed, the camera pointed at the ceiling, or a non-consented face walked through. Without reject codes you are scaling noise. Write the codes while everyone is still calm. Task not completed. Camera on the ceiling. Non-consented faces. Beauty filters. Digital zoom that destroys the hands. Automation can flag black frames. A person still has to apply the codes that mean “this is not the job.” If gold items do not exist before production, the first month of a larger cluster is the vendor – or the creator network – learning your kitchen on your budget, only faster.
Split eval by person so the same hands are not in train and test. A random split of clips is a test of whether the policy remembers those people. Public robot boards will not save you. They have the same contamination story the coding labs just admitted, with a mug. More GPUs make it cheaper to train on the eval by accident. That is not a reason to buy fewer GPUs. It is a reason to freeze the exam before the job scheduler starts.
What this means if you are not buying 100,000 GPUs
You will not copy the Nscale contract. You can still copy the sequence Figure keeps repeating when it is not talking about racks. Write the skill in one sentence. “Pick a mug off a cluttered counter, chest camera, indoor daylight, faces out, start to finish including the failures.” Decide teleop versus human demonstration versus first-person video. Lock the rig. Lock faces. Look at a catalogue match with the actual trainer before you reserve a cloud budget to replay the wrong files. Custom when the object set, the kitchen, or the faces decision is yours.
If you are standing up even a small capture loop this quarter, treat it as an operation with gold and reject codes, not as a film shoot that you will “scale later when we have compute.” Headcount and cluster reservations come after the sentence. Starting with a warehouse – of cameras or of GPUs – is how briefs balloon. Synthetic-only fills will not rescue a missing human baseline. They inherit the seed. They are not the kitchen. Label them. Do not pour a generator into a larger furnace and call the absence of a person a scaling law.
A familiar week, and it will happen in labs that cannot spell Nscale. Someone books capacity because a humanoid competitor booked capacity. The model card is ready. The hour-count on disk is a mix of hall video, a YouTube dump, and a week of people walking through houses. Friday looks like a training run. Monday is a mug the policy has never seen, a face that should not have been in the weight, and an eval split that contains the same hands as the train set. Nothing in the reservation of GPUs taught the policy to pick the mug. Nothing in the reservation cleared the face. The constraint Adcock named first was data. The screenshot was compute. Teams copy the screenshot.
Physical AI is not short of cluster announcements. It is short of specified, consented behaviour in the rooms the robots will enter. That is why Index exists alongside the Nscale deal, not underneath it. That is why a generic “egocentric for robotics” brief still produces lifestyle footage, even when the trainer finally has enough steps. Buying GPUs buys the right to be wrong at scale. The skill is a sentence. The kitchen is a room you were allowed to film. The rights are a pack a person can read. None of those arrive on a pallet of accelerators. Collect the skill, or license a set that already has one. Then, and only then, the furnace is a good purchase.
People4.ai would rather kill a bad spec on the first call than watch it train at cluster scale. hello@people4.ai · Contact
FAQ
If we buy enough compute, can we make up for a thin human-video set?
No. Compute scales what you already have. Helix 2.5’s own story was Index pretraining moving zero-shot success, not a larger cluster turning hall video into a kitchen. Thin data trained faster is still thin.
Is teleoperation enough if we cannot stand up an Index-like app?
Teleop is a different object from first-person human demonstration. Useful. Robot cameras, robot dynamics, a different rights conversation. It does not become crowdsourced kitchen video because you reserved GPUs. Spec it separately.
Do rights get easier once capture is internal and the cluster is ours?
No. Internal capture makes the legal problem yours. Homes, faces, screens and a booked workplace still need yes or no lines. A purchase order for compute is not consent.
Related reading
Humanoid labs are building data factories, not just models · Physical AI is suddenly hungry for first-person human video