What the most vertically integrated company on Earth just admitted about frontier AI models – and what your build-vs-buy debate should take from it.

On Monday, at his final WWDC keynote before handing the CEO role to John Ternus in September, Tim Cook unveiled the biggest overhaul of Siri in the assistant's 15-year history. The demos were slick. The new system-wide "Search or Ask" interface looks genuinely useful. And underneath all of it sits a brain Apple didn't build.

The rebuilt Siri runs on a custom version of Google's Gemini – a model reported to carry around 1.2 trillion parameters, which Bloomberg says is costing Apple roughly $1 billion a year to license. Apple hasn't published the commercial terms itself, but the partnership was announced jointly with Google back in January, with Apple stating that Google's technology offered the most capable foundation for its next-generation models.

Sit with that for a second. Apple holds roughly $162 billion in cash. It designs its own silicon, writes its own operating systems, and spent a decade telling the world that privacy meant keeping Google at arm's length. It evaluated OpenAI and Anthropic alongside its in-house models, took its time, and arrived at a conclusion that should reframe every enterprise AI conversation happening this quarter:

Building the frontier model is a bad investment, even for Apple. The value sits in owning everything around it.

What Apple bought, and what it deliberately kept

The lazy reading of this deal is "Apple gave up". Plenty of commentators went straight there this week. We think they've misread it, because the interesting part of the arrangement is everything Apple refused to hand over.

The Gemini model runs inside Apple's Private Cloud Compute infrastructure – Apple's servers, Apple's privacy controls, Apple's audit surface. Google supplies the weights; it doesn't get the relationship. Latency-sensitive and privacy-sensitive tasks still run on Apple's own smaller on-device models. The context layer – your emails, your messages, what's on your screen right now – belongs entirely to Apple, and the cross-app actions the new Siri can take are built on Apple's frameworks, not Google's.

So Google gets a licence fee. Apple keeps the customer, the data, the interface, and the workflows. One of those things is a commodity with a price tag. The rest is the business.

The build instinct is expensive, and it usually loses

Almost no enterprise was ever going to train a foundation model from scratch. Fine. But the same instinct shows up at smaller scale, and it burns real money. The bespoke fine-tune that took nine months to ship and was beaten by the next off-the-shelf release within six weeks. The internal "AI platform" that quietly duplicates what three vendors now ship as standard. The "our data is special" argument that, on inspection, means "our data is badly organised".

Frontier models are depreciating assets. Each generation makes the last one look slow and expensive within months, and the depreciation is paid by whoever owns the model – which is precisely why you want someone else owning it. Apple's own rebuild of Siri missed two promised ship dates over two years before the company changed course. If an organisation with that much engineering talent and that much cash couldn't out-build Google on Google's home turf, what exactly is the case for your in-house model being a point of pride?

We made a related argument in our recent piece on AI ROI: returns come from workflow redesign, not from the model itself. Apple has just spent a reported $1 billion a year agreeing with us.

The playbook worth stealing

Apple's decision is more useful as a template than as a headline. Four moves stand out.

Run an actual bake-off

Apple spent an extended evaluation period testing Gemini against OpenAI, Anthropic, and its own models before committing. Most enterprises do the opposite – they adopt whichever model their existing cloud contract makes convenient, then wonder why results disappoint. Evaluate against your real tasks and your real error tolerance, not against leaderboard benchmarks. The right model is the one that fails least on the work you actually do.

Wrap the vendor's model in infrastructure you control

Private Cloud Compute is the masterstroke here. Apple consumes Google's intelligence without surrendering its data path, its privacy posture, or its customer relationship. The enterprise equivalent is a model gateway: one abstraction layer through which every application consumes AI, with your logging, your access controls, and your routing logic. Build that once and swapping providers becomes a configuration change rather than a re-platforming programme.

Keep small models you own for the right jobs

Apple didn't replace everything with Gemini. On-device models still handle the high-frequency, low-complexity work, because routing a trivial request to a 1.2 trillion parameter model is poor economics and poor latency. The same routing discipline applies inside your organisation – we covered the unit economics of this in our article on token economics, and Apple's architecture is that argument running at consumer scale.

Treat context as your proprietary asset

Gemini knows everything in general and nothing about you. Siri's advantage is the calendar, the inbox, the screen – context no rival assistant can reach. Your enterprise version of that is the semantic layer over your data, your customer history, and your operational systems. The model is rented intelligence. Context is the part nobody can rent.

The dependency question nobody at Apple Park wanted to discuss

Renting was the right call. Renting without an exit is a different thing entirely.

