Intelligence belongs
in everyone’s hands.

01 / The conviction

No company should own humanity’s skill library.

Folding, fixing, growing, assembling, explaining: useful knowledge comes from people doing the work. Turning that knowledge into AI should not require handing a private company exclusive control over what everyone else can learn.

Our thesis is that edge inference and robotics will make intelligence part of everyday infrastructure. More of it will run close to the person, on a local machine, or inside a physical system. The skills behind that infrastructure should be inspectable, adaptable and available beyond any one platform.

Teach it once. Let more people build on it.

DVIDIA’s commitment is to open skill releases, portable artifacts and visible credit. Contributors keep the rights they hold. A workspace helps organize the work; it does not grant ownership of an everyday action. We will not make exclusive corporate ownership of shared skillsets our business model.

02 / The opportunity

The scale is real.
Access is the unfinished work.

Robotics already has a substantial installed base. Open weights already serve real inference demand. The opportunity is to make useful capabilities easier to share, adapt and run across people, software and machines.

These figures measure different things and cannot be added into a total market. We have no verified estimate of the worldwide private inference market or DVIDIA’s addressable revenue. Our market starts with specific tasks people can reproduce and services they choose to pay for.

03 / The raw material

Start with someone doing the work.

The supply comes from demonstrations people choose to contribute, consented robot teleoperation, openly licensed research, and clearly labeled simulation or scripted rollouts. Preserve who made each contribution, how it was collected and what its license permits.

These are foundations to learn from, with specific robots, tasks and licenses. Publicly visible footage is not automatically reusable training data. A person’s video can explain a task, but it cannot supply robot actions or sensor readings that were never recorded. Missing information stays missing.

04 / From data to skill

A useful skill has to survive contact with the world.

Distillation means preserving the parts that make a task work: the goal, conditions, steps, actions, feedback and recovery from mistakes. A collection of clips is a starting point. An executable, evaluated release is the result we are working toward.

1. 01

   Record with provenance

   Keep original media, the task, contributor credit and permission to use it. Sharing begins with an explicit choice.

2. 02

   Structure the evidence

   Curate episodes and annotations. Preserve real sensor and action data. Robot datasets must pass a LeRobot v3 load-and-replay check.

3. 03

   Build and test the skill

   Produce an agent workflow, script or trained robot policy. State the supported model, hardware and environment; measure outcomes and failures.

4. 04

   Release something reproducible

   Publish the version, source, component licenses, data references, setup instructions and evaluations so others can inspect, improve and run it.

05 / Private inference

Your model.
Your machine. Your rules.

Open weights let people choose where inference runs. A workshop can use a local machine. A team can operate a private server. A community can offer compatible compute and compete on price, performance and reliability.

By liberated, we mean the freedom to inspect, adapt and operate a model on your own terms—including deployments without a hosting provider imposing additional content filters. People should be able to choose unfiltered models and control their own runtime policies. Downloadable weights still carry their licenses and behaviors learned during training; the word “open” alone does not remove them.

Our direction is a competitive inference market built around portable models, clear terms and verifiable service. Keep private work private. Make switching providers possible. Drive the cost of useful intelligence down through shared work and competition.

06 / An open economy

Make access cheap.
Make contribution worthwhile.

The ambition is broad access at the lowest practical cost. Useful open releases should remain available to study, adapt and run. Original licenses and attribution travel with them, and anyone should be able to build compatible tools around them.

Premium skill spaces will be paid places to develop and maintain a skill together. Creators will be able to set prices for packaged downloads and useful services, while the underlying open-source releases remain accessible. People can pay for a maintained package, an adaptation, evaluation, support or delivered compute. Participation should not require surrendering the skill to a platform.

Contributors should share in the value of work that earns revenue. That requires recorded contributions, agreed terms and a clear calculation based on collected sales. A free download does not create an earnings balance. A paid workspace does not buy exclusive rights to an everyday human skill.

The founder skillspace is our proposed starting point: a paid workspace with a branded first-person recording necklace, where someone takes responsibility for a useful task, recruits contributors and helps improve the release. We are exploring an 80% founder share of a defined revenue pool. Pricing, hardware fulfillment and the complete contributor agreement must be established before that proposal becomes an offer.

Inference operators follow a separate path: open-source management software, verified endpoints, measured usage tied to an authenticated account and settlement address, and payment from collected customer revenue. Any network-token rewards need separate, published rules. A connected machine or a token counter is not an earnings balance.