Apple just rented Siri's brain
and that is the right call

Apple designs its own A20 Pro chip. It writes its own operating system. It runs its own stores. And on June 8, 2026²², it stood on the WWDC stage and told the world that Siri's brain now belongs to Google.

Bloomberg reported on November 5, 2025¹³ that Apple was nearing a deal to pay Alphabet roughly $1 billion a year for a custom Gemini model of about 1.2 trillion parameters. On January 12, 2026²³, Apple and Google confirmed a multi-year deal. The rebuilt Siri is expected to ship with iOS 27⁰⁴ this September. That is more than two years after Apple first announced Apple Intelligence on June 10, 2024²⁴.

Here is the damaging admission: I do not know the exact contract terms, and neither does anyone outside Cupertino and Mountain View. But the shape is clear. The most vertically integrated hardware company in history decided the frontier model layer was not worth owning. If Apple will not build it, you should think very hard before you do.

The Half-Life Rule of Layers

Every layer in a technology stack has a half-life, meaning how long it stays competitive before the market moves past it. Apple silicon has a half-life measured in years. iOS has one measured in decades. A frontier model has one measured in months.

That gap is the whole story. Between June 2024 and June 2026, frontier models turned over roughly three generations by most counts. The same week Apple's Gemini-powered Siri went public, GPT-6 Astra was reported to have posted roughly 98%⁰¹ on FrontierMath Tier 4 and 99.9%⁰² on ARC-AGI-3. Any model Apple started training in 2024 would have been a museum piece before it shipped.

So here is the rule. Own the layers whose half-life is longer than your build cycle. Rent the layers whose half-life is shorter. If it takes you 24 months to build something that depreciates in 6, you have not built an asset. You have built a liability with a launch party.

Apple's newly announced Siri architecture is the rule made concrete. Tier 1 runs Apple's own small models on-device for timers and simple actions. Tier 2 runs Apple's own models on Private Cloud Compute for moderate work. Tier 3 routes the hardest reasoning to a custom Gemini model on Google Cloud. Two owned, one rented. And the rented one is the layer that changes fastest.

Apple Already Rented Its Foundry

Start with a question that sounds naive. Does Apple own a chip fab? No. TSMC manufactures every A-series and M-series processor Apple has ever shipped, and nobody calls Apple un-integrated because of it.

Amateurs count what they own. Strategists count what they control. Apple controls the chip design, the instruction set, the thermal envelope, and the software that runs on it. TSMC controls a very expensive building. The same logic now applies to Gemini. Google controls a very expensive model. Apple controls the routing, the privacy layer, and the surface 2.5 billion devices see every day.

Now look at the cash, because only cash is real and the rest is accounting. Apple already receives roughly $20 billion⁰³ a year from Google for the Safari search default. The Gemini fee runs roughly $1 billion a year in the other direction. That is about one twentieth of the search money, a 20x difference, and it deepens a dependency Apple spent a decade claiming it was engineering away from. The public narrative was independence. The cash flows said partnership the entire time.

The Nvidia case is the mirror image. In 1996 Nvidia was months from death because its NV1 chip fought Microsoft's Direct3D standard instead of adopting it. The RIVA 128 saved the company by building on top of the standard rather than against it. Apple fighting the frontier model layer in 2024 was the NV1 move. Renting Gemini in 2026 is the RIVA 128 move.

Why Google and not OpenAI? Reuters reported Apple chose Gemini because, according to a joint Apple and Google statement, it judged it the most capable foundation. Both reasons matter, and neither has anything to do with owning weights. When you rent, you pick the landlord who will still be there in five years.

There is a real counterargument, and it deserves a straight answer. The antitrust attorney Luis Blanquez wrote that Siri's cognitive core "will no longer be Apple's. It will be Google's." True, at the model layer. But Apple kept the on-device models, the Private Cloud Compute tier, the anonymization step, and the decision of which query goes where. In iOS 27⁰⁴, Apple is also opening an Extensions framework so users can pick from assistants including Claude, Grok, Copilot, Perplexity, or a Gemini chatbot alongside ChatGPT.