Gemini is now the default cloud intelligence on Android and, shortly, on well over a billion iPhones – one company's model sitting between most of the world's smartphone owners and the services they use. Regulators spent years scrutinising the Apple-Google search payments; this arrangement hands them a sequel. Apple has hedged the risk by keeping its on-device models alive, running the workload on its own infrastructure, and (reportedly) structuring the deal so a future migration is possible.

European users got a concrete preview of all this within hours of the keynote. Siri AI launches with restrictions in the EU – initially arriving only on macOS and visionOS while Apple works through its regulatory position – meaning the product making the case for rented intelligence is also the first demonstration that jurisdiction now decides what ships, and where. If Apple can't fully deploy its flagship AI feature in a market of 450 million people, the idea that your own AI rollout can treat regulation as an afterthought doesn't survive contact with the evidence.

That hedge is the discipline to copy. Model concentration is becoming the new lock-in, and it's a more dangerous one than cloud lock-in because the switching costs are buried in prompts, evaluations, and behavioural quirks rather than in egress fees you can at least see on an invoice. If you operate in a regulated European industry, there's a jurisdictional layer on top – the question of whose model, running where, under whose legal authority, follows the same logic we set out in our analysis of the EU's Cloud and AI Development Act, just one layer up the stack. UK financial services firms face a sharper version of the same question: under the critical third parties regime introduced through FSMA 2023, the Bank of England, PRA, and FCA can now designate and directly oversee the suppliers the sector depends on. A single model provider sitting behind your customer-facing AI is precisely the concentration risk that regime was written for – and "the vendor's status page says everything is fine" will not satisfy an operational resilience review.

Apple negotiated from strength because it kept alternatives alive. That discipline costs a fraction of what building a frontier model does, and it's available to every organisation reading this.

Q&A: Build, Buy, or Rent?

Does this mean we should never build or fine-tune our own models?
No. Narrow, specialised models – fraud scoring, document classification, domain-specific extraction – can still justify in-house ownership, particularly where data sensitivity or unit costs demand it. What rarely justifies itself any more is building general-purpose capability that frontier vendors improve faster than you can. Build where you're genuinely differentiated. Rent everywhere else.

We've already invested heavily in a custom model. Is that money wasted?
Not automatically, but it does need an honest re-evaluation. Benchmark it against current frontier models on your actual tasks, including cost per outcome, and be prepared for an uncomfortable answer. Apple wrote off years of internal Siri development because the evidence said to. Sunk cost is not a strategy, and the gap compounds with every vendor release you ignore.

How do we rent a model without recreating vendor lock-in?
Put an abstraction layer between your applications and the model provider, so every system consumes AI through one gateway you control. Keep your evaluation suites and prompts portable, log everything on your side of the boundary, and negotiate contractual exit terms before you're dependent. Apple's Private Cloud Compute is exactly this pattern at consumer scale.

What does "owning the integration" actually mean in practice?
It means the things the model vendor can't supply: a machine-readable context layer over your data, orchestration that routes each task to the cheapest model that can do it reliably, evaluation pipelines tuned to your error tolerance, and workflows redesigned around the capability rather than decorated with it. That's where AI programmes succeed or stall – the model is rarely the deciding factor.

How should we choose between frontier providers?
Test on your work, not on benchmarks. Measure the error tail – the frequency and cost of confident, wrong answers in your specific domain – alongside cost per resolved task. Then weigh jurisdiction and data residency, which matter enormously for regulated European businesses. And re-run the comparison regularly, because the rankings change faster than annual procurement cycles do.

Working Through This With Vertex Agility

The decision Apple made this week – rent the frontier model, own the architecture around it – is the same decision sitting in front of most technology leaders we work with, usually in less glamorous form. Somewhere in your organisation there's a build-vs-buy debate running right now, and the honest answer depends on evaluation evidence, integration architecture, and exit design rather than on instinct.

Our AI Consultancy works with organisations on exactly this: AI strategy and model selection, the gateway and orchestration architecture that keeps providers swappable, and the governance frameworks needed for responsible adoption at enterprise scale. Our Software Consultancy and Data Consultancy practices sit alongside it, building the integration and context layers that turn a rented model into something that actually performs on your work. Having those disciplines in one practice matters, because the model decision and the architecture decision are the same decision – which is the point Apple just spent a billion dollars making.

If you want an independent read on where your organisation stands before that debate resurfaces, our free AI Readiness Mini Audit takes a few minutes. For the fuller conversation, get in touch with us directly below.