Business
The plan
We build the Robot Processing Unit (RPU) to run robot policies at 100 watts on the edge.
- Our first silicon targets a realistic >3× in energy efficiency, reaching >5× as the compiler matures, using proven methods. See technical roadmap.
- Long-term vision: co-design RPU once model architectures converge, achieving a >15× in energy efficiency. We will become the computational substrate for physical intelligence.
Edge inference is bottlenecked by energy efficiency, not memory bandwidth. A GPU spends <10% of its energy on arithmetic. Most of the energy is wasted on generality that is not needed as model architectures converge.
Milestones
| Milestone | Cost | Requirements |
|---|---|---|
| 1. Simulate virtually and on FPGA | $100 k, 3–6 months | Demonstrate a realistic path to a >5× energy-efficiency multiplier. |
| 2. Physical prototype + compiler stack | $5–10 M, 1 year | Validate the hardware on a test-chip shuttle; secure letters of intent. |
| 3. Production tapeout + integration | $50–100 M, 2 years | Mature the product into a drop-in replacement for Jetson Thor. |
| 4. Co-design | — | Capitalize on the convergence in model architecture by creating a highly specialized NPU that achieves a >15× energy-efficiency multiplier and dominates the market. |
Philosophy
At SpaceX, we learned that there is no such thing as pure software. Fundamentally, software is constrained by hardware. GPUs are impressive, but even at full utilization they cannot run state-of-the-art world-action models on the edge. We must dare to imagine the system as a whole, and create a specialized NPU for these workloads. This unlocks the ChatGPT moment for robotics.
We also don’t want to live in a world where our robots are puppets controlled by datacentersThis is a deeply unsatisfying vision of the future and falls far short of the scifi dream of true embodied intelligence. hundreds or thousands of miles away, at the mercy of network transients. Besides the reliability issues, remote inference cannot scalably achieve human-like reaction latencies. And the cost of renting expensive datacenter racks can quickly surpass the BOM cost of the robot itself.
We have personally felt the pain. We have built inference stacks at frontier robotics labs. Even with the most aggressive optimizations, real-time control still required two remote B200 Blackwell GPUs in a datacenter a thousand miles away. Also, extraordinary software complexity, like a custom ultra-low-latency QUIC transport and distributed relay system for bidirectional streaming of camera frames and trajectories. It worked, but it can never scale beyond the demo.
To keep ourselves accountable, we follow Elon’s algorithm.
- Make the requirements less dumb. Challenge every assumption. The industry assumes these workloads cannot run on the edge, and therefore require complex remote-inference stacks.
- Delete the part. Most of a GPU’s area and energy exists so it can run anything.
- Simplify and optimize — only what survives deletion: one dataflow, memory managed by the compiler, and optimization spent where the joules actually go.
- Accelerate cycle time. Invest in simulation and fail fast. Aggressively pursue tapeout timelines.
- Automate. Our long-term mission is to create a company-as-a-machine that allows us to capitalize on the convergence of model architecture and co-design an unbeatable chip.
The Long Term Vision
Once the first chip ships, we turn the whole company toward co-design: building each new generation of silicon around the model architecture the industry is converging on. Three curves are meeting at once — architectures are converging, model compression is maturing, and memory is moving onto the compute die.
A chip built for a known architecture reaches far past the first product. Our simulation bounds it above 15× a Jetson-class GPU. We will be able to run the largest world-action models of the future in real time and onboard, benefiting from improvements in both model efficiency and silicon. It is how we become the computational substrate for physical intelligence.
Challenging Our Assumptions
Our core bet is that robot policies keep getting bigger — and that is not guaranteed. How many parameters does it actually take to act usefully in the physical world? Nobody knows. There is no clean comparison to the brain: a synapse is not a parameter, and while the brain has on the order of 100 trillion synapses, a robot almost certainly needs far fewer. How many fewer, we can’t yet say.
But the demand for onboard intelligence looks effectively unbounded. Teams will always want to pack as much intelligence into a robot as they can, because more capability wins. Like a gas expanding to fill its container, models grow to use whatever compute a robot can carry — so we expect to stay hardware-constrained for a long time.
We are not betting against GPUs eventually running the most state-of-the-art robot policies on the edge — model efficiency and silicon both keep improving, and one day the gap may close enough for a general-purpose part to suffice. But even in that world, a specialized chip is an insurmountable advantage. GPUs will remain useful for prototyping, but we will dominate the market.