Read that Extensions move as counterpositioning. Apple is turning itself into the buyer every model vendor must court. A buyer with five suppliers has pricing power. A buyer with one supplier has a hostage situation. Apple learned this with modems and displays over 20 years⁰³ and is applying the same shoshin, the beginner's mind, to models.

It is unclear whether the reported $1 billion-a-year arrangement is the "stopgap" Bloomberg described or a permanent equilibrium. My read is that the stopgap framing is a face-saving story. The half-life of frontier models is not going to lengthen, so the economic logic that pushed Apple to rent in 2026 will still hold in 2028.

Half-life decides what you own

2031. Pull back five years. Estimates for the foundation-model market disagree wildly, and the disagreement is itself the signal. One industry overview projects revenue growing from $22 billion in 2025 to $125.4 billion by 2029. Another firm projects $21.72 billion in 2025 rising to $119.29 billion by 2031. When analysts cannot agree within a factor of two, the layer is being priced like a commodity utility, not like a moat.

This is the cloud playbook repeating. In 2008 serious companies debated whether to build their own data centers. By 2018 that debate was over for everyone except hyperscalers. The model layer is running the same arc on a compressed clock. Renting frontier intelligence in 2031 will feel as normal as renting servers does today.

Consider the asymmetry. Build your own model and win, and you gain independence plus a slightly better margin. Build and lose, and you burn two years shipping something obsolete, which is exactly what happened to Apple between 2024 and 2026. Rent and get squeezed by the vendor, and you switch vendors, which Apple's Extensions framework already makes structurally possible. The downside of renting is annoying. The downside of building is fatal.

Think about Costco's hot dog. It has cost $1.50²¹ since 1985 and Costco loses money on every one. Costco does not care, because it owns the membership relationship and the hot dog exists to feed that flywheel. Siri handles about 1.5 billion requests a day, according to Apple's own marketing, a figure that is not independently audited. Gemini is Apple's hot dog. The 2.5 billion active devices are the membership.

The asset that compounds by 2031 is the routing data, the device relationship, and the trust that comes from anonymizing every query before it leaves the phone. Not the model. Those layers have a half-life of a decade. The model powering them will have changed five times. Impermanence is not a risk to manage here. It is the environment you are building in.

Partition Your Stack Before iOS 27 Ships

You can copy Apple's partitioning this weekend without a CS degree. The principle is simple: own what depreciates slowly, rent what depreciates fast, and never let the rented part touch your customers directly.

First, list every layer of what you are building and give each one a half-life. Your customer data, your domain rules, and your user interface probably last years. Your model choice lasts months. Write the numbers down, even if they are guesses. Guessing badly beats not asking.

Second, put every model call behind a single routing layer. A routing layer is just one file or one function that decides which model handles which request. Apple has three tiers; you need two at most. Swapping Gemini for Claude should mean changing one line, not rewriting your product.

Third, log your own data on your own storage. When a query goes out and an answer comes back, keep both. That log is the layer with the longest half-life you own, and it is what makes switching vendors painless later.

Fourth, keep two vendors live from day one. Send 10% of traffic to your backup model and compare answers weekly. Things will break. That is the point. You want to discover the break in your test traffic, not in a customer's inbox.

The tools in today's digest all reward this thinking. Softr AI gives you a working web app from one prompt, typically after a few clarifying questions, and you should treat that output as scaffolding, not the finished product. Ajelix is marketed as able to run a bounded project end to end from a single instruction, so hand it something small and read the audit trail like a routing log. Robin AI automates prospecting, but the definition of a qualified lead is your owned logic, so write it before you switch anything on.

Meta's Business Agent is about to get more autonomous after the Stilla.ai acquisition on September 9. Set your reply rules now, because those rules are the layer you own. The model underneath will change. Your rules should not.

Apple took two years and a public humiliation to learn the Half-Life Rule. You can learn it in a weekend. Build one tiny thing, put a router in front of the model, and get your reps in.