Dragonwing IQ10 Explained [What It Changes]

Qualcomm's Dragonwing IQ10 is a deployment-ready "brain" for robots — a single processor delivering 700 TOPS of AI compute on-device so that an industrial machine or humanoid can see, plan, and reason locally instead of phoning the cloud for every decision.

That one sentence is the whole shift. Everything below unpacks why a chip announcement matters to anyone running an operation that touches physical work.

TL;DR: At Computex 2026, Qualcomm unveiled the Dragonwing IQ10 Robotics Reference Design (RRD), a full-stack package built around the IQ10 processor. According to Edge AI and Vision Alliance, the design delivers up to 700 TOPS, 18 Oryon CPU cores, and native support for 12 cameras. (report) The point is not raw speed — it is that perception, planning, and reasoning now run on the robot itself, with evaluation units seeding in June 2026 and global availability in September 2026.

What actually happened, in plain English

For years, a robot smart enough to handle a messy warehouse aisle or an unstructured factory cell needed a server rack of GPUs somewhere, with the robot streaming sensor data up and waiting on answers. That round-trip introduced latency, bandwidth cost, and a hard dependency on connectivity. The Dragonwing IQ10 collapses that loop onto a chip.

According to Edge AI and Vision Alliance, the Dragonwing IQ10 RRD delivers up to 700 TOPS of AI performance, 18 Qualcomm Oryon CPU cores, and native support for up to 12 GMSL2 cameras. (report) That is enough on-device horsepower to fuse a dozen camera feeds plus LiDAR and time-of-flight depth in real time — the sensing budget a robot needs to navigate a space built for humans.

According to PC-Tablet, the platform's baseline 700 TOPS scales to 2,000 TOPS via external expansion modules, and the in-package memory is 64GB LPDDR5x with 512GB of UFS 4.0 storage. (report) Those numbers matter because large vision-language-action models — the kind that let a robot follow a spoken instruction — are memory-hungry, and on-board memory is what keeps them from running in the cloud. The Dragonwing IQ10 delivers 700 TOPS on-device, scalable to 2,000 TOPS.

This is also a reference design, not just a chip. According to Edge AI and Vision Alliance, the RRD ships with an enterprise-grade Linux framework, native ROS2 support, and operating tolerance from -40 to 70 °C. (report) A robotics builder gets a near-complete starting point rather than a bag of parts.

Why now? The constraint that broke

The reason this lands in mid-2026 and not three years ago is that two constraints eased at once: on-device AI accelerators got big enough to run modern perception models, and "physical AI" software stacks (ROS2, simulation, VLA models) matured enough to be packaged. Qualcomm bundling them is the news.

According to PC-Tablet, the RRD includes low-latency on-device AI runtimes, native ROS2 support, and cloud-connected fleet monitoring through Qualcomm AI Hub — a stack that previously took a robotics team months to assemble. (report)

The market context explains the urgency. According to the International Federation of Robotics, the global average robot density stands at 132 units per 10,000 employees as of 2024 data published in April 2026, with the United States at 307. (report) The pressure to automate physical work is real and uneven. US robot density reached 307 units per 10,000 employees.

According to the International Federation of Robotics, South Korea leads the world at 1,220 robots per 10,000 employees while China reached 166 — figures that show how much headroom most economies still have. (report)

Dragonwing IQ10 at a glance

Figures per Edge AI and Vision Alliance and PC-Tablet.

Who shipped it — and who is building on it

Qualcomm announced the design; the ecosystem is what makes it real. According to PC-Tablet, early-access partners include NEURA Robotics, Advantech, and Thundercomm, with evaluation units going to enterprise customers in June 2026. (report)

According to Edge AI and Vision Alliance, the broader partner roster spans NEURA Robotics, Advantech, APLUX, Booster, Innodisk, MeiG, NEXCOM, Radxa, Thundercomm, and VinMotion — a 10-company list that signals a shared platform, not a single vendor's robot. (report)

Rollout timeline

Timeline per PC-Tablet.

The honest limit: a reference design is a starting point, not a finished robot. Builders still have to integrate it, train task-specific models, and certify the machine for their environment. Commercial availability in September 2026 means real fleets running this silicon are a 2027 story, not a this-quarter one.

Cloud robot vs on-device robot

The clearest way to grasp the change is to compare the old architecture to the new one. The shift is not just speed — it is where the decision happens and what the robot depends on.

On-device specs per Edge AI and Vision Alliance; the left column is the prior-generation architecture this replaces.

According to PC-Tablet, the design carries 64GB of in-package LPDDR5x memory and 512GB of UFS 4.0 storage. (report) That is enough to keep large models, maps, and logs resident on the robot so it never has to ask the cloud for its own brain. On-board memory hits 64GB LPDDR5x with 512GB storage.

