Qualcomm's Second Act: How Snapdragon Became the Most Important AI Chip Outside Data Centers

In May 2023, Qualcomm demonstrated Llama 7B running entirely on a Snapdragon 8 Gen 2 smartphone chip at 10 tokens per second. The demonstration was not particularly polished — it required a specific quantized model format, showed on a development handset rather than a shipping product, and the token generation speed was slower than most users would accept for a production chat application. But the demonstration proved something that had been theoretical: a consumer smartphone chip could run a capable language model locally, without a network connection.

Qualcomm has been building toward this moment since Paul Jacobs, who served as CEO from 2005 to 2014, made the decision to invest in a dedicated neural processing unit rather than relying on the GPU for machine learning inference. The Hexagon DSP, introduced in 2012 for audio and sensor processing, evolved through successive generations to become the Hexagon NPU — a specialized inference accelerator that now ships in every Snapdragon 8 series chip and claims 45 TOPS in the 8 Gen 3.

Understanding why Qualcomm is positioned as it is in the edge AI era requires understanding the architectural decision made fifteen years ago and why it was not obvious at the time.

The Hexagon Gamble and Why It Was Not Obvious

In 2012, building a dedicated processor for machine learning in a smartphone chip was not a response to the AI moment we are currently experiencing. It was a response to a much more prosaic problem: battery life. Always-on sensor processing — accelerometers, microphones for voice activation, GPS — was consuming meaningful power on the application processor because those tasks had to wake the main CPU to handle sensor data. A low-power co-processor that could handle sensor fusion and audio processing without waking the application processor would extend standby battery life measurably.

Related reading

The Hexagon DSP was designed for this: running always-on, low-power sensor processing. It was not designed for neural network inference because neural network inference at scale did not exist on mobile devices in 2012. But the architectural characteristics that made it useful for sensor fusion — optimized for matrix operations, designed for minimal power draw, capable of running while the application processor is asleep — turned out to be exactly the right characteristics for neural network inference.

When AlexNet in 2012 demonstrated that deep neural networks could be trained practically, and when mobile applications began to incorporate rudimentary ML features (image recognition, language translation), Qualcomm had an existing piece of silicon that could be evolved for these tasks. The evolution was not trivial, but the starting point was already there. Apple, building the A11 Bionic in 2017, designed its Neural Engine from scratch with similar architectural principles. Both companies were solving the same engineering problem from different starting points.

The Competitive Landscape in Android AI Chips

Qualcomm does not have the Android AI chip market to itself. Samsung’s Exynos chips include a Mali GPU and Samsung’s own NPU. MediaTek’s Dimensity series, which captures a significant share of mid-range Android devices, includes the APU (AI Processing Unit) architecture that competes directly with Hexagon on performance-per-dollar. HiSilicon’s Kirin chips, used in Huawei devices before U.S. export controls eliminated Qualcomm’s supply, incorporated their own Ascend Nano AI accelerators.

But Qualcomm’s position is structurally stronger than these alternatives for a reason that has nothing to do with transistor counts: developer tools. The Qualcomm AI Engine SDK, the Qualcomm Neural Processing SDK, and the integration with Android’s NNAPI (Neural Network API) have been maintained and updated for over a decade. When an Android developer wants to run a neural network model efficiently on Snapdragon, there is a well-documented, well-tested path. For MediaTek or Samsung Exynos, the equivalent tooling is younger, less consistent, and less integrated with the broader Android development ecosystem.

This is the same moat that Nvidia has in the data center — not just hardware performance, but the accumulated investment in software infrastructure that makes the hardware usable. Qualcomm has built this moat in mobile through consistent investment in developer tools while competitors focused more narrowly on chip performance metrics.

The Windows on Arm Bet

Qualcomm’s second act is not limited to smartphones. The Snapdragon X Elite, launched in 2024, is the PC chip that carries Qualcomm’s most serious challenge to Intel and AMD in the laptop market. The chip includes a 45 TOPS NPU specifically designed for Microsoft’s Copilot+ PC requirements — Windows AI features including real-time translation, on-device image generation via Windows Studio Effects, and the controversial Recall feature (screenshot history with semantic search).

The competitive context is revealing. Intel and AMD both scrambled to add NPU blocks to their laptop chips after Microsoft announced the 40+ TOPS NPU requirement for Copilot+ PC certification. Intel’s Meteor Lake and AMD’s Ryzen AI 300 both include dedicated NPUs. But Qualcomm shipped first, with higher NPU performance, because Qualcomm had been building mobile NPUs for a decade and had significant architectural head start.

