In November 2025, Apple and Google announced a deal. Around a billion dollars a year for the language model powering the new Siri. The same Google Apple pays 20 billion a year to be the default search engine in Safari.
If you’re looking for proof that Apple is losing the AI race - there it is. A company that can design its own processor and its own operating system has to buy its thinking from a competitor.
And still, I think Apple can win.
Not that race
OpenAI, Google and Anthropic are chasing the same thing: the biggest model, the best benchmarks, the most convincing chatbot. That’s the race you see in headlines. GPT-5.5 versus Gemini 3 versus Claude Opus. Hundreds of billions of dollars in infrastructure. Twenty thousand GPUs in a single data center.
Apple isn’t in that race. AFM - Apple Foundation Model - has around 3 billion parameters. The current frontier runs in the cloud on compute you can’t fit in a phone. Apple made that choice deliberately.
The better question is: what do most people actually use AI for?
Summarizing emails. Fixing messages. Looking up quick answers. Dictating notes. None of those tasks require a trillion-parameter model. They require something fast, cheap to run, and always available.
Apple has that. It runs locally, on the phone. The first word appears in half a millisecond.
Nine years of silicon investment
Apple has been building a dedicated AI accelerator since 2017. The first Neural Engine in the A11 chip handled Face ID and Animoji. Today the M4 chip runs 38 trillion operations per second - more than 60 times more than A11 in 2017.
The M5, which arrived in MacBooks and Private Cloud Compute servers in early 2026, distributes neural accelerators across every GPU core. Peak AI throughput is more than four times higher than M4.
Nobody else controls the full stack: silicon, OS, developer framework, model. Samsung relies on other people’s chips and other people’s models. Qualcomm supplies silicon without an OS. Google has its own Tensor, but Pixel sells in the tens of millions - iPhone in the hundreds. That’s a scale gap no contract closes.
The cloud Apple can’t wiretap itself
Not everything fits in 3 billion parameters. When a task is too hard for the local model, iPhone sends it to Private Cloud Compute - Apple’s servers running on Apple Silicon, in a system with no logs, no remote access, no persistent data storage.
Apple published the PCC source code on GitHub. Independent researchers can run a complete server in a virtual machine. Apple offers up to $1 million for a serious vulnerability.
Associate professor Matthew Green from Johns Hopkins wrote that if you gave an excellent team a huge pile of money and told them to build the best private cloud in the world, it would probably look like this.
Meanwhile, ChatGPT stores your conversations, Anthropic in August 2025 introduced an opt-in data training toggle that extends conversation retention from 30 days to five years for those who agree, and in May 2025 a court ordered OpenAI to preserve all logs indefinitely as part of the New York Times lawsuit.
PCC has nothing to preserve.
Privacy as a moat
Competitors can’t copy Apple’s approach. Their business models are built on the opposite.
Google makes over $350 billion a year from ads. That engine runs on user knowledge - and no privacy statement changes that as long as the revenue model stays the same. OpenAI and Anthropic live off API fees and subscriptions, which means conversation data is a training asset, not a burden.
Apple makes money on hardware. It can afford AI that doesn’t trade user data, because it never made money that way. A business position the others can’t occupy without destroying their own revenue.
The contrast got sharper in May 2026. Researcher Alexander Hanff documented that Google Chrome silently installs a 4 GB Gemini Nano model on user devices without consent - and reinstalls it automatically if deleted. The file, called weights.bin, was found on machines that had never received a single human input in Chrome. The part that’s hard to ignore: Chrome 147 shows an “AI Mode” button in the address bar that looks like local processing, but every query typed into it goes to Google’s servers. The on-device model and the cloud-backed surface are two separate things. Chrome doesn’t explain that distinction.
Regulation does the rest. In December 2024, Italy’s data protection authority fined OpenAI €15 million for GDPR violations. The EU AI Act, with fines up to €35 million or 7% of global revenue for the most serious violations, came into force in August 2025. Legal uncertainty around cloud AI is growing. Apple doesn’t have that uncertainty - structurally, there’s nothing to regulate.
Where the strategy breaks
The notification summary feature generated false headlines - that a CEO murder suspect had shot himself, that Nadal had come out as gay. The BBC publicly stated that Apple’s AI summaries contradicted original reporting. Apple disabled the feature for news apps.
The personalized Siri Apple showed at WWDC 2024 - the one that knows when mom landed and automatically plans the weekend - doesn’t exist. In March 2025 Apple admitted the features were pushed to an unspecified future date. John Gruber wrote an essay titled “Something is rotten in Cupertino.” Hard to argue he was wrong.
In July 2025, the head of the entire Apple Foundation Models team left for Meta for a package estimated at over $200 million. A dozen key people followed. In April 2026, John Giannandrea left - the AI chief Apple had recruited from Google in 2018.
And there’s that Google deal. The new Siri will think with Gemini. Apple will pay a billion dollars a year for it. A privacy platform, powered by the world’s largest data collector.
What a winning scenario looks like
Apple won’t win the race for the best chatbot. It doesn’t have the budget, the talent pipeline, or the training infrastructure - and the distance to frontier labs is growing faster than Apple is closing it.
But the chatbot race isn’t the only race.
The Foundation Models framework, which Apple opened to developers in June 2025, gives every iOS app a local model at zero cost per token. A small company building an AI notepad doesn’t need to pay OpenAI a few dollars per user per month. That changes the economics of thousands of apps at once.
iOS 27 goes further. Users will be able to set Claude, Gemini, or any other model as the default AI provider for system features - Writing Tools, Image Playground, and probably more. Apple isn’t trying to win with a model. It’s becoming a model aggregator. The same way it won music distribution without being a label, or search without building its own search engine. It controls the interface. Everyone else competes for placement in its store.
Visual Intelligence in iOS 27 shows how Apple thinks about quiet AI: you scan a food label, the data goes to Health. You scan a business card, the contact lands in Contacts. You photograph a gym membership card, it appears in Wallet. Nobody shouts “artificial intelligence.” It just works.
Apple has 2.5 billion active devices. Over 250 million already support Apple Intelligence today. Every iPhone with A19 and every Mac with M5 sold grows that fleet. No cloud provider has an installed base of hardware capable of local inference at that scale.
The winning scenario doesn’t look like “Apple built a better model than OpenAI.” It looks like: in five years, on-device AI is as normal as GPS, Apple controls the interface through which users choose models, and the iOS app ecosystem built itself around free local inference. In regulated environments - healthcare, finance, law - Apple is the default choice, because competitors have a problem Apple doesn’t.
The Google deal, in this reading, isn’t surrender - it’s tactical time-buying. Apple Maps ran on Google data for years before Apple built its own. Siri might take the same road.
Might. That’s an assumption.
A bet that’s hard to copy
AI as a superintelligent chatbot is one definition of winning. AI as infrastructure - always available, free to run, private by design - is another.
Apple bet on the second. Nine years in Neural Engine, unified memory architecture, Private Cloud Compute, Foundation Models framework - one coherent wager that the winner won’t be whoever has the biggest model, but whoever has the best distribution and the fewest reasons to hurt you.
The rest of the market can’t easily replicate that bet. They have different business models they can’t abandon.