Apple Reveals Google and Nvidia Power Its Most Advanced AI Cloud Model

At its annual Worldwide Developers Conference in Cupertino, California, Apple showed off something that no one had seen before: its most advanced AI model, called Apple Foundation Model Cloud Pro, runs on Nvidia GPUs that are set up in Google's cloud infrastructure, but Apple's end-to-end privacy architecture is still in place.

Apple also released a much improved Siri that can have back-and-forth conversations, complete jobs with multiple steps, and access personal information like calendars and location. All of this is done through a privacy-protecting system Apple calls the "system orchestrator."

Tim Cook's last developer meeting as CEO was WWDC. In September, he will become Executive Chairman and hand the stage over to John Ternus, who has been an executive for a long time and will become CEO in October.

The Three-Way Architecture Nobody Saw Coming

When Apple and Google announced their relationship in January, everyone in the business world thought Apple would use Google's Gemini models as a simple way to add AI from a third party, which is what other companies have done. In more than one way, that belief was wrong.

Amar Subramanya, an AI executive at Apple, revealed the real structure on Monday: AFM Cloud Pro is not driven by public Gemini. Apple's third-generation AFM cloud models are "trained using proprietary data with reinforcement learning and refined using outputs from Gemini frontier models." Google's technology was instead used to help train Apple's own models.

Then, Apple made a way for those models to run on Nvidia GPUs in Google's cloud infrastructure, while still following Apple's privacy rules instead of standard cloud data practices.

The relationship with Nvidia on hardware needed its own big step forward in engineering. Sebastian Marineau-Mes, VP of software, said that Apple wanted to use Nvidia's newest chips but needed them to be set up in a more private way, so the chips couldn't read what was on the servers they were handling. An up-and-coming technology from Nvidia called "ambiguous confidential compute" made the system possible.

"We wanted to avail ourselves of the latest technology from Nvidia, and so we set out to extend private cloud compute to third-party cloud," Marineau-Mes stated.

Apple's custom-trained models run on Nvidia hardware inside of Google's infrastructure. The data being handled can't be read by either Google or Nvidia because of Nvidia's confidential compute technology and Apple's Private Cloud Compute architecture.

Federighi said that the system orchestrator, which is the software that sends questions to the right model, was "key to the privacy architecture of our whole system."

The Strategy Apple Is Explicitly Rejecting

The way Apple executives decided to talk about the news was just as sharp as the technology itself. In the beginning, SVP Craig Federighi talked about how Apple's approach is different from that of rivals who are "racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people that it's ultimately meant to serve." It was clear that the speaker was talking about competitors who spend hundreds of billions of dollars on infrastructure.

Analysts have praised Apple's slow pace in a certain way, saying that the company didn't make the huge capital expenditure promises that have hurt Microsoft, Google, Amazon, and Meta's profit margins.

The irony that Apple needed infrastructure from both Google and Nvidia in the end to deliver its most advanced features doesn't completely destroy that argument, but it does change the way it's put together. Apple isn't skipping spending on AI technology; it's just using it selectively and on its own terms, since it doesn't own the data centers.

Apple is not building model power as a way to set itself apart, but privacy architecture. It is said that AFM Cloud Pro is "comparable to Google's Gemini frontier models," not better. Apple's AI will be chosen by users and businesses because of what it doesn't do with their data, not because of how well it does on benchmarks.

The WWDC Siri Demo and What It Actually Showed

When the remade Siri was shown at WWDC, it worked much better than earlier versions. During the demo, Siri looked up concert times, set a reminder to buy tickets, and helped the user find their way to pick up a friend on the way. This was a multi-step, context-aware interaction that older versions of Siri often failed to complete.

Apple has been promising a big improvement since the 2024 WWDC announcement, but the conversational back-and-forth feature that lets people talk easily and keep the conversation going is still a dream.

Apple's architecture sends queries through the system orchestrator to models that are either on the device or in the cloud, based on how much computing power is needed and what personal data is being sent.

You don't have to leave your phone to answer questions that need to access your calendar, messages, or position. More complex queries or queries that need to use the frontier model go to the cloud, but they do so through Apple's Private Cloud Compute infrastructure instead of normal cloud APIs.

Tim Cook's Final WWDC — and What Ternus Inherits

There was a farewell feel to the gathering. Cook took over as CEO after Steve Jobs died in 2011 and grew Apple's market value from $348 billion to over $4 trillion. He spoke to a group of workers and executives and seemed to wipe away a tear as he called his job "an honor of a lifetime."

Many people think that Apple's move to Ternus in September—an engineer instead of an operator—shows that the company knows that the next level of competition will depend on AI hardware differentiation rather than supply chain success.

Ternus takes over a company that has the largest installed base of privacy-preserving AI hardware in the world—2.5 billion active devices—as well as a strategy that finally makes sense: use that base as infrastructure for on-device and private cloud AI instead of competing on data collection.

The most important question for the next three years will be whether AFM Cloud Pro running on Nvidia chips in Google's cloud can catch up to future models in terms of capabilities while still offering the privacy benefits that set Apple apart.

Apple (NASDAQ: AAPL) investors can be sure that the company's AI plan is real, architecturally complex, and partner-dependent in ways that they hadn't said before WWDC.

Both Nvidia (NVDA) and Google (GOOGL) benefit from their relationship with Apple. Nvidia benefits from putting GPUs in Apple's private cloud, and Google benefits from the training data relationship and cloud infrastructure fees. This is because Apple doesn't have to take the competitive risk of running standard Gemini on standard Google Cloud.

The engineering problems with the foldable iPhone that were mentioned earlier this year and the AI strategy's slower-than-expected 2024 rollout are still the risks that come with moving into a Ternus-led era.

But Monday's WWDC gave more architectural detail than expected in response to the question of capability. The privacy differentiation plan is now operationally real instead of just an idea.