Related reading

The Snapdragon X Elite’s battery life advantage over equivalent Intel chips is real and substantial — in comparable workloads, Snapdragon X Elite laptops achieve 20-30% longer battery life. This advantage comes partly from the Arm instruction set architecture (which requires less complex decoder circuitry than x86) and partly from the Hexagon NPU handling AI workloads more efficiently than routing them through the CPU. When Windows Studio Effects processes your video conference background in real time, doing it on the Hexagon NPU rather than the CPU means the CPUs are free for your actual applications.

Where Qualcomm Cannot Win

Qualcomm’s position has real structural limitations that the most enthusiastic coverage ignores.

The company does not control its own silicon fabrication. Snapdragon chips are manufactured by TSMC and (for some versions) Samsung Foundry. This means Qualcomm’s chip performance is bounded by its foundry partners’ process capabilities. Apple, which also uses TSMC but has a closer relationship and reportedly reserves more leading-edge capacity, consistently shows that the same TSMC node produces better results in Apple Silicon than in Snapdragon chips. The A17 Pro and Snapdragon 8 Gen 3 both use TSMC N3 process, but the A17 Pro demonstrates meaningfully better power efficiency. Part of this is Apple’s architectural advantage; part is believed to be better TSMC process customization.

Qualcomm also cannot run the largest models that matter commercially. A 70B parameter model requires 40GB of memory in INT4 quantization. No Snapdragon chip has 40GB of on-chip memory, and the UFS 4.0 flash storage that phones use for model weight loading does not approach the bandwidth needed for real-time inference on models of that size. The practical ceiling for on-device Snapdragon inference is around 7B parameters with aggressive quantization — useful for many tasks, but not representative of frontier model quality.

This means Qualcomm’s edge inference position is complementary to, not competitive with, cloud inference infrastructure. The typical hybrid architecture — handle simple queries on-device with a 7B model, escalate complex queries to a 70B+ cloud model — suits Qualcomm’s hardware perfectly. It also means Qualcomm’s success in edge AI is partially dependent on cloud AI remaining expensive enough that the on-device alternative is attractive.

The Automotive and IoT Extension

Qualcomm has been explicit that automotive and industrial IoT are the two markets where it expects Snapdragon-derived AI chips to grow most significantly over the next decade. The Snapdragon Ride platform — the automotive adaptation mentioned earlier in the context of Tesla and Mobileye — is not a separate chip family but an evolution of the Snapdragon mobile architecture with automotive-grade reliability certifications.

This vertical extension strategy is elegant for the same reason the Hexagon DSP bet was elegant: Qualcomm amortizes its chip R&D across hundreds of millions of smartphone units and then deploys adapted versions into automotive and industrial markets at incremental development cost. An automotive OEM that wants to source AI chips for their infotainment and advanced driver assistance systems can choose between Qualcomm’s automotive-grade Snapdragon (leveraging the same developer tools used by Android app developers) or custom chips from Mobileye or Nvidia that require entirely different tooling.

The developer ecosystem advantage is not subtle. Volkswagen has announced vehicles using Snapdragon Ride, in part because the software platform allows more reuse of consumer Android development tooling. When a software engineer at a tier-one automotive supplier can deploy a neural network using the same Qualcomm SDK they use for Android development, the integration effort drops by months.

Related reading

The Long Game

Qualcomm’s story in the AI era is a classic example of prepared opportunism. The company built capabilities over fifteen years for reasons that had nothing to do with the AI moment we are living through, and those capabilities turned out to be exactly what the moment required. That is not luck — it is what happens when you correctly identify that computing is moving toward specialized processing and you build infrastructure for that transition before it is obvious.

The company that comes out of the AI hardware era as the dominant edge inference platform will be the company that combines hardware performance with software infrastructure and developer trust. Qualcomm has all three, at scale, in mobile. Extending that position to automotive and PC is not guaranteed, but the foundation is stronger than most coverage acknowledges.

The AI chip story is not just about who builds the fastest data center accelerator. It is about who builds the infrastructure for the billion devices that will run AI locally, without data centers, in places where no cloud connection is available or desirable. Qualcomm built that infrastructure before anyone knew it would be needed.