VI
Robot brains & edge AI compute
A humanoid is a robot wrapped around an AI computer; the perception-and-policy silicon comes from one dominant vendor and a thin second tier.
Robot brains & edge AI compute
Who owns the choke. A humanoid is, mechanically, a power supply and 30-odd actuators wrapped around an AI computer — and the perception-and-policy silicon that turns camera frames into joint torques inside a battery budget comes from one dominant vendor and a thin second tier. NVIDIA owns the top of it. Its Jetson Thor / Jetson AGX Thor T5000 puts a Blackwell GPU with 2,070 FP4 teraflops onto a robot, and the moat is not the TOPS number — it is software gravity. The training data, the simulation (Isaac Lab), and the policy models (Isaac GR00T) are CUDA-native, so the brain and the dev stack are sold as one bundle. At CES 2026 the named adopters read like a census of the field — Boston Dynamics, NEURA, AGIBOT, LG and Humanoid among them (Figure runs in-house Helix compute and is a Qualcomm-named partner) — and the open GR00T reference humanoid (Unitree H2 Plus body, Sharpa hands, Jetson Thor brain, shipping late 2026) hard-wires the next generation of academic roboticists to Jetson before they graduate.
The second tier is the interesting trade. Qualcomm is the credible challenger: at CES 2026 it launched the Dragonwing IQ10 Series — a premium robotics processor explicitly for humanoids and AMRs — and named Figure (platform architecture), Booster (K1 Geek) and VinMotion (Motion 2 on IQ9) as adopters. Qualcomm's pitch is performance-per-watt and a phone-scale supply chain. Below that, the choke fragments by function. Ambarella sells perception-per-watt: its CV72S edge-AI SoC anchors the e-con eRCP robotics computing platform for industrial AMRs and outdoor robots. Lattice plays the deterministic glue — low-power FPGAs for sensor-bridging and motor/motion control — and has wired itself to the leader, joining NVIDIA's Halos safety ecosystem and pairing CrossLink-NX FPGAs with NVIDIA's Holoscan Sensor Bridge for humanoid 3D-vision pipelines. The tell across all three is that the second tier increasingly sells into NVIDIA's stack rather than against it.
What breaks it. Two forces. First, vertical integration by the largest robot programs caps the merchant story. Tesla designs its own AI5 inference silicon — taped out, mass production 2027, with Optimus among its applications — and Figure runs its Helix policy on proprietary onboard chips. If the two highest-volume Western humanoids both bring the brain in-house, the merchant TAM shrinks to everyone else. Second, China. The reducer and actuator layers of this sector are already being undercut by Chinese suppliers (Tesla sources harmonic drives from Suzhou Green Harmonic), and Chinese SoC alternatives will follow the same cost-down playbook into domestic robot programs — which is precisely the volume NVIDIA and Qualcomm are counting on.
Be honest about size and earliness. This is a narrative-rich, revenue-thin chokepoint. NVIDIA does not break out robotics: it folds into a new Edge Computing segment that booked $6.4B in Q1 FY27 but bundles PCs, gaming, workstations, AI-RAN and automotive — robotics is a rounding line inside it, and management says the quarter was driven by Blackwell workstations, not robots. Qualcomm discloses no robotics revenue. Ambarella's robotics line is real but small and still drone/AMR-led, with total company revenue guided to roughly $105–111M for the July quarter. Lattice's robot exposure is design-win press releases, not yet a disclosed revenue segment. So the structure is: one balance-sheet fortress (NVDA) whose robot revenue is currently invisible, one optionality name (QCOM) where robotics is immaterial to the model, and two small-caps (AMBA, LSCC) that are levered call options on the buildout actually arriving. The risk is robot-washing — count only disclosed revenue or a named, dated design win, not a CES slide.
What forces the reprice. Three clocks inside twelve months. The August prints — NVDA Q2 FY27 on Aug 26 and Ambarella's late-summer quarter — are the cleanest reads on whether physical-AI silicon demand is inflecting versus still being a demo. The GR00T reference humanoid shipping from Unitree in late 2026 converts the Jetson developer base into deployed units. And CES 2027 (Jan 6–9), where NVIDIA has said it will reveal its next computing platform, is where the merchant-vs-vertical-integration question gets re-litigated with new humanoid unveils. None of this is investment advice; it is a map of where the silicon bottleneck sits and which dates move it.
Who owns the choke
Ambarella, Inc.
Its CV72S edge-AI SoC anchors e-con Systems' robotics computing platform (eRCP) for industrial AMRs and outdoor robots, announced at CES 2024 (now ~2.5 years old). Total company revenue is guided to roughly $105–111M for the quarter ending July 31, 2026; robotics is a growth line, not yet separately sized.
- Bull
- Ambarella's CV72S edge-AI SoC anchors a shipping robotics computing platform for AMRs, making it a levered small-cap call option on the physical-AI buildout.
