Tesla officially tapes out its next-gen AI5 chip, a specialized 10x performance NPU designed for Optimus Gen 3 humanoids and the Dojo supercomputer.
What "Tape Out" Actually Means for AI5
Taping out is the point where a chip's design is finalized and handed to a foundry to be manufactured. It marks the end of the design and verification phase and the start of fabrication, when the layout is frozen and committed to silicon. For a part like Tesla's AI5, reaching tape out signals that the architecture, memory layout, and instruction set for its neural processing unit are locked, and the remaining work shifts to producing physical wafers, testing them, and validating that the manufactured chip behaves as the simulations promised.
This is a meaningful milestone precisely because it is expensive to reverse. Any design flaw discovered after tape out typically requires a new revision and another manufacturing cycle. Announcing a tape out is a statement of confidence that the design is ready to become real hardware for both Optimus Gen 3 humanoids and the Dojo supercomputer.
Why a Purpose-Built NPU Instead of General Hardware
A neural processing unit is designed specifically to run the matrix and tensor operations that dominate machine learning workloads. Compared with general-purpose processors, an NPU can dedicate far more of its silicon to the multiply-accumulate math that neural networks rely on, and less to the branching and control logic that CPUs need. The claimed 10x performance for AI5 reflects this specialization: rather than making a broad chip faster, Tesla is building hardware whose layout matches the exact shape of the models it will run.
Advertisement
Specialization also pays off in efficiency, which matters differently across AI5's two targets. Some of the tradeoffs a dedicated NPU lets a designer tune include:
- How much on-chip memory sits next to the compute units, reducing slow trips to external memory
- The numeric precision the hardware supports natively, trading accuracy for speed and lower power
- How data flows between processing elements, so intermediate results stay on-chip
- Power draw per operation, which is critical in a battery-powered robot and in a data center paying for electricity
One Chip, Two Very Different Homes
The interesting part of AI5 is that the same silicon is aimed at both a humanoid robot and a supercomputer. Optimus Gen 3 needs inference that runs locally, within a tight power and thermal budget, so the robot can perceive and act without waiting on a network connection. Dojo, by contrast, is about scale: many chips working together to train and process models where total throughput matters more than the constraints of a single mobile device.
Serving both from one design forces disciplined engineering. The team has to pick an architecture efficient enough for an untethered robot yet capable of ganging together into a large training system. When it works, the payoff is that models developed and trained on Dojo run on hardware closely related to what executes them inside the robot, narrowing the gap between how a model is trained and how it is deployed.
What to Watch Next
Tape out is a beginning, not a finish. The practical questions now are how the manufactured chips perform against the simulated targets, how yields hold up, and how quickly the part moves into shipping hardware. For anyone tracking this, the useful signals are real-world efficiency in Optimus, how AI5 scales inside Dojo, and whether the promised performance survives contact with physical silicon rather than staying a design goal.