How this connects to your software, not just your hardware

Most readers running an operation will never order a chip. The relevant question is what a smarter, locally-reasoning robot does to the software that surrounds it — the dispatch, exception-handling, and document workflows that decide whether a machine is useful.

A robot that reasons on-device still needs to be told what to do, hand off what it can't do, and log what it did. That orchestration layer is software, and it is where most of the operational value lives. Teams already routing tasks, exceptions, and documents through US Tech Automations workflows can treat a new on-device model as a swap at the edge, not a rebuild of the surrounding process.

Concretely, the value of a smarter robot is unlocked at three software seams: the instruction it receives, the exception it escalates, and the record it writes. Each of those is a workflow step, not a hardware feature. A firm that has already defined those steps — what triggers a human handoff, where the log lands, how a failure routes — gets to plug a more capable machine into an existing slot. A firm that hasn't ends up writing that orchestration under deadline pressure after the robot is already on site. The robots running on US Tech Automations workflows inherit the routing and audit logic the business already trusts, which is why the readiness work matters more than the silicon spec.

For the industry-specific picture, the spokes in this cluster go deeper: what Dragonwing IQ10 means for manufacturers covers the factory floor, what it means for logistics operators covers warehouses and yards, and what it means for home services companies covers field and residential work.

Signal vs Speculation

Everything above is sourced fact. This section is our forecast — read it as analysis, not as reported news.

Our read: the demonstrated facts are narrow and concrete. Qualcomm shipped a reference design with verified specs, real partners, and a real ship date. According to PC-Tablet, that includes a 700 TOPS baseline and a September 2026 availability window — those are facts, not projections. (report)

Our read on the next few years: if the partner ecosystem holds and pricing lands where reference designs usually do, the practical effect for small and mid-size operators is not robots-everywhere — it is that the robots arriving in your facility from vendors will be meaningfully more capable, and far less dependent on your network. The bottleneck shifts from "can the robot see?" to "is our process ready to delegate to it?"

Our read on the trap: the businesses that lose here are the ones that treat physical AI as a hardware purchase. The ones that win treat it as a workflow change — defining clean handoffs between machine, human, and software before the machine arrives. That readiness is buildable today, with or without a single robot on site. Given the US density gap the International Federation of Robotics measured at 307 against South Korea's 1,220, the room to grow is real, and the early movers will be the ones whose processes were ready first. (report)

Key Takeaways

• Dragonwing IQ10 moves robot perception, planning, and reasoning on-device (up to 700 TOPS locally, per Edge AI and Vision Alliance), ending the cloud round-trip.

• It is a full reference design (Linux + ROS2 + fleet monitoring), so builders start near the finish line, not from scratch.

• Evaluation units seed in June 2026 with global availability in September 2026 (PC-Tablet) — real fleets are a 2027 story.

• The operational value sits in the orchestration software around the robot, not the silicon itself.

• The smart move now is process readiness: defining clean machine-human-software handoffs before hardware lands.

Frequently Asked Questions

What is Dragonwing IQ10 in one sentence?

Dragonwing IQ10 is Qualcomm's robotics processor and reference design that runs AI perception and reasoning on the robot itself. According to Edge AI and Vision Alliance, it delivers up to 700 TOPS and supports 12 cameras natively. (report)

When can businesses actually buy it?

Evaluation units are seeding to enterprise customers now. According to PC-Tablet, global commercial availability is scheduled for September 2026, which means production robots using it reach facilities through 2027. (report)

How powerful is 700 TOPS, really?

It is enough to fuse a dozen camera feeds plus depth sensors and run modern perception models locally. According to PC-Tablet, the baseline 700 TOPS scales to 2,000 TOPS with external modules for larger workloads. (report)

Do I need to be a robotics company to care?

No. Any operator that will eventually receive vendor robots is affected. The robots arriving will be more capable and less network-dependent, which changes how you design the surrounding dispatch and exception workflows.

Who is building robots on this platform?

According to Edge AI and Vision Alliance, partners include NEURA Robotics, Advantech, Booster, Thundercomm, and several others — 10 named companies in total, signaling a shared platform. (report)

Does this replace cloud AI entirely?

No. It moves latency-sensitive perception and control on-device while cloud-connected fleet monitoring through Qualcomm AI Hub remains, according to PC-Tablet, in a design that scales to 2,000 TOPS. (report) It is a rebalancing of where compute happens, not the end of the cloud.

The practical next step

The chip is Qualcomm's problem. Your process is yours. The operators who benefit first from on-device robotics are the ones whose dispatch, exception-handling, and documentation already run as defined workflows — so a new machine plugs into a slot instead of forcing a redesign. If you want to see how that orchestration layer is built, explore the agentic workflow platform or browse the rest of the resource library for industry-specific playbooks.