App Economy Insights — The Trillion-Dollar Off Switch

Apple has been embarrassingly slow at AI — the smart Siri it promised in 2024 slipped more than two years and even drew a lawsuit. This article argues that lateness accidentally protected it. Instead of building its own giant AI model, Apple is renting one: it pays Google about $1 billion a year to run a custom version of Google's Gemini model (with ~1.2 trillion "parameters," a rough measure of a model's size/capability) as the new brain behind Siri, keeping only a tiny model on the phone itself.

Why that's a hedge: building a frontier model means spending tens of billions on data centers ("CapEx," long-term capital spending) and owning an asset a government could order switched off — exactly what just happened to Anthropic. By renting, Apple sidesteps that bill and that risk; if one model becomes a problem, it can swap it. The cost is control — the smartest version of Apple's flagship assistant now runs on a direct rival's model, it won't launch in the EU, and Apple's edge narrows to privacy and how neatly everything integrates. The piece reads it as a strategic stance, not a buy/sell call.

Google builds its own frontier AI model (Gemini), which normally looks like the riskiest position to be in — if a government forces a model offline, you lose your crown jewel. But Google is cushioned: Gemini sits inside a company that also runs Search, Android and Google Cloud, so a model shutdown wouldn't take the whole business down with it. That's the opposite of a pure-play AI lab.

Google is also quietly winning from everyone else's caution: it's collecting ~$1 billion a year to power Apple's Siri with a custom Gemini, and that model runs on NVIDIA chips inside Google's own data centers. So Google both owns the frontier and rents it out — getting upside from the buildout while having other businesses to absorb a regulatory shock.

Anthropic is a private AI lab (maker of the Claude models) whose paper valuation just rocketed to $965 billion — on paper the most valuable standalone AI company in the world. Then the article's whole point landed: the US government issued an export-control order restricting who could use its two most powerful models. Because Anthropic can't reliably check every user's nationality in real time, it had to switch those models off for everyone, everywhere.

That's the danger of being a "pure-play" lab: the AI model basically is the entire business, so when it gets switched off there's no Search engine or app store or cloud division still earning money to cushion the blow. The same models, ironically, are valuable enough that elite partners (Apple, Google, Microsoft, CrowdStrike) test them to find security flaws — genuinely dual-use technology. The takeaway for anyone eyeing Anthropic's coming IPO: its biggest asset can be turned off by a phone call from Washington.

OpenAI (maker of ChatGPT) is the other giant pure-play AI lab, and it has filed confidentially to go public — potentially worth over $1 trillion. The article groups it with Anthropic because it carries the exact same structural risk: the model is the whole company, so a government recall or export restriction would hit it with no other revenue to fall back on.

The piece's bottom line is that for investors weighing these labs, the old question ("who has the best model?") matters less than a new one: "who can keep their model online?" Owners with broad businesses (Google) or renters (Apple) are more insulated than a standalone lab like OpenAI.