Surf's up- Catch the AI Smartphone Wave
Technological breakthroughs are set to ignite growth in AI consumer devices.
Artificial intelligence and machine learning are already woven into the fabric of our lives. Although large language models (LLMs) and generative AI have significantly expanded AI's capabilities, the increasing integration of AI into our lives is not new. This trend has been ongoing for years, even before ChatGPT's arrival, powering the apps on our smartphones, the software we use at work, and countless other aspects of our daily routines.
This context highlights why the widely anticipated AI-driven smartphone upgrade cycle hasn't materialized. Our smartphones already offer exceptional utility for both work and personal use. Currently, there simply isn't a compelling "killer app" for generative AI that motivates consumers to upgrade their devices.
However, just as investor attention has progressed from GPUs to AI networking and most recently to custom AI chips, the next wave of AI to garner investor interest is AI consumer devices, particularly smartphones.
Most new AI consumer apps have been "nice-to-haves" rather than "must-haves." They often enhance existing features rather than offering fundamentally new capabilities. This incremental improvement has not been enough to drive a significant upgrade cycle.
However, several technical innovations addressing power consumption, memory and storage limitations, and the use of larger on-device models are poised to contribute to stronger-than-anticipated smartphone sales in the second half 2025 and throughout 2026. These advancements will create investment opportunities across the smartphone hardware and semiconductor supply chain.
Apple and DeepSeek have developed groundbreaking techniques that will make larger AI models more viable for on-device processing and facilitate cheaper, more efficient inference. These innovations will reduce the computational load on devices and increase processing efficiency to reduce memory demands, which is of particular importance given the the amount of memory require for inference. Furthermore, ongoing improvements in algorithm efficiency and compute power will continue to drive down the cost of AI inference.
This rapid decline in the cost of AI inference has led many to refer to the Jevons Paradox, an economic principle suggesting that as technology increases the efficiency of using a resource, consumption of that resource often increases rather than decreases.
In the context of AI, this means that even as AI models become more efficient, the overall demand for computing power will continue to rise. This is because more efficient AI will enable a wider range of applications and use cases, driving increased adoption and usage.
To fully utilize the latest advancements in AI, most consumers will need to upgrade their smartphones. While current devices can handle some small-scale AI models, larger and more advanced models will likely require new hardware with upgraded processors, memory, and storage.
Stronger than expected end market demand could set the stage for AI device leveraged stocks to outperform already highly valued AI compute infrastructure plays. Memory stocks, in particular, would be prime beneficiaries as stronger consumer device demand and signs of positive price elasticity in these end markets are essential for memory stocks to recover from their current mid-cycle correction.
Technology Breakthroughs: Enabling On-Device AI
Technical breakthroughs often ignite growth in the tech hardware market, and are crucial for an AI-driven smartphone upgrade cycle. Apple and DeepSeek have recently made significant strides in this area.
Apple's Innovation: Optimizing Memory Management
Apple's new technique overcomes the limitations of "on-device" LLMs, enabling them to run larger AI models on devices with limited memory. This is important because larger models, while capable of powering more sophisticated AI features, typically require a lot of DRAM (dynamic random-access memory) to operate efficiently. The technique allows for effective inference of LLMs on consumer devices with limited memory by storing model parameters in the flash storage and loading them into DRAM as needed.
This approach utilizes two techniques to drive more efficient and productive use of on-device memory:
“Windowing”: This method reduces the amount of data transferred from flash storage to DRAM by strategically loading only the essential parts of the model into the faster memory when needed.
Row-column bundling: This technique optimizes how data is accessed from flash storage, making the process faster and more efficient.
This approach focuses on optimizing memory management rather than compute resources, enabling devices to run AI models up to twice the size of the available DRAM.
DeepSeek's Innovation: Enhancing Model Efficiency
In DeepSeek’s V3 and R1 models, it has built on and further optimized existing innovations such as Mixture of Experts (MoE) and Multi-head latent Attention (MLA) to support greater model efficiency and lower inference cost.
Mixture of Experts: This technique splits a larger neural network into multiple "expert" subnetworks and dynamically selects the most relevant expert for each task. This increases model capacity, reduces compute cost, and improves specialization.
Multi-head latent attention: This technique reduces memory usage, allowing for efficient processing of long sequences by compressing key-value pairs, which reduces memory usage during inference.
DeepSeek's model-level improvements are distinct but potentially complementary to Apple's DRAM efficiency techniques. Together, they could enable significant on-device efficiency gains. The common element is that both approaches target reducing the computational resources needed without sacrificing performance.
