How Will John Ternus Change Apple’s AI Strategy? Siri, Apple Intelligence, and Gemini

Short answer: Ternus became CEO on 1 September. The 9 September debut cannot unwind the AI stack Cook already locked. Siri AI still runs in three layers — on-device Apple Intelligence, Private Cloud Compute, and a custom Gemini. The reported ~$1 billion-a-year Google license, English-first access, the waitlist, and the EU / China holds are not contracts a two-day CEO can shred. The question Cook left him on the 30 July earnings call is the one that matters: once people actually use Siri AI, who pays the compute bill.

This piece is current as of 3 September 2026. IDC’s read still holds: the foldable iPhone and the new Siri at “Surprise and Shine” were scheduled on Cook’s watch. Ternus’s own direction will not be readable until 2027. If you ship App Intents or call Foundation Models, the useful fact is not whether he swaps the model. The JSON contract Siri uses to invoke your app is already set. Models can move. Loose schemas cannot.

Handoff inventory: what Ternus inherits

Ternus is Apple’s third CEO since 1997. The succession was announced in April; he took the chair on 1 September, the same day Cook became executive chairman and kept government and policy relationships. He is a mechanical engineer who joined in 2001, ran hardware through a 2021 SVP title, and owns the résumé for Apple Watch, AirPods, Vision Pro, and the Intel → Apple Silicon shift. “Engineer-CEO” contrasts Jobs’s product narrative and Cook’s supply-chain execution. It does not mean he will train a frontier model himself.

His day-one staff memo thanked Cook, then called next week’s launch “phenomenal.” Bloomberg, MacRumors, and 9to5Mac all quoted the four-paragraph note. What it does not contain is a new AI roadmap, or a line that says Apple will replace Gemini with an in-house flagship. Week one is still a product launch, not a strategy pivot.

The ~$1 billion and 1.2-trillion-parameter figures are not official Apple specs. They are high-confidence reporting from Bloomberg, CNN, and others. Apple’s own language is only that cloud features run on advanced models inside Private Cloud Compute. Treat them as reported numbers in product docs, not a price list.

Three layers: Siri, Apple Intelligence, and Gemini

If Ternus changes strategy, he is changing who eats traffic, who pays, and what the user sees — not a slogan. WWDC 2026 on 8 June rebranded the assistant as Siri AI. The stack is still the 2024 Apple Intelligence story, except the cloud ceiling moved from Apple’s mid-size cloud models to a custom Gemini.

Users can still hand a single session to ChatGPT, as they could in 2024. The default spine is Gemini, not OpenAI. Apple’s 2026 trade-secret suit against OpenAI makes that optional exit a hedge, not the main path. On the developer side, the Foundation Models framework exposes both on-device and supported cloud inference — so the “Apple model” you call from Xcode is already Gemini-related in the cloud lane.

The waitlist exists for the same reason Apple licensed Gemini: trillion-parameter inference at the top of the stack does not scale overnight. Apple Intelligence has been capacity-gated since iOS 18. In the iOS 27 developer beta, the new Siri queue ran as long as a week. 9 September will still be “join a list, English, US first,” not a global switch.

9 September: what he can change and what he cannot

“Surprise and Shine” is 9 September at 10 a.m. PT, Apple Park. Expected hardware: the first foldable iPhone Ultra (about 5.5-inch cover, 7.6–7.8-inch inner display, A20 Pro, Touch ID in the power button, analyst-consensus starting price around $1,999), plus iPhone 18 Pro / Pro Max on TSMC’s 2nm A20 Pro. Software dates for iOS 27, watchOS 27, and macOS 27 Golden Gate should land the same day. The standard iPhone 18 slips to spring 2027 — not an AI decision, but it does change holiday-quarter revenue recognition.

Siri AI is the software headline and Ternus’s opening exam. Gene Munster set the bar at “10 out of 10” before the event; that score now attaches to the new CEO, not a feature he can blame on his predecessor. Do not listen for “we will train our own frontier model.” Listen for whether he makes three things explicit:

• Who gets it: English, waitlist, iPhone 15 Pro and newer; no timeline for EU iPhone / iPad; China still in regulatory review.
• Who pays for compute: a free allowance, iCloud tiers, or a later Apple Intelligence+ SKU. Cook left “still uncertain” on the table.
• How Gemini is named: only “advanced models inside PCC,” or Google said out loud on stage. The wording is the strategy signal.

IDC’s Francisco Jeronimo told DW the obvious thing: 9 September matters, but nothing yet shows Apple behaving differently because Cook left. Reading the foldable and Siri as Ternus’s “new strategy” mis-dates the timeline. That is the inherited debut, not the roadmap he designed.

What an engineer-CEO is likelier to change

The only extrapolation that holds is the job he already did: pull a critical capability back from a supplier, while Apple still defines the product. Apple Silicon is that script — ship on someone else’s CPU, then make the experience uncopyable on your own silicon. AI is the same logic with a longer fuse, and the counterpart is Google instead of Intel.

