Soon that mobile phone in your pocket (or your hand) will become a low-cost, secure, AI powerhouse. When most people think of AI, they think of hyperscalers burning massive amounts of energy and money to do their magic. That will soon change as more of that work moves to your mobile phone, increasing your privacy, performance, and efficiency (think cheap). This will be accomplished by a combination of technologies I’ll explain below: knowledge distillation, small language models (SLMs), and ASICs. The company I believe is in the lead to benefit from this evolution is Qualcomm. Let me explain.
A process called knowledge distillation enables the huge LLMs, with up to 2 Trillion parameters to teach SLMs with a mere 10 Billion parameters (200-times smaller). These SLMs retain 70-90% of the smarts of the LLMs while running super fast and they can fit on your phone. In the tech business, standards are created by user counts, so the mobile phone will define the standard in AI. That standard will run an SLM.
DeepSeek introduced a Mixture of Experts (MoE) where queries are divided up and sent to smaller “experts”. Your mobile phone’s SLM will act like your general practitioner doctor, routing you to experts as needed. These experts will be LLMs running in the cloud. The local SLM might handle 70% of the work while routing the remaining 30% to specialized LLM experts. The mobile phone AI will then assemble the responses into a coherent single response. It essentially acts as an orchestrator for your AI needs, making it all work seamlessly.
In the early days of crypto, the GPU replaced the CPU as a better solution for high-performance crypto mining. Then along came Application Specific Integrated Circuits (ASICs) that were purpose-built for crypto mining providing faster performance and lower costs. In the AI world, training is evolving far too rapidly for ASICs, this requires high-end GPUs from the likes of NVIDIA. However, inference (responding to prompts) is stable and repetitive, ideal for ASICs. ASIC adoption with bring down costs for compute, energy and money. Expect to see AI-focused ASICs handling inference in mobile phones and laptops. An added benefit is they process your data locally assuaging any fear that the AI companies will be monitoring you or training on your data. So, just like the crypto mining business, ASICs will provide local inference on the SLM on your mobile phone.
ASICs for AI inference on your mobile phone will be integrated into what is called a System-on-Chip (SoC). No one's better positioned than Qualcomm (NASDAQ: QCOM) to lead the future of SoCs for mobile phones. With 40-50% share in premium Android SoCs, they've built unbreakable trust through decades of powering Samsung Galaxies, Google Pixels, and beyond. Their Hexagon NPUs are already SLM-optimized, blending CPU, GPU, and ASIC-like accelerators for seamless on-device AI. They have the soup-to-nuts solution, the market share, the technical leadership and the trust of the mobile device vendors. They have the pole position for this race to mobile AI dominance.
I’m not calling for NVIDIA to lose their iron grip on AI-training, this is more about market expansion as AI moves to the mobile phone. In fact, if I were NVIDIA, I would use their $5T capitalization to acquire Qualcomm in anticipation of the next wave of AI unfolding on the mobile device.
This thesis isn't just hype; it's backed by trends. SLM distillation makes AI accessible; ASICs accelerate low-power inference; Qualcomm has complete SoCs to execute in the mobile space. Expect explosions in apps like augmented reality assistants or health monitors, all powered from your pocket. If you're in tech or investing, watch QCOM—they're not just riding the wave; they're shaping it. I invest based on a thesis of the future. Sometimes it takes a while to play out, but I like this thesis, it rests on a confluence of enabling technologies that are all ripening at the same time. I’ll revisit this in 3 years and see if I was right.
Subscribe to our newsletter to receive the latest updates and promotions from MPH straight to your inbox.
