You used edge AI three times today. You probably didn't notice.

Hi {{first_name|Investor}} -

You probably used AI on your phone today without thinking about it.

When you took a photo and your phone instantly sharpened the faces and fixed the lighting, that was AI running on a chip inside your phone. No cloud upload. Processed locally, in milliseconds, on dedicated silicon that's vastly more capable than what phones had even five years ago.

Your car nudging you back into your lane or braking before you saw the pedestrian? That was an AI chip inside the vehicle making a real-time safety decision, because a braking system that has to call a server first is a braking system that doesn't work.

And when your Ring camera told you a person was at the door instead of just "motion detected," newer models run that recognition directly on the device, cutting cloud costs and keeping your data off someone else's server.

These are shipping products, not future demos. In most modern devices, these tasks run on purpose-built AI accelerators (often called Neural Processing Units, or NPUs). This is edge AI, and it's creating an investment opportunity most people haven't noticed yet because they're still watching the cloud.

For three years, the AI investment story has been about one thing: training. Bigger models, more GPUs, more power. That story made NVIDIA the most important company in tech and rewarded investors who understood the moat underneath it.

But training is only half the equation, and increasingly the smaller half. The other side is inference: actually running AI models to do useful things like recognize images, translate speech, and make decisions. Training builds the model once. Inference runs it millions of times a day, across every user and every device. By most estimates, inference accounts for 80 to 90 percent of what it costs to operate an AI model over its lifetime.

A growing share of that inference is leaving the cloud and moving to the device itself. Your phone, your car, your doorbell camera, factory sensors, medical instruments. The edge AI market is projected to grow from roughly $25 billion today to more than $118 billion by 2033.

This is already happening. Every major smartphone chipmaker (Qualcomm, MediaTek, Apple, Samsung, Google) includes a dedicated NPU in their flagship processors. AI-enabled smartphones have crossed an estimated 30% of global shipments, and AI PCs with built-in NPUs are forecast to exceed 40% of PC shipments in the coming years.

The hardware is here. The shift is underway. The question is where the investment opportunity sits.

The move to on-device AI isn't driven by one factor. Four forces are pushing in the same direction, and they reinforce each other.

Speed matters more than convenience. Sending data to the cloud and back adds a delay, anywhere from tens to hundreds of milliseconds depending on conditions. For a chatbot, that's fine. For a car making a steering decision, a surgical robot adjusting mid-procedure, or a factory sensor catching a safety event, those milliseconds are the difference between working and failing. The applications that need AI most urgently are the ones that can't afford to wait.

Privacy is becoming law. When AI runs on the device, your data never leaves it. No transmission, no cloud storage, no third-party access. This isn't just a consumer preference anymore. GDPR in Europe, the EU AI Act, DORA for financial services, and a wave of new U.S. state privacy laws all increase pressure to keep data local. Companies that can process on-device are better positioned for where regulation is heading.

Cloud costs add up fast. Every time an AI model runs in the cloud, there's a cost: compute time, bandwidth, data transfer fees. Across millions of users, those costs compound quickly. With edge AI, the chip is already in the device and the model is already loaded. No per-query fee. In some modeled deployments, a hybrid approach has shown roughly 75% cost savings and more than 60% energy savings compared to running everything in the cloud.

Some places don't have reliable internet. Factory floors, remote infrastructure, agricultural equipment, offshore rigs, vehicles in transit, medical devices in rural clinics. These environments need AI that works whether or not there's a network connection. On-device processing is the only approach that doesn't break when connectivity does.

Every example I opened with exists because of at least two of these forces working together. That overlap is what makes this a lasting shift, not a temporary trend.

If you remember the smartphone era, Apple got the headlines, but enormous value was also created in chip architecture (ARM), wireless modems (Qualcomm), display glass (Corning), and manufacturing (TSMC). On-device AI has the same structure: companies collecting royalties regardless of which chip wins, foundries manufacturing everyone's silicon, chip designers building for phones and cars and IoT, and a wave of startups designing purpose-built edge AI hardware.

Then there's the layer I keep thinking about to complete the moat: the software layer. There's no standard developer toolkit for edge AI yet, nothing that unifies the ecosystem the way NVIDIA's CUDA did for cloud training. Whoever builds it could capture more long-term value than any chip company. I want to be upfront: I'm actively researching this layer, and I don't have a clear answer yet on who the winner will be or when it will consolidate. The companies, the watchlist, the way I think about where each one fits: all of it will develop and evolve as the market does. This is the beginning of a thesis I plan to keep refining, not a final verdict.

In my next email, I'm looking at which companies are positioned to grow into this cycle, the advantages that could make them formidable as the market develops, and the strongest arguments for why this thesis could be wrong.

I'm launching the premium tier of Reset and Invest, and early subscribers lock in Founder pricing for life. That means $12/mo or $120/yr instead of $24/mo or $250/yr, and it stays that way as long as you're subscribed. This rate goes away on May 25.

Founder members get every weekly deep dive unlocked, plus access to research reports like the 25-page one I’m releasing at the end of this email. When a topic warrants going deeper than the newsletter can, I publish a full report covering the thesis, the data, and the risks in detail. Those are only available to paid members.

You also get the full archive of every issue and report I've published, plus my upcoming e-book Make Your Own Alpha, a complete walkthrough of my investing strategy, research process, and the framework behind everything I write.

LOCK IN FOUNDER PRICING → CLICK HERE

Stay disciplined - Koh