More likely

• Push traffic down a layer: let the on-device 3B and the smaller PCC models eat timers, summaries, and short commands, and keep daily chat off Gemini. That saves money and strengthens the “data stays on device” story. Ternus’s résumé fits this layer, not an Arena chase against Gemini 4.
• Distill into Apple Foundation Models: reporting says Apple trains its on-device / PCC weights on Gemini outputs, and user data does not touch Google by default. Every gain in those weights is leverage at renewal.
• Hedge across models instead of betting one vendor: the ChatGPT ramp is still there; Anthropic is widely described as another iOS 27 fallback. A “Search or Ask” panel that can route to Siri, ChatGPT, Gemini, or Claude is how you refuse a single-supplier fate.
• Talk products, not model cards: foldables, Watch, and AirPods are his home field. Ambient compute on a headset or a wrist looks more like Ternus than another chat app.
• Put AI cost into iCloud and hardware ASP: Cook already named “use it a lot, move up an iCloud plan.” An engineer-CEO is likelier to publish quotas, tiers, and device gates than to drop $100 billion on training clusters to match Meta or Microsoft capex in 2026.

Readable only in 2027

• How large a training fleet Apple builds: PCC already extends onto third-party cloud (Google Cloud + Nvidia confidential compute). The Silicon script points to more Apple-owned inference halls, not a 2026 announcement that “we will stop buying Gemini.”
• A public line on regulation and antitrust: he entered the job with almost no congressional record on AI governance. The DOJ case, the DMA, and China still sit with Chairman Cook. Wedbush’s Dan Ives said it cleanly: he knows the hardware; the global stage is on-the-job learning.

What will not move in the near term

• Killing the Gemini contract in September: Siri AI’s cloud lane is bolted to that license. Switching vendors in debut week is how you cancel your own keynote.
• Full Siri AI in the EU or China on day one: the DMA wants rival assistants to get the same deep device access as Siri; Apple says no technically safe path exists, and the Commission says that is Apple’s choice. China still needs local approval. A keynote cannot flip either switch.
• Making “we have our own GPT-5” the 2026 story: Apple’s public story is still privacy, on-device, PCC. If Ternus reframes the company as a model-arms-race player, he throws away the differentiation he just inherited.
• Ripping out ChatGPT to settle a lawsuit: the case is the case; a second optional backend is a hedge. Users already know how to hand a hard question to someone else.
• Letting Siri become a Gemini wrapper with a logo: the license only works if the product still feels like Apple. PYMNTS had the right sentence: he does not need the best model in the industry; he needs the best products around a model Apple does not own, at a cost that does not crush the margin.

The JSON contract to watch: App Intents

For developers, whether Ternus swaps Gemini is a 2027 question. How Siri AI calls your app is homework for September. The system assistant maps natural language onto structured actions through App Intents: entities, parameters, required fields — a JSON Schema in all but name. Tool declarations in the Foundation Models framework are the same shape as Gemini / OpenAI Function Calling, which we already covered in Tool Calling and JSON Schema validation and OpenAI vs Gemini Structured Output.

Once Gemini makes Siri better at cross-app and multi-turn work, a loose schema fails louder: dates arrive as prose, party size exceeds the cap, a required restaurant id is omitted. The action does not fail because the model is “not smart enough.” It fails because the contract was never tight. The on-device 3B model will not paper over your types either.

{
  "name": "bookRestaurant",
  "description": "Book a restaurant by time and party size",
  "parameters": {
    "type": "object",
    "additionalProperties": false,
    "properties": {
      "restaurantId": { "type": "string", "description": "Stable in-app restaurant id, not a free-text name" },
      "partySize": { "type": "integer", "minimum": 1, "maximum": 12 },
      "date": { "type": "string", "format": "date" },
      "time": { "type": "string", "pattern": "^([01][0-9]|2[0-3]):[0-5][0-9]$" }
    },
    "required": ["restaurantId", "partySize", "date", "time"]
  }
}

Drop that schema into the JSON toolbox and validate it locally. It is the same discipline as tools.parameters / response_schema on the Gemini API. When Siri fills a parameter wrong, inspect required and enum first, and only then blame “Apple’s model.” If Ternus does push more traffic back on-device, the smaller model is less forgiving of a loose schema, not more.

What to do now

1. Ship App Intents before you bet on keynote wording: unregistered actions are invisible to Siri AI. Once English users come off the waitlist after 9 September, apps without Intents get skipped.
2. Write parameters as JSON Schema, not as “the model should get it”: stable ids, integer ranges, date formats, additionalProperties: false. Run the same fixtures through JSON validation and JSON Diff in the browser.
3. Accept against three layers, not a chat transcript: keep a few real paths for on-device shorts, PCC cross-app, and heavy reasoning. The waitlist and the English gate mean you will not see “all users” data for a while.
4. Do not make EU / China iPhones depend on Siri AI: there is no timeline. The core flow must still work with taps.
5. Degrade cloud calls by iCloud tier and daily caps: the bill formula is unpublished, but capacity gating is already real. Timeouts and quota errors need an on-device fallback.
6. Keep the model string in config: today’s Foundation Models cloud lane is Gemini-related; 2027 may add Anthropic or more Apple weights. Swapping a backend must not mean rewriting Intents.

FAQ

Takeaways

Two days into the job, Ternus is still running Cook’s AI skeleton: Apple on device, PCC for privacy, Gemini for the ceiling, ChatGPT as an optional exit. 9 September is an inherited product show, not a new roadmap. What he is likelier to do is replay Apple Silicon — use someone else’s capability, then pull the experience back: keep daily chat off Google, make the iCloud bill explicit, and distill until Apple’s own weights need the license less each year.

For developers, the 2026 job is not guessing whether the CEO changes vendors. It is writing App Intents and tool JSON that survive a schema check. The more capable Siri gets, the more a loose parameter costs. Treat EU and China iPhones as if Siri AI is absent, and keep a tap-only path. This site already covers Gemini structured output and Tool Calling validation; this article only adds the Apple-side strategy read.