Qualcomm CEO Cristiano Amon used his Computex 2026 keynote on June 1 in Taipei to do something Qualcomm has not previously done at this scale: ship a robot brain in a box. The Dragonwing IQ10 Robotics Reference Design (RRD) is a single fully enclosed compute unit that combines the company’s latest Oryon CPU silicon, an NPU and GPU stack, 64GB of in-package memory, 512GB of UFS 4.0 storage, and native interfaces for 12 simultaneous camera streams, LiDAR, time-of-flight depth sensors and IMUs — and it is meant to plug into an industrial, autonomous-mobile or humanoid robot as the single piece of compute that everything else hangs off.
The number that matters is 700 TOPS of on-device AI performance, scalable to 2,000 TOPS by attaching dedicated hardware accelerator modules. The 18-core CPU is X Elite class — the same Oryon silicon Qualcomm is shipping in Snapdragon laptops, repackaged for industrial thermals. The reference design is rated for -40°C to +70°C operation, forced-air cooled, with PCIe Gen5, 10Gb Ethernet, Time-Sensitive Networking, and EtherCAT built in. Software stack is ROS2 on Ubuntu Linux with cloud fleet management through Qualcomm AI Hub. Early access starts this month. Global availability is targeted for September 2026.
Why a reference design and why now
The pitch is unsubtle. Until this week, a team trying to build a humanoid robot had to take a generic NVIDIA Jetson or a custom SoC and then bolt on a separate motion-control board, a separate camera-fusion ASIC, a separate sensor I/O bridge, and a separate networking subsystem. The patchwork created data latency, drove up costs, and added months to development cycles. Qualcomm is now selling the whole compute stack as one SKU that ingests sensors and emits motion control natively. “Reduces prototype-to-production cycles by months” is the line Amon used, and for once it is not marketing-only — the bill of materials list and the I/O density actually back it up.
The strategic frame matters more than the spec sheet. On the same morning, NVIDIA announced its Isaac GR00T reference humanoid for academic research, using Unitree’s H2 body, Sharpa hands, and NVIDIA compute. NVIDIA’s reference is aimed at universities. Qualcomm’s reference is aimed at companies that ship product. Same week, two different “here’s the standard humanoid platform” announcements, from the two companies that intend to own the compute layer of the next decade of robots.
What it means for the integrators
The list of US humanoid builders that previously had to design their own compute board is exactly the cohort Qualcomm is courting. Apptronik’s Apollo, Figure’s 03, Agility’s Digit, Boston Dynamics’ new electric Atlas, and the Chinese cohort — Unitree, AGIBOT, Skild — have each spent the last two years building roughly the same compute brick from different parts. Qualcomm is offering them a way to skip that step. Some will accept and ship faster; others will decline because the compute board is the moat they don’t want to give up.
The chip is also a direct response to the Taiwan supply chain story that landed at the same Computex. Hiwin’s pitch is mechanical: gears, bearings, actuators, harmonic drives — the Taiwanese components that have to live inside any humanoid that ships. Qualcomm’s IQ10 is the brain that has to live on top of those components. Between the two, a “Taiwan stack” for humanoids — TSMC silicon on a Qualcomm reference, with Hiwin mechanics underneath — becomes a more concrete proposition than it was two weeks ago.
The 2,000-TOPS scaling story
The headline 700 TOPS is enough for most current industrial robots. The interesting number is the 2,000 TOPS scaling path. Qualcomm is conceding, in its own spec sheet, that current LLM-driven manipulation and reasoning workloads will not fit in 700 TOPS once a robot has to do real semantic mapping, multi-step task planning, and natural-language interaction in the same loop. The accelerator-module path means a robot can ship with 700 TOPS today and 2,000 TOPS tomorrow without redesigning the chassis or the harness — which is the same modularity story automotive electronics solved a decade ago and that Qualcomm explicitly says it is reusing here.
The pricing tells
Qualcomm has not published a price. Early access is by invitation. Global commercial availability is September. The pattern is reference-design-first, volume-pricing-once-design-wins-stack-up. Watch the next two quarters: if Apptronik, Figure, or Agility publishes a teardown showing an IQ10 inside their next-revision board, the reference design has won. If they keep rolling their own, the integrators are betting their compute IP is worth more than Qualcomm’s reduced time-to-market.
For now, the most interesting line in Qualcomm’s release is the one that’s not about specs: “ROS2 support, cloud-connected fleet monitoring via Qualcomm AI Hub.” Fleet monitoring is what an industrial customer buys after the robot ships. Selling chips into individual robots is a one-time sale. Selling the cloud back-end that monitors a fleet of those robots — for the life of the robot, billed monthly — is the Qualcomm-as-platform move that the spec sheet only hints at. The IQ10 is the foot in the factory door. The recurring revenue is what comes through it.
The June 1 keynote was Qualcomm telling the humanoid industry that it doesn’t have to design its own compute layer anymore. The September shipment date is when the industry has to decide whether it agrees.
