The Deal That Happened Instead
The structure of the Apple–Google deal is the part nobody called correctly in 2025. It isn't a licence for off-the-shelf Gemini. It's a custom model built specifically for Apple, hosted on Apple's Private Cloud Compute, running behind the Apple Intelligence branding users already see.
The commercial terms, as reported by Tech Insider, are roughly $1 billion per year for the inference capacity and the model weights. That's a rounding error on Apple's services line, but it's a structural admission: Apple has decided the 2026–2028 window is not the one in which it closes the foundation model gap on its own.
Crucially, this is not the end of Apple's internal AI work. The company is still shipping its own 150-billion parameter cloud models for lighter tasks. Gemini slots in for the workloads where scale matters — the “rebuilt Siri” workloads, and the web-search-adjacent ones that used to be Google's home turf.
The 1.2 Trillion Parameter Custom Gemini
The most striking number in the deal is the parameter count. 1.2 trillion parameters is roughly eight times larger than Apple's existing 150B cloud model. It is also larger than any model Apple has publicly indicated it can train and serve itself.
That ratio is the story. Apple's internal models were good enough for on-device summarisation, mail triage, photo captioning and the 2024-era Siri upgrades. They were not — and, on current trajectory, would not be — good enough for the assistant Apple wants Siri to become: one that can hold multi-turn conversations across apps, plan actions, summarise the live web, and operate as a genuine competitor to ChatGPT and Perplexity.
To get to that capability class, you need frontier-scale compute, frontier-scale data, and frontier-scale model architecture. Apple has the first. Google has all three. The $1B/year figure is what it costs Apple to close the gap without waiting three years to train its own 1T+ model from scratch.
Why the Mistral and Perplexity Talks Fell Apart
The original MacRumors reporting from August 2025 had Apple exploring both acquisitions. 9to5Mac confirmed that the conversations were serious enough to reach Tim Cook's desk. Winbuzzer framed it as Apple's response to an internal AI crisis.
Three things seem to have killed both deals:
- Google proximity overshadowed everything. Once a $1B/year custom-model partnership with Google was on the table, the case for buying a $10–30B AI lab became much harder to make internally. Why own a smaller model when you can rent a bigger one?
- Integration risk. Perplexity's product is a consumer-facing brand. Mistral's value is its European research culture. Neither bolts cleanly into Apple's “invisible intelligence inside our products” philosophy.
- Regulatory exposure. A large acquisition in AI in 2025–2026 means at least 18 months of antitrust review on both sides of the Atlantic. Apple wanted Siri fixed before iOS 27. The timelines didn't work.
Eddy Cue reportedly kept pushing for Perplexity even after the Google term sheet was signed — and according to MacRumors' April 2026 follow-up, Apple is still considering pulling in Anthropic or Perplexity for specific features. The door isn't closed. It's just no longer the front door.
“World Knowledge Answers”: Apple's Search Play
The second half of the deal is the part that should worry OpenAI and Perplexity more than the Siri refresh. Apple is launching World Knowledge Answers in Spring 2026 — a web search tool, baked into Apple Intelligence, powered by the custom Gemini model.
World Knowledge Answers is Apple's first real attempt to own the “ask a question, get a synthesised answer” surface that Perplexity pioneered and ChatGPT scaled. The twist: Apple has 2 billion active devices, default placement in Safari and Siri, and a user base that has never needed to install an app to use a new Apple feature.
If World Knowledge Answers ships cleanly with iOS 26.4, Perplexity's distribution advantage — being the search engine you actively chose — goes from a moat to a footnote. And if Apple's interface is good, ChatGPT's mobile search habit is under pressure too.
What Siri Actually Does Now
The first wave of Gemini-powered Siri features is expected with iOS 26.4 in Spring 2026, with the larger relaunch positioned for WWDC 2026. From Apple's communications and the reporting around them, the rebuilt Siri handles:
- Multi-turn conversations that retain context across apps.
- Document and email summarisation at far higher quality than the 150B model could deliver.
- Live web synthesis through World Knowledge Answers.
- Action chaining (“book the restaurant Maya sent me, reply to her, add it to my calendar”) with much lower hallucination rates than the 2024–2025 Siri.
- A fallback to Apple's own on-device and 150B cloud models for anything that doesn't need frontier scale — which keeps latency, cost and privacy surface area down.
This is the Siri that was promised at WWDC 2024 and didn't ship. It's now shipping because Google's model is doing the heavy lifting behind the Apple wrapper.
Pragmatism Over Ownership: What This Says About Apple
For a company whose entire brand is vertical integration, the Gemini deal is uncharacteristic in a way worth dwelling on. Apple built its own chips. Its own modems (eventually). Its own maps. Its own music service. Its own payments rails. Its own ad network. Its own search index, quietly, inside Spotlight.
On AI, it looked at the cost of catching up — the years, the GPUs, the talent war — and decided that on this specific technology, on this specific timeline, ownership was not worth the delay. That's a genuinely new posture from Cupertino.
The takeaway isn't that Apple is weak at AI. It's that Apple has accepted that the frontier of generative AI moves faster than Apple's product cycle can absorb, and that the right strategy for this decade is to integrate the best model available and focus Apple's own work on the layers where differentiation is still possible: silicon, on-device privacy, interface, and distribution.
What This Means for Marketing Teams
If you're running marketing in 2026, Apple's decision is the decision you're about to face — at a much smaller scale, with much less negotiating leverage, and much faster.
You are not going to train your own foundation model. You are not going to build your own AI agent framework from first principles. You are not going to own the whole stack, because the stack is moving too quickly and the frontier is too expensive.
What you can own is the layer Apple is keeping for itself: the integration, the interface, the brand voice, the distribution, and the data feedback loop. Everything else — the model, the embeddings, the search, the summarisation — is rented, and the question is only which vendor and on what commercial terms.
The marketing teams winning in this environment are the ones who have already stopped trying to build proprietary AI and started building proprietary workflows on top of rented AI. That's the shift Anjin exists to make cheap and fast.
Anjin: The Marketing Operating System for a World Where You Rent Intelligence
Anjin is the Marketing Operating System for a world where you rent intelligence rather than build it. If Apple can't justify the cost and timeline of building its own 1.2T parameter model, your marketing team definitely can't justify building bespoke AI tooling either. The winning move — the Apple move — is to integrate the best available models behind a single operational layer that is yours.
Anjin is that layer for marketing. A single Marketing Operating System that handles content generation, campaign planning, channel distribution, SEO, performance tracking and brand consistency — running on top of frontier models you'd never sensibly train yourself.
What Anjin replaces:
- Your content agency (drafts, revises, publishes across channels).
- Your SEO consultant (monitors, optimises, rewrites continuously).
- Your paid media planner (briefs, tests, reports).
- The seven disconnected AI subscriptions your team currently expenses.
- The £8–15k/month of coordination overhead holding it all together.
What Anjin does that none of them can:
- Runs 24/7 on the same class of models Apple just paid $1B a year to use.
- Learns your brand voice in hours, not months.
- Ships the campaign the same day a news moment breaks — the way Apple is about to ship Siri features the same quarter Google pushes new Gemini weights.
Apple's answer to the AI gap was integrate, don't build. Anjin is the same answer, for marketing.
Sources: MacRumors (April 2026), Tech Insider, WebProNews, Information Age, MacRumors (Aug 2025), 9to5Mac, Winbuzzer.
