애플의 우연한 해자: AI 패배자가 최종 승자가 되는 법

@adlrocha - How the "AI Loser" may end up winning Apple’s accidental moat adlrocha Apr 12, 2026 1 Share A few weeks ago I wrote about how I thought intelligence is becoming a commodity . The idea is quite straightforward, and widespread now: when everyone races to build the best model, the models get better, but so does every other model eventually. Every dollar spent on a bigger training run makes the previous one cheaper. The distance between frontier, second-best, and open-source alternatives is collapsing fast (actually Gemma4, Kimi K2.5 and GLM 5.1 are becoming my bedside models these days) . Even more, as models become better, the unit of intelligence that can be deployed in local hardware with lower hardware capabilities increases significantly. The irony of this situation is that this commoditisation of intelligence is benefiting the company that everyone was framing as the “AI loser”: Apple The company that “lost” There’s a version of the last three years where Apple genuinely failed at AI. They had Siri before anyone had a serious voice assistant, and then watched how ChatGPT ate their lunch already since their first release (even before they had introduced their native voice interaction) . Apple didn’t have a flagship frontier (or even a vanity open-source) model, no $500B compute commitment with the usual suspects. Meanwhile, the rest of the AI labs and big tech companies were racing to win the next state-of-the-art benchmark by burning bags of cash. What this also meant is that while these companies were burning money at a rate that would make a sovereign wealth fund uncomfortable, Apple was (and still is) sitting in a pile of undeployed cash (to the point of even increasing their stock buybacks) giving them optionality . To me, OpenAI is the most paradigmatic example of this “infinite money burning machine”. OpenAI raised at a $300B valuation and then shut down Sora, the video product they’d been positioning as a creative industry flagship, because it was running at roughly $15M a day in costs against $2.1M in daily revenue . Disney had already signed a three-year licensing deal for Sora to generate content from Marvel, Pixar, and Star Wars characters. They were finalising a $1B equity stake in OpenAI . When Sora died, so did the billion. A $1B investment evaporated, because the product it was staked on couldn’t pay for itself (reducing their buffer that accommodates their daily burn). On the infrastructure side: OpenAI signed non-binding letters of intent with Samsung and SK Hynix for up to 900,000 DRAM wafers per month, roughly 40% of global output . These were of course non-binding. Micron, reading the demand signal, shut down its 29-year-old Crucial consumer memory brand to redirect all capacity toward AI customers. Then Stargate Texas was cancelled , OpenAI and Oracle couldn’t agree terms, and the demand that had justified Micron’s entire strategic pivot simply vanished. Micron’s stock crashed . I don’t know about you, but I don’t see these behaviours as those of someone that is winning the AI race, independently of how good their models do in benchmarks, and how much they are burning in infrastructure. A small miscalculation in the expected revenue, and you are out of the game (I am actually of the opinion that without some kind of bailout, OpenAI could be bankrupt in the next 18-24 months, but I am horrible at predictions). From intelligence to capabilities My sense is that the labs’ bet was always that raw model capability, i.e. intelligence, along with the infrastructure required to run them would stay scarce. Those who manage to secure the best model and the infrastructure to run it at scale would get the best moat . But I am afraid that having the best model in itself may not be enough moving forward. Less capable models are becoming as capable as previous versions of the frontier models. The best recent example I can think of is Gemma 4 , Google’s open-weight model. It was built to run on a phone, scores 85.2% on MMLU Pro and matches Claude Sonnet 4.5 Thinking on the Arena leaderboard. 2 million downloads in its first week . Models that would have been state-of-the-art eighteen months ago now run on a laptop, and they get better every quarter. If you haven’t tried Gemma4 yourself I highly recommend it. I am running it on my AMD Ryzen AI Max+, and its performance in terms of tokens per second and intelligence are so good that I have already migrated some of my personal tools to use this model as the backend without visibly impacting their output. This trend can really change in the next few months way we access intelligence. I feel that some of the labs see this coming. Anthropic has been particularly aggressive about it and they are releasing new (actually useful) tools every day that work like a charm with their models in order to lock users into their ecosystem. Claude Code for developers, Claude Cowork for teams, the recent Claude Managed Sessions to orchestrate agents , all designed to put Claude inside workflows people are already in. The logic behind it: if the model itself won’t hold the moat, capture the usage layer and make switching painful. I think this is brilliant, and seeing how much Anthropic is growing in number of users and revenue, it seems to be paying off. The economics of their plans are still rough, though. One analysis found a max-plan subscriber consuming $27,000 worth of compute with their 200$ Max subscription. The labs are subsidising the demand they’re chasing, which justifies their level of burn (let’s see for how long they can afford these subsidies). Apple, by contrast, has spent almost nothing on AI infrastructure and subsidising users’ token burn. And this may be giving them more optionality and leverage than any of the other companies that jumped heads first into the AI race. Context is all you need In that earlier post , I argued tha t if intelligence becomes abundant, context becomes the scarce resource. A model that can reason about anything but knows nothing about you or the environment it operates in is a generic tool. What makes AI genuinely useful day-to-day is reasoning plus personal context: your messages, your calendar, your code, your tools, your health data, your photos, your habits. I think here is where Anthropic is making an amazing job with their “Claude suite”. But Apple already has all this context and access to your environment through their 2.5 billion active devices. Each one is a context mine that users have been filling for years. Health data from Apple Watch. Every photo taken on an iPhone. Notes, messages, location history, app behaviour, emails, and awareness of your environment through the pool of sensors of your device. Why build a commodity when they already have the context that can become their moat? And they even have the ability to keep all this data on-device, which is where the “Privacy. That’s iPhone” positioning becomes something more than a PR strategy, and which could actually make a comeback to become one of their core value propositions. Apple spent years using privacy as a differentiator against the ad-driven models of Google and Meta. It worked, but it always felt a bit abstract and, honestly, fake. Now it could become really concrete. Would you hand OpenAI your medical records and fifteen years of photos to get better AI answers? Probably not. Some are, but I personally wouldn’t like Sam to have that personal data from me. Would you let a model running entirely on your device (no network request, no data leaving your phone) access all of that? That’s a different question. The on-device model gets full context because it never leaves the hardware. Apple built the reputation and the architecture for this when no one else thought it mattered. Of course, there are still technological barriers to make this possible, but I feel like we may be getting there. In this context, the Gemini deal , where Apple signed a $1B to license Google’s frontier model f