Can a Smartphone and Automotive Giant Learn to Serve Thousands of Customers?

Key Highlights

• At Dragonwing IoT Day in San Diego, Qualcomm described moving its industrial and embedded IoT business from high-touch engagements with a few large OEMs to a standardized platform built on upstream Linux, Ubuntu, Debian and self-service developer access.
• Cognex presented its In-Sight 3900 vision system, built on Dragonwing, which inspects coffee capsules in about 30 milliseconds each, or roughly 2,000 parts per minute, a rate Cognex says its previous compute generation could not reach.
• Axon and Geotab described long-running engineering relationships with Qualcomm. Axon's newly announced Body 5 camera splits AI work between the device and the cloud, and Geotab, which claims more than 700 integrated partners of its own, has moved to joint roadmaps for devices that stay in vehicles a decade or more.
• Tosh Tambe, who leads Qualcomm's industrial manufacturing vertical, said autonomous mobile robots in factories typically do productive work only 40 to 45 percent of the time, and described plans to apply Qualcomm's automotive V2X, ultra-wideband and Bluetooth technologies to fleet coordination.
• A week after the event, Qualcomm reported fiscal fourth-quarter IoT revenue of $1.9 billion, with industrial, networking and robotics up 20 percent year over year, while handset revenue fell 26 percent.

The News

Qualcomm hosted its North America Dragonwing IoT Day at its new San Diego Customer Experience Center on October 28 and 29, with CEO Cristiano Amon and EVP Nakul Duggal laying out a reworked strategy for industrial IoT, embedded systems and robotics. The goal is to make Dragonwing silicon, software and tools reachable by thousands of developers and product companies without the bespoke engineering support that defined Qualcomm's smartphone and automotive businesses. Zebra, Cognex, Axon, Geotab, Skydio and PickNik presented production deployments, and Qualcomm's industrial automation team described a push from peripheral devices into core factory control, which suggests the model is already carrying commercial workloads. Duggal's keynote recap is available on Qualcomm's OnQ blog.

Analyst Take

Qualcomm used two days in San Diego to argue something most chip companies avoid saying out loud: the customer model that built its smartphone franchise does not easily scale into physical AI. Duggal described Snapdragon as a business where Qualcomm knows every customer and their families. Dragonwing is designed for a market where that kind of intimacy cannot work, so the company has reorganized around standard software, documentation and a partner bench that carries the message to customers. The numbers explain the urgency. In the fiscal fourth quarter, handset revenue fell 26 percent to $5.1 billion, while IoT revenue reached $1.9 billion, up 4 percent, led by 20 percent growth in industrial, networking and robotics. The bear case deserves a fair hearing. Horizontal IoT platforms have a long record of stalling in the long tail, where volumes are small, requirements are bespoke and support costs eat the margin. Asking a smartphone silicon vendor to serve thousands of small builders is a bit like asking a great white to start filter-feeding. The presenters in San Diego do not settle that question. Zebra, Cognex, Axon and Geotab are the scale end of the market, and they show the foundation can carry demanding production workloads. They are not the long tail. The better long-tail evidence is Arduino. Qualcomm said roughly 280,000 Uno Q boards have sold in about a year, each one a developer who started without a sales call. Boards are not design wins, and the conversion rate from bench to product remains unknown, but that is the right kind of number to watch.

What Was Announced

The platform has four layers. The silicon runs from entry-level parts up to the IQ10 robotics processor, which Qualcomm rates at up to 700 TOPS (350 dense TOPS at INT8) and supports until 2038. Duggal stressed that the industrial line inherits what the automotive business taught the company: operation from minus 40 to plus 125 degrees Celsius, safety-grade design, compliance and auditability, and supply commitments measured in decades. For an operator, that last point matters more than peak performance. Plant equipment and fleet hardware outlive the org charts that bought them.

The software layer may be the bigger shift. Qualcomm says it now treats upstream Linux, Ubuntu and Debian as standard alongside Android, with Qualcomm Linux as a common base across Dragonwing processors. Upstreaming means a customer's software investment is architected to outlive any single chip generation, which is the main fear industrial buyers have about mobile-heritage vendors.

The developer layer is built from acquisitions. Arduino provides the entry point, and its Ventuno Q board offers a path to a production module based on the IQ8. Edge Impulse handles data and model workflows, Foundries.io handles fleet lifecycle and updates, and Modular aims to give AI workloads a common runtime across CPUs, GPUs and NPUs. PickNik, whose MoveIt framework is widely used for robot-arm motion planning, fills what Duggal acknowledged was the robotics capability Qualcomm knew least: manipulation. Qualcomm also previewed agentic development tools that use MCP to act on host and target devices, so a developer can describe what to build instead of digging through SDKs.

