Smart Utilities
Why Qualcomm Spent 18 Months Buying Software, Not Chips
Published by IOT San Jose Research & Editorial Team

- Why Training a Model Is the Easy Part
- Why a Chip Company Bought a Software Platform
- A Specific Example of the Deployment Gap
- What This Means for South Bay IoT Product Teams
- Implementation Advice for Embedding AI Into IoT Firmware
- Common Challenges in Embedded AI Development
- What's Next for Embedded AI Tooling
Qualcomm spent roughly eighteen months leading up to early 2026 acquiring five separate companies, Augentix, Arduino, Edge Impulse, FocusAI, and Foundries.io, to assemble a single coherent answer to a problem that has quietly determined whether most AI-for-IoT projects actually ship: getting a trained machine learning model to run reliably inside existing embedded firmware is a fundamentally different discipline than training the model in the first place. Edge Impulse, the San Jose company at the center of that acquisition spree, spent the six years before its purchase building exactly the tooling that discipline requires, and its story is a useful lens into why embedded AI for IoT has remained a harder, slower problem than the machine learning hype cycle generally acknowledges.
Why Training a Model Is the Easy Part
Most public conversation about AI treats model training as the hard problem and deployment as an afterthought, largely because that is how AI development works in cloud and mobile contexts, where compute and memory are relatively abundant. Embedded IoT development inverts that difficulty. A microcontroller running a real-time firmware stack has kilobytes, not gigabytes, of available memory, a power budget measured in milliwatts rather than watts, and a real-time operating system with strict scheduling requirements that cannot tolerate the kind of unpredictable execution time a poorly optimized AI inference call can introduce. Converting a model trained on a workstation into something that runs reliably inside that environment, correctly quantized, properly integrated with the RTOS scheduler, validated against real sensor data rather than a clean training dataset, is where most embedded AI projects actually stall.
Edge Impulse, founded in San Jose in 2019 by Zach Shelby and Jan Jongboom, built its platform specifically around closing that gap, letting embedded engineers collect real sensor data, train models, and deploy optimized inference code onto microcontrollers without needing deep machine learning research expertise on staff. The company's customer base, spanning industrial, healthcare, and consumer applications, reflected a market need that existed well before most large semiconductor and cloud companies took the category seriously enough to build competing tools internally.
The company also built a developer community around its Imagine conference, held in person at the Computer History Museum in nearby Mountain View, a detail that underscores how much of the embedded AI tooling ecosystem has grown up inside the same dense South Bay community that already supports the region's broader semiconductor and IoT hardware industry. Zach Shelby's own background, a former Arm executive and tinyML Foundation board member, reflects how tightly connected the region's embedded AI tooling, standards, and semiconductor communities actually are, with the same small set of people and organizations showing up repeatedly across what might look from the outside like separate companies and initiatives.
Why a Chip Company Bought a Software Platform
Qualcomm's acquisition pattern is worth examining closely, because it reveals what the chip industry has concluded about where the actual bottleneck in embedded AI sits. Buying Arduino addressed developer accessibility, since Arduino's hardware and software ecosystem reaches one of the largest embedded developer communities in the world. Buying Edge Impulse addressed the machine learning tooling and deployment pipeline specifically. Buying Foundries.io addressed secure, scalable firmware deployment and device management. None of these acquisitions were primarily about chip technology itself, Qualcomm already had that. They were about the software layer that determines whether a company with strong silicon can actually get developers to build working products on top of it.
That pattern reflects a broader lesson about embedded AI for IoT: raw chip performance, measured in TOPS or clock speed, is necessary but not sufficient. A chipset with an impressive on-paper NPU specification is only useful to a product team that can actually get a working model onto it, validated, and shipping reliably, which is precisely the gap tooling platforms like Edge Impulse were built to close.
A Specific Example of the Deployment Gap
Edge Impulse's collaboration with STMicroelectronics on the STM32N6 microcontroller illustrates what closing this gap actually looks like in practice. Rather than leaving developers to figure out independently how to fit a vision or audio model onto a new microcontroller's specific neural processing hardware, the two companies built official, tested support so that a model trained in Edge Impulse's platform could be deployed directly onto the STM32N6's NPU with a validated, documented path rather than ad hoc trial and error. That kind of vendor-to-vendor integration work, unglamorous and rarely covered in AI hype coverage, is precisely the layer that determines whether a product team building on a new chip can actually ship a working embedded AI feature within a reasonable development timeline, rather than spending months reverse-engineering an undocumented deployment pipeline themselves.
What This Means for South Bay IoT Product Teams
For hardware companies and product teams building embedded AI into IoT devices, the practical implication is that platform and tooling choice deserves at least as much evaluation effort as chip selection itself. A team that selects a technically excellent microcontroller but lacks a mature deployment pipeline for getting models onto it reliably will likely ship slower and with more field issues than a team that prioritizes a well-supported toolchain, even on marginally less capable hardware. This mirrors a pattern seen across the broader embedded operating systems landscape, where real-time operating systems like FreeRTOS and Zephyr continue gaining traction specifically because they offer a mature, predictable foundation that embedded AI tooling can build on top of reliably.
Implementation Advice for Embedding AI Into IoT Firmware
- Validate against real field data early, not just a clean training dataset, since sensor noise, mounting variation, and environmental conditions in actual deployment consistently degrade model performance in ways a lab dataset does not capture.
- Choose an RTOS with established AI/ML integration support rather than treating real-time operating system selection and AI tooling selection as separate decisions, since the two increasingly need to work together for deterministic, reliable inference timing.
- Budget realistic engineering time for quantization and compression, since achieving acceptable accuracy within a microcontroller's memory and compute budget typically requires meaningfully more iteration than training an unconstrained model.
- Plan for over-the-air model updates from the start, not as a later addition, since embedded AI models, unlike traditional firmware, often need retraining and redeployment as real-world data reveals edge cases the original training set missed.
Common Challenges in Embedded AI Development
| Challenge | Description |
|---|---|
| Memory and compute constraints | Fitting a useful model within a microcontroller's kilobyte-scale memory budget requires specialized compression expertise |
| Real-time scheduling conflicts | AI inference calls can introduce unpredictable execution time that conflicts with an RTOS's deterministic scheduling requirements |
| Field data versus training data mismatch | Models trained on clean datasets often underperform when deployed against noisy, real-world sensor conditions |
| Fragmented tooling across hardware vendors | Different microcontroller platforms often require different deployment toolchains, complicating cross-platform product strategies |
These challenges connect directly to the TinyML development constraints and multi-protocol chipset architecture covered elsewhere on this site, since embedded AI tooling, connectivity, and silicon selection increasingly need to be evaluated together rather than as independent decisions made by separate teams.
What's Next for Embedded AI Tooling
Qualcomm's consolidation of Edge Impulse, Arduino, and Foundries.io under one roof suggests the next phase of embedded AI development will increasingly happen through fewer, more integrated platforms rather than a fragmented landscape of point solutions each handling one piece of the pipeline. For South Bay IoT product teams, that consolidation cuts both ways: integrated tooling can meaningfully reduce development friction, but it also concentrates platform dependency risk in a smaller number of vendors, a tradeoff worth weighing explicitly when choosing which embedded AI toolchain to build a new product generation around.
