Apple just paid Google roughly a billion dollars a year to be Siri's brain. The most vertically integrated company in history evaluated three AI providers, rejected two, and rented the intelligence it couldn't build. Twenty years of owning the full stack ended with a cloud computing contract.
In January 2026, Apple signed a multiyear deal with Google to power the next generation of Siri with Gemini's large language models. The reported cost is roughly a billion dollars a year. The custom Gemini system will handle Siri's summarizer and planner functions, running on Apple's Private Cloud Compute infrastructure to maintain privacy standards. Google's branding will be invisible — the deal stipulates that Gemini's role will be white-labeled with no Google name visible to end users. Siri will answer as Siri. The intelligence behind the answers will be Google's.
Apple is the most successful vertically integrated company in history. It designs its own chips, builds its own operating systems, manufactures its own hardware, runs its own services, and controls its own distribution. For twenty years, this strategy produced the most valuable company on Earth. It worked because Apple could master every layer of the stack and polish the seams between them into an experience no competitor could match.
AI broke the model.
The Shopping Trip
Apple did not go directly to Google. It evaluated alternatives. Anthropic had custom versions of Claude running on Apple's servers and was in contention to become the backbone for Siri. But Anthropic reportedly demanded several billion dollars a year, with the amount doubling annually for the next three years. OpenAI declined — the two companies are increasingly competitors, with Apple developing its own AI search capabilities and OpenAI pursuing device manufacturing. That left Google.
The elimination process reveals something structural about the AI market. Apple — a company that could acquire most AI startups outright — found itself with exactly one viable partner for its most important product. Anthropic priced itself as the scarce resource, which it may be. OpenAI recognized Apple as a rival, not a customer. Google won partly by default, which is another way of saying Google won because it already had leverage.
Apple already pays Google roughly twenty billion dollars a year to be the default search engine on iOS. That deal survived antitrust proceedings in 2025 when Judge Amit Mehta ruled it could continue despite Google's monopoly finding. It is the single largest business-to-business payment in the technology industry. Now Apple is adding a billion-dollar AI dependency on top. Google's influence over Apple is not shrinking. It is compounding.
The pricing tells its own story. Anthropic wanted several billion annually, doubling each year. Google reportedly settled for roughly a billion. The gap is not explained by model quality — Anthropic's Claude is competitive with Gemini by most benchmarks. The gap is explained by an existing relationship. Google already runs Apple's search. It already operates at Apple's scale. It already has the cloud infrastructure. The marginal cost of adding AI to an existing twenty-billion-dollar partnership is lower than starting a new multibillion-dollar one from scratch. Incumbency compounds.
The Confession
Before the deal was signed, Apple's Siri team already knew the overhaul was failing. In March 2025, Robby Walker, then senior director for Siri and information intelligence, told employees in an all-hands meeting that Siri's new features were working properly only sixty to eighty percent of the time. He called the delays ugly and embarrassing, acknowledging that Apple had shown off features publicly before they were ready. This was not one of these situations where we get to show people our plan after it's done, Walker told the staff. He acknowledged that employees might be feeling angry, disappointed, burned out, and embarrassed.
Walker left Apple in October 2025. Three months later, Apple signed the Google deal.
The reliability number matters more than it appears. Sixty to eighty percent accuracy in a demo is impressive. Sixty to eighty percent accuracy on two billion devices is a different category of problem. If Siri mishandles one in five requests across Apple's entire user base, the failure surface is measured in hundreds of millions of bad interactions per day. The gap between making AI work and making AI work reliably at planetary scale is where Apple's vertical integration advantage was supposed to deliver. It did not.
Apple Intelligence was announced at WWDC in June 2024 with an agentic Siri promised for spring 2025. That deadline slipped. Then slipped again. In February 2026, the first beta of iOS 26.4 was released to developers with no new Siri features included. Bloomberg reported that some capabilities may ship in iOS 26.5 in May. Others may wait for iOS 27 in September — more than a year past the original target.
Intelligence Versus Compute
This journal previously covered Meta signing a multi-billion-dollar deal to rent Google's TPUs — compute infrastructure — even after building the largest private AI compute cluster in the world. That was a commodity decision. You rent compute the way you rent warehouse space: the capacity matters, the brand does not.
Renting intelligence is structurally different. When Apple rents Gemini, it is not renting a fungible resource. It is renting the capability that makes Siri competitive. Without Gemini, Siri is the assistant that works sixty to eighty percent of the time. With Gemini, it competes. The dependency is not on infrastructure but on the thing that generates value for the end user.
This distinction has implications that extend well beyond Apple. For two decades, the platform owners — Apple, Google, Microsoft, Amazon — captured disproportionate value because distribution was the scarce resource. Building a great product was necessary but not sufficient. You needed the app store, the default setting, the pre-installed slot. The platform controlled access to users, and the companies that controlled platforms extracted rents from everyone who wanted that access.
AI may be inverting that relationship. If the intelligence layer becomes the value layer — if the model's quality matters more than the device it runs on — then the companies that build the best models will extract value from the companies that own the most devices. The Apple-Google deal is early evidence. Apple has the distribution. Google has the intelligence. Apple is paying Google, not the other way around.
The existing search deal already suggests the direction. Google pays Apple twenty billion for access to Apple's users. But Google makes considerably more than twenty billion from those users' search activity. The payment is a distribution tax — significant, but smaller than the value extracted. The AI deal adds a second channel through which Google provides intelligence and Apple provides users. In both channels, the intelligence provider captures disproportionate value relative to the distribution provider.
What the Label Hides
Apple insisted that Google's branding be invisible. Siri answers as Siri, not as Gemini. This is not vanity — it is strategic necessity. If users learn that Siri's intelligence is Google's, Apple's brand premium erodes. The vertical integration story — we build everything, that is why it is better — loses its force when the most important layer is rented.
But white-labeling intelligence has a structural weakness that white-labeling components does not. When Apple uses Samsung-manufactured displays or TSMC-fabricated chips, the component is frozen at the point of purchase. A display panel does not get better or worse after it ships. An LLM does. When Google improves Gemini, Siri improves. When Google deprioritizes Apple's workload, Siri degrades. The intelligence layer is not a part — it is a living dependency whose quality is determined by the supplier's ongoing investment.
This creates a control problem that no contract fully resolves. Apple can specify performance benchmarks and service levels. It cannot specify that Google continue to pour its best researchers and its latest techniques into a white-labeled product that competes with Google's own Assistant. The incentive tension is fundamental: Google profits from Apple's dependency, but Google also profits from its own AI products. Every improvement to Gemini-for-Apple is simultaneously an improvement to the technology that Google could use to compete with Apple directly.
For twenty years, Apple resolved this kind of tension by bringing the capability in-house. It designed its own chips when Intel's roadmap diverged from Apple's needs. It built its own maps when Google's terms became unfavorable. It launched Apple Music, Apple TV+, Apple Pay — each time replacing a dependency with an owned capability. The AI deal is the first time in Apple's modern history that it moved in the opposite direction: from attempting to build in-house to renting from a rival.
The question is whether this is a temporary bridge or a permanent condition. Apple is almost certainly investing in its own foundation models. But the pace of frontier AI development — major capability jumps every six to twelve months, billions of dollars in training costs per run — may outstrip any single company's ability to catch up. Intelligence may be the first technology that moves faster than Apple can integrate it. If that is true, then the outsource is not a stopgap. It is the new architecture.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)