Three ways to play the on-device AI cycle, and the bear case against all of them

Hi {{first_name|Investor}}

Last time, I mapped the on-device AI shift: the four forces moving inference from the cloud to the device, and the supply chain being built underneath it.

Today I want to look at which companies are positioned to grow into this cycle, why each one has a different kind of advantage, and then I'm going to try to break the thesis. You deserve to hear the strongest arguments against it before you put any research time to work.

In 2017, I bought NVIDIA because I understood something most investors didn't. The moat wasn't the GPU. It was CUDA, the software layer that made switching away from NVIDIA nearly impossible for developers. Hardware plus software lock-in. That was the thesis, and it played out.

I've been looking for similar patterns in on-device AI, and ARM is one of the most interesting.

ARM doesn't manufacture or sell chips. Instead, it designs the fundamental blueprint that other companies license to build their own processors. Think of it as the common language that nearly every mobile chip speaks. Apple's A-series and M-series, Qualcomm's Snapdragon, MediaTek's Dimensity, Samsung's Exynos, Google's Tensor: every single one runs on ARM's architecture.

Here's why that matters for edge AI. Most of the AI-specific processors in phones, tablets, and smart devices are built right alongside ARM cores, plugged into ARM's software ecosystem. ARM doesn't need to win the chip race. It collects a royalty on nearly every chip that enters it. And those royalties are growing. The latest version of ARM's architecture (Armv9) charges meaningfully higher rates than the previous generation, so revenue is growing faster than the number of chips being shipped. Costs stay relatively flat because the architecture is already designed. That's a powerful economic setup.

The tension: ARM is expensive, and the market already knows this story. I'm still developing my conviction here, but the positioning is hard to ignore.

Almost every company designing chips for edge AI pays TSMC to actually build them. ARM, Qualcomm, Apple, MediaTek, smaller startups like Hailo: they all send their designs to TSMC's factories.

That gives TSMC something unusual. You don't need to guess which chip designer will win if you own the company that manufactures for all of them. Edge AI demand stacks on top of the cloud AI demand TSMC already serves, making it one of the broadest ways to participate in this theme.

Building cutting-edge chips is also extraordinarily hard to replicate. The capital, the engineering talent, and the decades of manufacturing know-how create a barrier that new competitors can't realistically clear anytime soon.

These two might be the most fascinating companies in this entire thesis, and the ones I go back and forth on the most.

Apple and Google are the only companies that control both the chip and the software for on-device AI. Apple is the clearest example: the Neural Engine, Core ML, and a tightly controlled developer ecosystem make it a true walled garden where hardware and software are designed to work together and nothing else. Google's approach is different. Android is open source, but Google has been steadily integrating vertically with its own Tensor chips paired with LiteRT. The result is that Google increasingly controls the full stack on its own Pixel devices while still licensing the broader Android ecosystem out. Both could become the unified platform edge AI is missing right now, just through very different models.

If either one builds the standard toolkit that developers rally around (the way NVIDIA's CUDA did for cloud AI), the rest of the supply chain accelerates around them. That's enormous potential leverage.

The catch: these are multi-trillion-dollar companies, so edge AI alone will never move their stock price in a meaningful way. And historically, both have been slow to roll out platform-level features at scale. The pieces are there, but the timeline is measured in years, not quarters. I watch them for signals, especially around developer adoption of LiteRT and Core ML, rather than for direct portfolio exposure.

The large names above are important to understand, but the most interesting research is happening one layer down. There are smaller companies with direct, concentrated exposure to the edge AI buildout that most investors have never heard of.

One example: a legacy touchscreen company pivoting hard into edge AI and IoT, with its core IoT product revenue up 53% year over year in its most recent quarter. Another is structurally similar to ARM but at a fraction of the size, licensing processor designs rather than selling finished chips. If the ARM royalty model appeals to you but you want a smaller name with more upside potential, it's worth knowing about.

These are the kinds of names I cover in the full report, along with the risks, the milestones to watch, and where each one sits in the supply chain. Premium members get all of it.