Competition is for losers.

The Verdict

Down 2%, on an up tape.

Siri AI shipped in beta at WWDC—rebuilt on a custom Gemini, with no consumer timeline. AAPL touched an all-time high during the keynote, then closed down 2% while the S&P rose. Daniel Ives of Wedbush read the deal as an admission that Apple "cannot win the AI arms race alone"; Gene Munster of Deepwater watched the stock sell off on the missing Siri date. A stock at roughly forty times trailing earnings, against a ten-year median near twenty-six, is priced as an AI competitor. The verdict wrote itself: Apple lost AI.

The Deal

One billion dollars a year.

About $1 billion annually for a custom 1.2-trillion-parameter Gemini that runs Siri's hardest queries on Apple's own Private Cloud Compute servers. Apple is paying its oldest rival to think for Siri. Agreed in January, shipped in June—fast work for a company supposedly five years behind.

The Scale

Across 1.5 billion requests a day.

Siri fields about 1.5 billion requests a day, by Apple's own disclosure—roughly 547 billion a year, up from 25 billion a month in 2020. Whatever the frontier costs, Apple spreads it across a planetary denominator.

The Unit Cost

A fifth of a cent.

$1B ÷ 547B answers ≈ $0.0018 per request, at Apple's disclosed June 2024 volume. If only 10–30% of requests ever reach Gemini, the bill is 0.6–1.8¢ per answer—still trivial. Per device it lands near forty cents a year across 2.5 billion; Apple discloses >1B iPhones, so call it under a dollar per iPhone. The cheapest iCloud tier costs more per month than the frontier costs Apple per device per year.

The Other Bill

Or build it for a hundred billion.

The other path is being the frontier. Training one Gemini model ran about $190 million (Gemini Ultra, 2023, the last public estimate); being the provider costs Alphabet roughly $49 billion a year in company-wide R&D and ~$185 billion in 2026 capex—"more than $100 billion over time" on AI, by DeepMind chief Demis Hassabis's own count. Different ledgers, not a sum. On every one of them, Apple's rent is a rounding error on a war it declined to fund.

The Invisible Engine

You never see Google.

Gemini runs stateless inside Apple's Private Cloud Compute and is not labeled anywhere in Siri. The user sees Apple. Apple keeps the brand, the interface, the relationship, the data. Its supplier ships weights and gets a check.

The Leverage

And it isn't locked in.

The license is non-exclusive. Apple ran a bake-off (Gemini against OpenAI and Anthropic), keeps ChatGPT integrated, and dangles ~2 billion devices over every renewal. Google didn't pay to be generous; it paid to keep OpenAI off two billion devices. Apple monetized the labs' fear of each other.

The Trade

The number the verdict missed.

Here's the number the verdict missed: Google pays Apple ~$20 billion a year—twenty times the Gemini fee—for the same users. Call it a retcon if you like—Apple spent three years failing to build this in-house. So did Maps, in 2012. The playbook is judged by where it lands. Apple isn't renting its brain; it's leasing the only shelf that matters, and the brain's owner is its biggest tenant.

Apple licensing Google's Gemini (~$1B/yr) to power Siri is not AI capitulation but interface-monopoly strategy — Apple monetizes a planetary-scale interface while Google funds the hyperscale infrastructure.

That $20 billion has its own twenty-year story—the founding play →

The Door

The newest play of fourteen.

Siri-on-Gemini is one run of a playbook Apple has executed for over a decade: kill a default, tax a path, own a prompt, rent a fab. Every tile below is a door, and the doors are real. Choose where to go next.