Every robot AI stack covered so far in this series — GR00T, Cosmos, teleoperation and video training pipelines — eventually has to run somewhere inside the robot's own body, in real time, without a data center connection, and that's the job of edge inference chips: NVIDIA's Jetson Thor line and Tesla's in-house AI5 processor represent two fundamentally different bets on how to solve that problem, one selling compute to the entire robotics industry and the other building it exclusively for one company's own fleet. The chip you choose here determines what your robot can actually decide on its own, without phoning home.
NVIDIA Jetson Thor: Selling the Industry Standard
NVIDIA's Jetson Thor platform, expanded in mid-2026 with new T2000 and T3000 modules alongside the existing T5000, is built to be the industry-wide default rather than a single company's proprietary tool. The flagship Jetson Thor delivers up to 2,070 FP4 TFLOPS of compute with 128GB of memory at 40 to 130 watts of power, which NVIDIA describes as 7.5 times the AI performance and 3.5 times the efficiency of its previous-generation AGX Orin platform. The new T2000 module extends that architecture down to more affordable edge systems, offering 400 FP4 TeraFLOPS and 16GB of memory as an entry point for autonomous mobile robots and industrial manipulators, while the T3000 is positioned as achieving similar inference performance to the larger T5000 for multimodal workloads specifically, at a lower cost during a period of high memory prices. NVIDIA also released Cosmos 3 Edge alongside this hardware lineup — a 4-billion-parameter, lightweight version of the Cosmos world foundation model covered earlier in this series, specifically built to run directly on Jetson Thor hardware for real-time on-device robot policy and vision reasoning, rather than requiring a cloud connection.
The commercial breadth here is real and measurable: NVIDIA reports that robotics companies including UBTECH, Agile Robots, and Connect Tech reduced memory usage by as much as 15GB through software optimization tools, letting them shift from a 64GB Jetson AGX Orin module down to a 32GB module — a genuine cost reduction passed on to companies building on NVIDIA's platform rather than their own custom silicon.
Tesla AI5: A Moat, Not a Product
Tesla's AI5 processor, which completed tape-out (finalized its physical chip design and sent it to a foundry) in mid-April 2026, targets roughly eight times the raw compute and nine times the memory capacity of its AI4 predecessor. Crucially, AI5 is designed primarily for Optimus robots and Tesla's own supercomputer clusters, not simply as a faster chip for existing vehicles — despite most initial press coverage framing it as "faster FSD." Unlike Jetson Thor, Tesla has no plan to sell AI5 to other companies; analysts frame the chip explicitly as a competitive moat rather than a product, similar in spirit to how Google built its own TPU chips and Amazon built Trainium and Inferentia specifically to reduce dependence on NVIDIA's pricing and supply constraints for their own massive internal workloads. The practical effect described by industry analysts isn't that Tesla threatens to become a chip vendor — it almost certainly won't — but that Tesla's success accelerates the broader custom-silicon trend, setting a public performance benchmark that other large-scale robotics and autonomy companies now have to either match with their own chip investment or accept paying NVIDIA's premium instead.
The Market These Chips Are Actually Competing In
The edge AI hardware market overall is tracked at $26.14 billion in 2025, projected to reach $58.90 billion by 2030 — a 17.6% compound annual growth rate that reflects how central this specific hardware layer has become across robotics, autonomous vehicles, and industrial AI. NVIDIA's approach bets on capturing the broadest possible share of that growing market by selling compute to virtually every robotics company that isn't building its own chips. Tesla's approach bets that owning its full stack, from chip design through the robot's neural network, is worth forgoing that external revenue entirely in exchange for cost control and independence from NVIDIA's roadmap and pricing.
Where This Actually Matters for Building Robots
For nearly every company in the humanoid robotics space outside Tesla itself — Figure, Boston Dynamics, Agility Robotics, and the broader partner ecosystem covered in the first entry of this series — Jetson Thor and its edge variants are the realistic, available option, backed by a full software stack including Cosmos 3 Edge for on-device world-model inference and TensorRT for model optimization. Tesla's AI5 represents what vertical integration at massive scale can achieve, but it's simply not an option available to purchase, which is the core structural difference running through this entire series: NVIDIA's stack is designed to be adopted broadly, while Tesla's is designed to be an advantage no competitor can access.
Edge Inference Chips at a Glance
| Chip platform | Peak compute | Availability | Strategic purpose |
|---|---|---|---|
| NVIDIA Jetson Thor (T5000) | 2,070 FP4 TFLOPS, 128GB memory | Commercially available to any robotics company | Capture the broadest share of the edge AI robotics market |
| NVIDIA Jetson Thor T2000/T3000 | 400 FP4 TFLOPS (T2000); T3000 near-T5000 performance | Commercially available, lower-cost entry points | Extend Jetson architecture to budget-sensitive robotics |
| Tesla AI5 | ~8x compute, ~9x memory vs. AI4 (targets, unconfirmed final specs) | Tesla-exclusive; not sold externally | Vertical integration moat; independence from NVIDIA pricing |
Frequently Asked Questions
What is NVIDIA Jetson Thor used for?
Jetson Thor is NVIDIA's edge AI computing platform for robotics, offering up to 2,070 FP4 TFLOPS of compute for running vision-language-action models, world foundation models, and other AI workloads directly on a robot without needing a cloud connection.
Will Tesla sell its AI5 chip to other robotics companies?
No. Tesla's AI5 processor is designed exclusively for Tesla's own Optimus robots and internal supercomputer clusters. Industry analysts describe it as a competitive moat rather than a product, similar to Google's TPU or Amazon's Trainium chips.
How much faster is Tesla's AI5 chip compared to its predecessor?
Tesla targets roughly eight times the raw compute and nine times the memory capacity of the AI4 chip it replaces, though these are stated targets from the April 2026 tape-out announcement rather than independently confirmed final specifications.
Why does edge AI hardware matter for robots specifically?
Robots need to make real-time decisions without depending on a constant cloud connection, especially for safety-critical tasks. Edge inference chips like Jetson Thor let a robot run large AI models directly onboard, and the overall edge AI hardware market is projected to grow from $26.14 billion in 2025 to $58.90 billion by 2030.