TheChek News Flow

Apple could deepen its reliance on Google for cloud services after years of underinvesting in its own infrastructure.

The Takeaway

Now there’s a chance Apple could deepen its cloud dependence on Google even more. At Apple’s request, Google has investigated setting up servers inside its data centers to run an upcoming version of Siri, Apple’s digital assistant, that will be powered by Gemini while also abiding by Apple’s privacy standards, said two people with knowledge of the talks. Apple already relies on Google Cloud for services like online storage and training of Apple’s in-house AI models.

In the past, Apple has said it sends complex AI queries by its users to an Apple system called Private Cloud Compute that runs on Apple servers equipped with its own custom silicon chips (Apple processes simpler AI queries on its devices).

For years, Apple aspired to be self-sufficient in cloud computing. Various leaders at the company have attempted to build up its internal infrastructure to reduce its reliance on cloud providers like Google and Amazon Web Services.

But those cloud efforts have been stymied in the past by Apple finance executives that treated cloud computing as an irksome cost center rather than a strategic priority, according to roughly a dozen former Apple executives and engineers. Apple’s resistance to making big investments in its own infrastructure has led to ongoing departures of cloud experts from the company, some of those people said.

In recent years, as questions have mounted about Apple’s lack of data center spending in comparison to others, the company has told investors that its “hybrid” approach to its infrastructure needs—which combines public cloud services and Apple-operated infrastructure—is serving it well.

Apple’s cloud difficulties parallel and in many ways are linked to its stumbles in AI. Over the past year, it has struggled to deliver the AI-powered overhaul of its Siri digital assistant on time, and the AI features it has delivered have received a lackluster response from the public.

Reflecting that struggle, only 10% of Apple’s Private Cloud Compute capacity is in use on average, according to former employees. The usage is low enough that some servers intended for Apple’s AI cloud are still in warehouses and haven’t been installed, the people said. Apple’s need for AI compute could change very quickly, though, if the new version of Siri, which Apple says will debut this year, takes off with users. That could explain why it’s talking with Google about hosting the assistant.

Apple’s reliance on Google in cloud computing stands in contrast to the company’s focus on controlling many of the most essential ingredients in its products. It is well known for designing the most important parts of those products, including the hardware, software and chips, so it can give its devices an edge over those of rivals.

At the same time, Apple is also known for its tightfistedness, especially when it comes to making the huge up-front capital expenditures required for efforts like building data centers. While Meta Platforms, Microsoft, Google, Amazon and others have gone on history-making spending sprees in recent years to construct data centers that accommodate soaring AI use, Apple has largely sat on the sidelines.

Instead, its financial leaders are more comfortable renting AI computing capacity and other services from outside cloud providers. To be sure, owning your own data centers isn’t a prerequisite to be considered a leader in AI, as OpenAI and Anthropic have shown. But at some point, they may regret depending so much on external firms if the cloud providers choose to raise costs, potentially forcing the AI companies to build out more of their own internal server capacity.

Perhaps the bigger issue, said former Apple cloud engineers, is that Apple’s culture still revolves around selling devices, which accounts for the vast majority of its sales. That remains true even as it has begun to make more money from music, its App Store and other services, all of which run either in its own cloud infrastructure or on servers it rents from other providers.

Apple’s messy back-end internal infrastructure only adds to its difficulties. Many parts of the company use their own servers or various cloud offerings, a contrast to other large tech firms like Google that have a single pool of computing resources for engineers to pull from.

“Apple and Google have two very different engineering cultures,” said Igor Naverniouk, who left Apple in December after working on a next-generation version of Siri and previously spent more than a decade working on Google’s infrastructure. “At Google, most things are centralized. Everybody is using the same supercomputer. At Apple, technology choices are siloed.”

Controlling Costs

Apple’s cloud troubles go back decades.

With the launch of the iTunes Music Store in 2003, the company began expanding its small data center footprint to handle what would become a booming new business in selling digital music.

As Apple began expanding its online offerings, it did so using a hodgepodge of different systems to run them. For example, it built iTunes Genius, a personalized playlist feature it launched in 2008, on technologies and servers that were separate from those of iTunes. The reason: Apple wanted to anonymize the data from its analysis of users’ music libraries.

Around that time, Apple began to rely increasingly on public cloud services, a then-novel concept pioneered by Amazon Web Services that involved renting out its huge data centers to external business customers. AWS was an early provider of online storage capacity for iCloud, the storage service Apple launched in 2011.

But when Apple wanted to release features for iCloud photo backup, it struggled with the huge costs of storing data in the public cloud as droves of iPhone users began snapping pictures. To save money, it began building out its own servers to back up iCloud photos. Doing so had other benefits: When Apple told AWS what it was up to, Amazon slashed the cloud fees it was charging the company by half, a former Apple executive involved in the process said. Apple continued to use its own servers, along with those of cloud providers.

By 2013, Apple’s finance department began getting frustrated with the company’s ballooning costs from both the public cloud and its own servers. Managers from that department began asking members of Apple services teams if they were making full use of the internal infrastructure the company already had at their disposal.

It became clear those teams had not coordinated with each other on their server efforts, resulting in duplicate infrastructure and unused capacity, former Apple engineers said. For example, when iTunes’ servers had unused cloud capacity, other Apple teams couldn’t tap into it, said the people.