However, there could be trade-offs when combining these techniques due to increased complexity that could impact performance. Integrating these techniques would likely require additional fine-tuning and optimization work. Nevertheless, these innovations are set to support the use of more robust, larger on-device models, whether on a standalone basis or combined.
These are significant breakthroughs that will enable more sophisticated AI applications on smartphones and other consumer devices. Bigger LLMs on device could enable enhanced features such as real-time language translation, more natural and context-aware virtual assistants, personalized content recommendations, and improved chatbot interactions.
Although recent progress in consumer generative AI has focused on text and image generation, as well as advanced search and answer engines, more powerful on-device AI could drive stronger smartphone sales as consumers seek out devices with more advanced capabilities.
Overcoming recent concerns of high inference cost
While OpenAI's new frontier model, o3, has generated excitement among AI enthusiasts, its reliance on test-time scaling and Chain-of-thought (CoT) reasoning increases computational intensity and cost. This presents a challenge, as AI applications must be economically viable to gain traction. This is where o3 falls short as the increase in compute intensity comes at an increased cost to answer a query.
However, DeepSeek's innovations have addressed this concern with its breakthroughs translating into lower inference costs, which should support increased adoption of AI applications. Although it appears that DeepSeek had access to significantly greater compute resources than initially reported, they were still likely resource-constrained compared to Western peers. This serves as a good example of how limited access to resources can stimulate innovation.
AI Application Demand Inflection on the Horizon
While cloud providers and enterprises have discussed various AI use cases, broad-based demand has yet to materialize. Consumer and enterprise demand has been lackluster relative to more bullish expectations, with monetization primarily focused on cost savings rather than revenue generation. However, this is poised to change as AI inference costs decline and application development matures.
A wave of investment in AI application and agent startups is underway. While many will fail, this "spray and pray" approach to investment and experimentation should eventually yield a more compelling range of AI consumer apps that will support a more robust smartphone upgrade cycle.
DeepSeek's release of its models V3 and R1 has called attention to the benefits of smaller models for both enterprise and consumer AI applications. Model distillation compresses a large model by training a smaller model to mimic it, reducing the model size and, consequently, the compute resources needed, which translates into lower inference costs while maintaining performance.
Databricks CEO Ali Ghodsi has emphasized that the next stage of AI innovation lies in building applications on top of existing models and utilizing smaller, more specialized models. This approach can lead to faster adoption in both consumer and enterprise markets. Smaller, more specialized models are easier to deploy and integrate into existing workflows.
Enterprise Adoption
Enterprise adoption of advanced AI applications may indirectly contribute to improved smartphone sales throughout 2025 and 2026. Employees will likely need new smartphones with the necessary capabilities and compatibility to utilize new, more advanced AI-powered applications and productivity tools, as well as edge AI in field operations.
While the widespread adoption of AI tools by businesses has been gradual, it represents a significant long-term growth opportunity. As AI systems improve and address concerns about reliability and accuracy, businesses will increasingly integrate AI applications into their workflows. First in low risk applications such personal productivity and team based workflow automation, before progressing to more complex tasks as systems improve.
Technical innovation has historically driven productivity gains and economic growth, and artificial intelligence is no different. Whether by boosting worker productivity or impacting employee headcount, AI will drive investments in applications that offer positive returns on investment.
Where to Invest in the AI-Driven Smartphone Sales Recovery
The anticipated improvement in smartphone sales presents a compelling opportunity for investment in the hardware supply chain and semiconductor companies leveraged to smartphone sales.
Some of these companies offer a more attractive risk-reward profile than many "AI-related plays," which often have high valuations and lofty expectations. Memory providers, in particular, stand to benefit significantly from stronger smartphone sales.
Apple, will remain front and center of any improvement in global smartphone sales, and is a prime beneficiary of the discussed technical innovations. However, investors seeking greater upside potential may want to look beyond Apple to the companies in the broader smartphone supply chain.
Below are a select group of companies that will benefit from stronger AI smartphone growth -
TSMC - Smartphones account for 35% of revenue.
Broadcom - Dominant position in RF front-end components and Wi-Fi and Bluetooth chips.
Micron and Western Digital - To support more advance AI features, more DRAM is needed for AI processing and flash storage for models and larger media files.
Silicon Motion - Increased demand for NAND flash controllers.
Qualcomm - In the past as one of the more levered semiconductor stocks to smartphone growth the company would be top of this list. However, in the near to medium term the company faces the headwind of Apple’s transition to using its own in-house modem.