Tesla Optimus — Architecture Analysis

Tesla's published portfolio tops at 180 Nm for rotary joints. EYOU PH20 at 182 Nm is effectively identical. Harmonic FHA-25C at 108 Nm is a precision-first alternative. Green Harmonic (Suzhou, 688017) is Tesla's actual supplier — single joint module drops from 2.5 kg (traditional RV) to 1.2 kg using harmonic drive. Tesla buys from Green Harmonic at 30–40% below Japanese list price.

Tesla's linear actuators use inverted planetary roller screws (static screw, rotating nut) — 2–3× higher load capacity than ball screws, 85–90% efficiency. Portfolio: 500 N / 3,900 N / 8,000 N. No commodity linear actuator matches 8,000 N in a humanoid package. Planetary roller screws cost $1,350–$2,700 each (Morgan Stanley est.); 14 per robot = $19K–$38K at today's prices — the single biggest BOM driver at ~19% of total cost.

Tesla uses parallel mechanisms at the ankle — multiple actuators sharing load for higher stiffness and precision. Linear actuators handle 2 DOF (dorsiflexion + inversion) within a compact volume. Commodity rotary options deliver 25–50 Nm which is insufficient for dynamic push-off forces. Ankle parallel mechanism architecture also appears in Optimus knee patent (parallel linkage, second link coupled via linear actuator).

Tesla's 20 Nm small rotary is achievable with commodity servos — Dynamixel XH540 at 10.6 Nm covers ~53% or pair two for 21+ Nm combined. Gen 3 wrist has 2 DOF controlled by 2 dedicated linear actuators in the forearm assembly (not the 20 Nm rotary), making this the closest commodity match in the entire stack.

Mechanistically incompatible. Tesla's forearm houses 25 linear actuators in concentric rings around a central rotary axis — all 25 actuators per arm relocated OUT of the hand. 3 cables per finger route through wrist router and guided phalanx channels. This enables hand mass < 300g (estimate) vs. servo-in-hand designs. The concentric ring patent (WO2024/072984A1) shows 23 linear + 1 rotary + 2 wrist linear. Commodity servo-in-joint designs cannot replicate this. Closest analogy: Shadow Robot Hand (tendon-driven) at $150K+.

Commodity EXCEEDS Tesla HW4 in raw TOPS (275 vs ~100–150). But AI4 is purpose-built for Tesla's vision pipeline with HW-accelerated softmax and FSD-specific memory architecture (384 GB/s bandwidth). AI4 runs Tesla's proprietary end-to-end NN — 48 co-trained sub-networks sharing learned representations across FSD and Optimus. Orin runs GR00T/Isaac ecosystem. Gen 3 transitions to AI5: 8× raw compute, 9× memory (est. 144 GB), 5× bandwidth — rivals NVIDIA H100 for inference.

Capacity match (2.4 vs 2.3 kWh). But Victron is ~19 kg vs Tesla 4680 pack at estimated 8–10 kg — roughly 2× mass penalty. 4680 cells enable cell-to-pack integration in the torso, providing structural function beyond energy storage. At 500W walking draw, 2.3 kWh gives ~4.6 hr theoretical runtime; Tesla targets 8 hr with efficiency optimizations. Gen 3 current challenge: battery life remains a key hardware limitation (joint motor overheating, transmission wear cited in 2025 production halt).

Tesla uses pure vision-only perception — 8 cameras trained on 8.2 billion vehicle miles of real-world data. No LiDAR, no radar. Open builds use depth cameras (RealSense D435i) to compensate for fewer cameras and less training data. Tesla's moat is not the camera hardware — it's the 8.2B miles of training data that gives the vision model unmatched environmental understanding. Camera hardware itself is replicable; the trained model is not.

Tesla's structural advantage is process, not material. Giga Press single-piece castings eliminate assembly labor by ~70%. PEEK polymer in joints saves 40% mass vs. metal alternatives. CF/PEEK limbs provide torsional stiffness at reduced weight. Harmonic drive modules weigh 1.2 kg each (vs 2.5 kg traditional RV). Commodity CNC frames can match material quality but require 10–50× more assembly steps and higher per-unit labor cost.