WWDC 2026: Apple Confirms Gemini-Powered Siri in Tim Cook's Farewell Keynote
Apple confirmed a multi-year, roughly billion-dollar-a-year deal to run a custom Gemini for Siri's cloud intelligence. The privacy story is strong; the dependency story is bigger.
Apple used its June 8 WWDC 2026 keynote — widely framed as Tim Cook's farewell — to confirm the worst-kept secret in Cupertino: the cloud intelligence behind the long-delayed new Siri will be powered by a custom, Apple-tuned version of Google's Gemini, a roughly 1.2-trillion-parameter model running under a multi-year license that industry estimates put at around a billion dollars a year. The demos were strong. The strategic admission underneath them was the actual headline, and Apple delivered it in the calmest register it could manage.
How the architecture splits
Apple's pitch is a tiered system. Light, latency-sensitive, privacy-intimate tasks — the expressive new voices, dramatically improved dictation, on-screen awareness, and quick lookups against personal context — run on-device, on Apple's own next-generation foundation models executing on Apple Silicon. Heavy world-knowledge and complex-reasoning requests route to the cloud through Apple's Private Cloud Compute, where Apple says data is processed without being stored and without Apple itself being able to read it. The cloud tier is branded "AFM Cloud Pro" — Apple Foundation Models on Cloud — and positioned as comparable in quality to Gemini's frontier models. It should be: it is a custom-tuned Gemini.
The naming is the strategy in miniature. Users will not see "Gemini" anywhere. They will see "Apple Intelligence" and "Apple Foundation Models," and the reasoning doing the impressive work will be Google's.
Estimated request routing: on-device vs. Gemini cloud tier (illustrative)
| task | onDevice | gemini |
|---|---|---|
| Set a timer | 100 | 0 |
| Dictate a note | 95 | 0 |
| Summarize a thread | 15 | 50 |
| Plan a trip | 5 | 80 |
| Open-ended research | 0 | 90 |
The shape matters. The easy, frequent tasks stay on-device, where Apple's models are genuinely competitive. The hard, memorable tasks — the reason anyone will call the new Siri good — lean on the Gemini tier. Apple owns the floor; Google supplies the ceiling.
The rest of the keynote
It was a large release. iOS 27, macOS 27 — codenamed "Golden Gate" — iPadOS 27, watchOS 27, tvOS 27, and visionOS 27 all shipped, unified under a sweeping new "Liquid Glass" design language that reworks materials, depth, and translucency across the platforms. homeOS got its first formal preview. Much of the runtime went to design and to the on-device tier — the parts Apple unambiguously owns — which had the convenient effect of surrounding the one announcement that punctures the self-reliance narrative with a wall of competence.
Why this is a dependency story
The privacy engineering in Private Cloud Compute is real and impressive, with attestable environments and verifiable software images. But the privacy frame answers "can Google read my messages?" — not "who supplies Apple's intelligence?" Those are different questions, and the second is the strategic one. You can have a perfectly private pipeline to a capability you do not own.
A recurring model-license line item is a new kind of cost for Apple. Its services business has thrived on recurring, high-margin revenue built on infrastructure Apple already owns. A roughly billion-dollar annual payment to Google for capability, scaling loosely with usage of the marquee feature, runs against that grain. The bigger issue is bargaining power over time: as Apple rebuilds Siri around the assumption that the Gemini tier exists, and as developers build App Intents on top of it, the cost of removing it rises and renewal leverage drifts toward Google.
Estimated negotiating leverage over the dependency's life (illustrative)
| phase | appleLeverage | googleLeverage |
|---|---|---|
| Initial term | 55 | 45 |
| Mid-term | 42 | 58 |
| First renewal | 30 | 70 |
| Deep dependency | 22 | 78 |
There is a sharp irony in the comparison to Apple's other big Google deal. For years Google has paid Apple roughly twenty billion dollars a year to be Safari's default search engine — pure-margin distribution rent Apple collects for owning the device. The Gemini-Siri deal reverses the direction: now Apple pays Google for capability. The magnitudes are not symmetric, but the directions tell the strategic story.
Two Google deals, opposite directions ($B/yr, approximate)
| deal | appleReceives | applePays |
|---|---|---|
| Search default | 20 | 0 |
| Gemini-Siri license | 0 | 1 |
What it means for the services margin
The market reaction will hinge on whether investors read the deal as shrewd procurement or as the start of structural margin erosion. Both readings are defensible. Apple's services gross margin has run north of seventy percent, and that figure is the engine behind the company's valuation multiple — recurring, high-margin revenue that mostly rides on infrastructure Apple already owns. Layering a model license that accrues to a competitor, plus the compute to serve growing cloud-query volume, introduces a cost of revenue that scales with the success of the feature. A one-point compression will not frighten anyone in year one. The concern is directional: for the first time, a meaningful and growing slice of the cost of delivering Apple's intelligence is set by a price Google controls.
There is also a quieter accounting wrinkle. A multi-year license with committed minimums behaves, in substance, like a long-dated operating commitment — the kind of obligation that shapes how analysts model the durability of free cash flow. Apple will not capitalize a model it does not own. What it holds instead is a contractual dependency with a duration and, almost certainly, a price escalator, entered into against the alternative of shipping a Siri without competitive cloud reasoning. That is a debt in everything but name, sitting on the income statement rather than the balance sheet.
What developers should watch
For developers building App Intents and Siri integrations, the practical takeaway is a new flavor of platform risk. The ceiling of what Siri can do is now partly a function of a contract between Apple and Google whose terms developers cannot see and whose renewal they cannot influence. If the Gemini tier improves, integrations get better. If the economics push Apple to throttle cloud routing to protect margin, integrations quietly degrade — and developers will not be told why. The most ambitious, most agentic integrations carry the most of this third-party exposure, because they lean hardest on the rented tier.
The succession optics
That this landed on Cook's farewell keynote is the part that will age. The CEO who made Apple a machine for controlling its own fate exits on a keynote whose centerpiece is a concession that the defining technology of the decade is one Apple could not build in time, and so chose to rent — from its largest mobile rival, on terms that will only get harder. The new Siri is genuinely good. It thinks with Google's mind. And every quarter it ships, the price of getting that mind back to itself goes up.
For the deeper accounting and strategy analysis — why this is a balance-sheet story wearing a product's clothes — see the companion piece, Apple's One-Billion-Dollar Dependency: WWDC 2026 and the Gemini Siri Deal.