- Bear
- The eRCP design win is ~2.5 years old, robotics is unsized inside ~$105-111M quarterly revenue, and Chinese SoC cost-down threatens the second tier.
Figure AI Inc.
US humanoid developer named by Qualcomm as a Dragonwing platform-architecture collaborator, yet Figure runs its Helix AI policy on proprietary onboard chips with all inference performed onboard (Helix-02 demos, May 2026; Figure 03 introduced Oct 2025). Illustrates how the largest Western humanoid programs partly bring the brain in-house.
- Bull
- Figure runs its Helix AI policy on proprietary onboard chips with Helix-02 demos logging full autonomous shifts, illustrating a vertically integrated Western humanoid leader controlling its brain.
- Bear
- Private and pre-revenue; in-housing its compute caps merchant silicon TAM, and listed access is only via ARKVX and NVDA minority stakes.
KraneShares Global Humanoid Robotics and Physical AI Index ETF
First US-listed humanoid-robotics ETF (NASDAQ, launched ~June 5, 2025), tracking ~50–58 companies across robot manufacturing, semiconductors, actuators and sensing. Target equal-weight positions (NVIDIA/Tesla ~2.5% each) drift between quarterly rebalances; expense ratio ~0.69% after waiver.
- Bull
- KraneShares' KOID was the first US-listed humanoid-robotics ETF, an equal-weight ~50-58 name basket spanning manufacturing, semiconductors, actuators and sensing across the supply chain.
- Bear
- Target equal weights drift between quarterly rebalances, the ~0.69% expense ratio relies on a waiver, and it captures only listed names, not private makers.
Lattice Semiconductor Corporation
Joined NVIDIA's Halos AI-systems safety ecosystem (Mar 16, 2026) and pairs low-power CrossLink-NX FPGAs with NVIDIA's Holoscan Sensor Bridge for humanoid and robotic 3D-vision pipelines; a collaboration with Airy3D targets next-gen humanoid sensor fusion (Embedded World demo). No robot-specific revenue is disclosed.
- Bull
- Lattice's low-power CrossLink-NX FPGAs bridge sensors in NVIDIA's Halos and Holoscan humanoid 3D-vision pipelines, wiring it into the dominant stack as deterministic glue.
- Bear
- Robot exposure is design-win press releases with no disclosed robot-specific revenue, and the second tier increasingly sells into NVIDIA rather than capturing value itself.
NVIDIA Corporation
Isaac Sim/Isaac Lab plus Cosmos world-foundation models and GR00T humanoid models form the reference sim-to-real stack; GR00T N1.7 is in early access under commercial licensing and Isaac Lab 3.0 ships on the new Newton physics engine 1.0 (co-developed with Google DeepMind). Named users include 1X, Agility, Boston Dynamics, Figure, AGIBOT and NEURA. Disclosed Automotive and Robotics revenue was $567M in Q1 FY26 (~1.3% of $44.1B), now folded into a broader Edge Computing platform.
- Bull
- NVIDIA's Isaac Sim/Lab, Cosmos, GR00T and Jetson Thor form the sim-to-real reference stack nearly every named humanoid program trains on, monetized through GPUs and model licensing.
- Bear
- Robotics is invisible — folded into a broad Edge Computing platform driven by Blackwell workstations — so it is a free option on a data-center stock.
QUALCOMM Incorporated
At CES 2026 launched the Dragonwing IQ10 Series, a premium robotics processor for humanoids and advanced AMRs, and named Figure (platform architecture), Booster (K1 Geek humanoid) and VinMotion (Motion 2, on the IQ9 Series) as adopters. No robotics revenue is broken out.
- Bull
- Qualcomm's Dragonwing IQ10 robotics processor for humanoids, with named adopters Figure, Booster and VinMotion, makes it the credible second-tier challenger on performance-per-watt.
- Bear
- No robotics revenue is broken out and it is immaterial to the model; even named partner Figure runs inference on its own proprietary onboard chips.
Catalyst calendar
- 2026-08-26Ambarella Q2 FY2027 earnings (quarter ended July 31, 2026)mediumPure-play tell on whether robotics/AMR edge-AI SoC revenue is scaling beyond drones for the choke's second tier.
- 2026-08-26NVIDIA Q2 FY2027 earningsmediumCleanest read on whether physical-AI/robot silicon demand is inflecting inside the Edge Computing segment versus still demo-stage.
- 2026-10-01NVIDIA Isaac GR00T reference humanoid ships from Unitree (late 2026)mediumConverts the Jetson Thor developer base into deployed humanoid units, hard-wiring NVIDIA compute into the next research generation.
- 2027-01-06CES 2027 (Jan 6-9, Las Vegas)lowNVIDIA has said it will reveal its next computing platform; the venue where the merchant-vs-vertical-integration question for robot brains gets re-litigated with new humanoid unveils.