The fourth layer is access. Duggal set an internal bar: a developer should be able to click a link, have a kit arrive and get running without calling Qualcomm. Qualcomm cited more than 200 hardware partners and 35 distribution partners to carry that model, and one executive openly acknowledged the industry joke that Qualcomm is hard to work with.

Market Analysis

The partner stories carried the argument better than any slide.

Zebra framed edge AI as cost control. Its customers worry about what cloud AI bills look like across 10,000 store associates. Its on-device proof-of-delivery model for a global courier handling about 15 million packages a day captures only the data the procedure requires, discards faces, and is designed to add one or two deliveries per driver per day. Cognex described a compute gap. In 2022 it had a new inspection algorithm and no processor that met its targets for speed, price, thermal budget and size. Four years later, the In-Sight 3900 runs on Dragonwing.

Axon's case is about battery math. Its body cameras have been built on Qualcomm silicon since the Body 3 generation. Body 5, announced October 24, offers what Axon calls 25 times the AI processing of earlier cameras and is designed around a hybrid split: capture, connectivity and location stay on the device, while contextual analysis for its Guardian escalation alerts runs in the cloud. Anyone who has run a field organization knows the device budget is set by the battery, not the spec sheet, and a camera that dies before the shift ends has failed. Skydio extends the same pattern to infrastructure, with docked drones that public agencies and utilities launch remotely, no pilot on site, for emergency response and asset inspection.

Geotab may be the most instructive partner on stage. It connects roughly 6 million commercial vehicles generating more than 100 billion data points a day. It has moved from an indirect relationship with Qualcomm to joint roadmaps, and it exposes its insights to enterprise AI assistants through MCP. Geotab also claims more than 700 integrated partners of its own. That pattern of one ecosystem sitting on another deserves attention. Each Geotab partner that builds on Geotab data is, a step removed, building on Dragonwing inference, so platform value appears to compound faster than Qualcomm's own partner count would suggest.

Tosh Tambe's session moved the story onto the factory floor. Tambe, five months in after seven years at Siemens, said AMRs typically do productive work only 40 to 45 percent of the time. In his telling they are individually intelligent but unaware of each other, so they slow at every blind corner and struggle to dock. Qualcomm's proposed fix borrows from the car business: V2X peer-to-peer links between robots, ultra-wideband for precise docking, and the Bluetooth technology behind its electronic shelf labels. He described work on a common control plane at the factory edge for discovering, updating and versioning AI models across equipment from many vendors, and said Qualcomm is moving from peripheral devices into PLCs and low-latency wireless for motion control. Rockwell Automation demonstrated a Qualcomm-based product at the event, a sign that a major US automation vendor is already putting Dragonwing into its roadmap. Tambe also pointed to Advantech's large share of the industrial PC market as evidence that customers want a neutral silicon supplier. He stated a real constraint plainly: automation must be deterministic while AI is probabilistic, so he expects collaborative robots and human oversight, not dark factories everywhere.

The platform still has gaps against established competitors. Vendors with GPU-centric robotics stacks hold a multi-year lead in simulation and synthetic-data tooling for training robots, and PickNik does not close that gap. Incumbent PLC and industrial PC suppliers own the control layer and its certification history, and Qualcomm is only now entering their roadmaps. At the motor and actuator tier, long-established microcontroller suppliers bring decades of design-ins and functional safety credentials. And at the low end of edge silicon, comparably priced, long-lifecycle parts from other merchant vendors are the variable we would watch most closely. Qualcomm's differentiation case appears to rest on combining connectivity, power efficiency and long-lived software rather than on leading every category.

The brand is also stretching. SpaceX's selection of Dragonwing radio-access platforms for a terrestrial 5G network takes Dragonwing well beyond IoT into carrier infrastructure, alongside the industrial edge in Qualcomm's fiscal 2029 target of about $8 billion for industrial, networking and robotics.

Looking Ahead

The key trend we'll be monitoring is whether the 20 percent growth in industrial, networking and robotics holds as handset revenue stays under pressure and Qualcomm guides combined automotive and IoT revenue roughly flat for the December quarter. The calendar offers several checkpoints. Qualcomm's Japan robotics center is set to open in early December, the IQ10 is entering its commercial ramp, and Axon has yet to announce general availability for Body 5 and Guardian. Tambe said the factory automation roadmap should look quite different within six months, and Embedded World in March will show how many Arduino developers and hardware partners have moved from boards to shipping products. The harder test is cultural. A self-service model only works if documentation, toolchains and distributor support hold up when no Qualcomm engineer is on the call. If the long tail starts converting while the scale operators keep expanding, we would read that as evidence the platform model is working.