In 2013, Apple tasked Patrick Gates, then a director of engineering at the company, with figuring out how to consolidate its sprawling server infrastructure into a common pool every team could access. He began leading a new group, dubbed Platform Infrastructure Engineering, to build that pooled resource, modeling it on how modern cloud systems worked at Amazon and Google.

‘ACDC’

But Gates struggled to get various parts of the company working on the centralized platform and left Apple in 2018, former Apple engineers said. In 2019, Mike Abbott, a former vice president of engineering at Twitter and an early engineering leader of Microsoft’s cloud, took charge of what was left of the Platform Infrastructure Engineering group to continue the project.

Abbott sought to instill a more cloud-oriented culture at Apple. In 2021, he started a virtual event inside Apple, Infrastructure Summit, with the goal of encouraging the entire company to collaborate on shared infrastructure.

He pushed other new ideas, most notably Project ACDC, an effort to bring Apple’s chip expertise to its servers for its own data centers. The project—the name of which stood for “Apple chips in data center”—began as a means to conform the data centers running Apple online services to the same strict privacy standards Apple used on its devices.

He also proposed that Apple consider eventually renting out its servers to external developers, similar to the public cloud services Amazon and Google sell, The Information previously reported. Johny Srouji, Apple’s chip chief, was a big backer of the project, seeing it as an opportunity to offer the chips his team had built to business customers, a former leader of the project said.

But Abbott struggled to move his various projects ahead due to pushback from Apple’s finance team, which didn’t want to invest in more Apple-owned cloud services because not enough of the company’s existing server capacity was in use, former employees on the team said. Relying on outside cloud vendors would give Apple more control over its infrastructure costs, the finance team believed.

In 2023, Abbott left Apple and joined General Motors. His departure soon prompted a cloud brain drain from Apple as many of the people he had hired soon followed him to the automaker.

Warming to Google Cloud

The launch of ChatGPT 3.5 in late 2022 was a seismic moment for the entire tech industry. It also shook up Apple’s ideas about AI and the cloud.

Before ChatGPT came out, Apple’s priority with Siri was making it run as efficiently as possible so it could handle user interactions on Apple devices rather than in the cloud. That approach would better protect the privacy of user data, Apple felt. After ChatGPT showed the possibilities of huge cloud-based models, Apple realized it couldn’t stick with its old approach and needed to tap the power of the cloud more.

There was a problem, though: Apple’s internal AI infrastructure was beginning to decay. In 2023, the company was in the process of decommissioning a large number of older Nvidia chips in its data center that were starting to fail, people involved in the effort said. This was a job the company had wanted to do for years, but it had fallen by the wayside. Apple needed to replace the older Nvidia chips with newer models better suited to the latest AI.

Pushed by its finance team, Apple chose instead to move ahead with its AI plans primarily using outside cloud providers, as it had done for other services such as storage. AWS, already a longtime Apple cloud provider, became the company’s first big partner in its AI push. Apple was also an early customer for Amazon’s in-house alternatives to Nvidia AI chips, Inferentia and Trainium2.

Partnering with Google proved trickier. For years, Apple had banned its AI engineers from using Google’s cloud because of privacy concerns. Siri handled personally identifiable information from users of Apple devices, and Apple didn’t want to run any risk of exposing such data to outside companies. Apple software chief Craig Federighi, who also acts as Apple’s de facto privacy czar, repeatedly vetoed Google Cloud as an option for its AI compute needs.

But in 2023, Google made updates to its security systems that satisfied Apple’s privacy concerns. Apple quickly began adopting Google Cloud for its AI needs, including using Google’s custom chip, the tensor processing unit—which Apple calculated was significantly less expensive to operate than Nvidia’s comparable chips.

Private Cloud

Apple needed to show the world it was taking AI seriously. ChatGPT’s explosive growth had forced Google, Amazon and nearly every major tech company to come up with their own plans to deliver more intelligent, conversational forms of AI to users. At Apple’s annual Worldwide Developers Conference in 2024, the company made its move, announcing Apple Intelligence, a suite of AI tools based on the types of generative models popularized by ChatGPT.

Apple tasked the team working on Project ACDC, the cloud effort previously led by Abbott, to help out with the Apple Intelligence launch. While that project wasn’t initially AI focused, the team working on it rushed to build Private Cloud Compute, code-named Project Thimble, to create a server system that would power Apple’s next-generation AI products more privately. Although Apple announced Private Cloud Compute alongside Apple Intelligence in June 2024, the system wasn’t actually working yet and was six months behind schedule, said former employees who worked on it (it finally launched in late 2024).

In the following months, Apple began trickling out parts of Apple Intelligence—such as AI writing tools and notification summaries—but the public and technology critics mostly greeted the features with disappointment. A bigger issue for Apple was that an overhauled, more conversational version of Siri simply wasn’t ready to ship.

Apple’s talks with Google about hosting Siri could be a sign that Apple wants to be prepared for a surge in AI activity on its devices when the new Siri comes out later this year.

Another factor in those talks could be that Private Cloud Compute hasn’t worked very well so far inside Apple’s own data centers. Updating software on its AI servers can take much longer than on other kinds of servers, said former employees. In addition, the Apple chips on Private Cloud Compute servers weren’t designed for AI and are not well equipped to run big models like Google’s Gemini, said former Apple cloud engineers and AI employees.