a16zWhat iPhone Packaging Reveals About the Robotics Opportunity
a16z
Supply chain reality 0:00
Global supply chains are still tightly linked in 2026, and both U.S. and Chinese companies are thinking hard about long-term industrial strategy. The point is not just better robots. It is what those robots are for. Most robot pitches start with the model, but almost no one starts with the job at the end. Most of the market is greenfield, because almost nobody buying a robot has bought one before.
Full stack robotics 4:01
Josh Gruenstein says Tutor began with the idea that robots need to be deployed first, so the company can build the data, models, and supply chain needed for scale. He describes robotics as a systems problem that includes intelligence, hardware, software, customers, product, and manufacturing. In his view, the market is too fragmented for one layer alone to work, so the company has to make trade-offs across the whole stack to build a real platform.
Data and model tradeoffs 14:01
Real customer sites make robot data economically valuable, because you know the task is worth paying for. But the data can still be hard to get if the solution space is too high dimensional and humans must finish the last 10 percent for a while. The team is not strict about purity. It will use simulation and internet scale data to widen the data distribution and fill gaps where real robot data is thin.
Frontier models and hardware 15:02
The rise of frontier models has changed robotics fast. Labs now post-train on 3D spatial data like CAD and 3D modeling because there is a real market for it, and that same spatial reasoning helps action models. The business sees this as an accelerator, not a threat. It has lived through several shifts since 2021, from convolutional networks onward, and still sees the data flywheel as the lasting advantage.
Naming the robots 28:00
People name the robots after football players or pop stars because they get attached when the machine makes their jobs easier. A line worker is spared backbreaking work and can move into a higher role, while a factory manager gets a much easier day. The naming started with hurricane names, then moved on after the alphabet ran out. Now technicians pick names that bring them joy, and a large staff in the United States helps build the robots.
Serving smaller customers 29:00
The company found that most robotics firms chase huge Fortune 50 customers, but most of the market is made up of small and mid-market businesses. That pushed it toward a broader product. It wanted a general robot, not a one-off machine for one customer. By signing a mid-market contract in two weeks and putting a robot to work right away, it avoids the custom projects that can fail, stall, or disappear when one customer’s problem changes.
Real world robot learning 42:00
The team is building its research around real robot experience, much as top LLMs rely on real-world reinforcement learning traces. It has a 100-robot fleet, Data Factory 1, in Boston, which they say is the largest semi-humanoid data collection farm in the United States. They use it as a test bed for scaling reinforcement learning.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
