In January of this year, Apple announced its intention to integrate Google's Gemini large language model into its products, a move that effectively acknowledged the iPhone maker's inability to compete independently in the artificial intelligence arena over recent years. However, this agreement also signaled that Apple is once again turning to external firms to address another weakness: its cloud computing capabilities.
Apple may now be poised to deepen its reliance on Google Cloud. According to two individuals familiar with the discussions, Google has been exploring the deployment of servers within its data centers to operate an upcoming version of Siri, following a request from Apple. This Gemini-powered digital assistant would also adhere to Apple's stringent privacy standards. Apple already depends on Google Cloud for services such as online storage and the training of its internal AI models.
Historically, Apple stated that it would route complex AI queries from users to its proprietary system, known as Private Cloud Compute, which operates on servers equipped with Apple's own custom chips. Simpler AI queries are processed directly on the device.
For many years, Apple has aspired to achieve self-sufficiency in cloud computing. Several senior executives have attempted to build internal infrastructure to reduce dependence on cloud service providers like Google and Amazon Web Services.
However, according to over a dozen former Apple executives and engineers, these cloud-related initiatives have repeatedly encountered obstacles because the company's financial leadership viewed cloud computing as a burdensome cost center rather than a strategic priority. This reluctance to invest heavily in proprietary infrastructure reportedly led to a steady departure of cloud computing specialists.
Amid growing external skepticism about Apple's comparatively lower investment in data centers versus its peers, Apple has assured investors that its "hybrid" infrastructure strategy—combining public cloud services with its own infrastructure—is performing effectively.
Apple's challenges in cloud computing have coincided with its struggles in AI, and the two issues are interconnected in many ways. Over the past year, Apple has faced difficulties in launching a comprehensive, AI-driven overhaul of Siri on schedule, and the AI features it has released have received a lukewarm response.
Former employees revealed that Apple's Private Cloud Compute infrastructure has been operating at an average utilization rate of just 10%. The usage is so low that some servers intended for Apple's AI cloud remain uninstalled in warehouses. Nevertheless, if the new version of Siri, which Apple has promised for later this year, gains significant user adoption, the demand for AI computing power could surge rapidly. This potential spike may explain why Apple is negotiating with Google to host the assistant.
Apple's reliance on Google for cloud services stands in stark contrast to its longstanding philosophy of controlling core components of its products. The company is renowned for designing its own hardware, software, and chips to create competitive advantages for its devices.
Concurrently, Apple is known for its fiscal prudence, particularly regarding large upfront capital expenditures, such as building data centers. While companies like Meta, Microsoft, Google, and Amazon have made unprecedented investments in data centers to accommodate the AI boom, Apple has largely remained on the sidelines.
Instead, Apple's financial leadership has preferred renting AI computing power and other services from external cloud providers. While owning data centers is not a prerequisite for AI leadership, as demonstrated by firms like OpenAI and Anthropic, over-reliance on external companies could become problematic if those providers decide to raise prices, potentially forcing AI companies to build more of their own server capacity.
A former Apple cloud engineer suggested that a more fundamental issue may be that Apple's corporate culture remains centered on device sales, which constitute the bulk of its revenue. This focus persists even as the company generates increasing revenue from music, the App Store, and other services—services that run either on its own cloud infrastructure or on servers leased from other providers.
The situation is further complicated by Apple's fragmented internal backend infrastructure. Various internal departments operate their own distinct servers or use different cloud services, unlike the centralized, shared resource pools commonly utilized by engineers at other large tech firms like Google.
Igor Nafanyuk, a former Google infrastructure employee with over a decade of experience who worked on the next-generation Siri project at Apple before leaving in December, noted the stark contrast: "The engineering cultures at Apple and Google are completely different. Most systems at Google are centralized; everyone uses the same supercomputer. At Apple, technology choices are siloed and fragmented."
**Cost Control** Apple's cloud computing challenges trace back decades.
Following the launch of the iTunes Music Store in 2003, Apple began expanding its small-scale data centers to support the burgeoning new business of digital music sales.
As Apple's online services expanded, they were operated on a variety of disparate systems. For instance, the iTunes Genius personalized playlist feature, introduced in 2008, was built on technology and servers separate from the main iTunes platform because Apple wanted to anonymize data analyzed from users' music libraries.
Around the same period, Apple began increasingly relying on public cloud services—a novel concept pioneered by Amazon Web Services, which involved renting out capacity in large data centers to external business customers. AWS was an early storage provider for iCloud, Apple's cloud storage service launched in 2011.
However, when Apple planned to introduce iCloud Photo backup, the high cost of public cloud storage for the vast number of photos uploaded by iPhone users became a significant burden. To reduce expenses, Apple started building its own servers for iCloud Photo backup. This move had an additional benefit: according to a former Apple executive involved in the project, after Apple informed AWS of its plans, Amazon halved the cloud service fees it charged Apple. Apple subsequently continued using a mix of its own servers and those from cloud providers.
By 2013, Apple's finance department, dissatisfied with the soaring costs associated with both public cloud and owned servers, began questioning service teams about whether they were fully utilizing the company's existing internal infrastructure.
Former Apple engineers indicated that these teams clearly lacked coordination in server deployment, leading to redundant infrastructure and idle resources. For example, when iTunes servers had spare cloud capacity, other Apple teams could not access it.
In 2013, Apple assigned Patrick Gates, then a director of engineering, the task of consolidating the disparate server infrastructure into a shared resource pool accessible to all teams. He led the formation of a new division, "Platform Infrastructure Engineering," aimed at building a shared resource system modeled after the modern cloud systems of Amazon and Google.
**The ACDC Project** However, according to former Apple engineers, Gates struggled to persuade various company departments to adopt this centralized platform and ultimately left the company in 2018. In 2019, Mike Abbott, former Vice President of Engineering at Twitter and an early engineering leader for Microsoft's cloud, took over the department and continued the project.
Abbott attempted to foster a more cloud-oriented culture within Apple. In 2021, he initiated an internal "Infrastructure Summit" to promote company-wide collaboration on shared infrastructure.
He also championed several new initiatives, the most notable being the ACDC project—"Apple chips in data center." The project's goal was to equip servers in Apple's own data centers with Apple's chip technology, aiming to bring the same rigorous privacy standards applied to Apple devices to the data centers running its online services.
As previously reported, he also proposed that Apple consider eventually renting out server capacity to external developers, similar to the public cloud services offered by Amazon and Google. A former leader of the project stated that Johnny Srouji, Apple's head of chips, was a strong proponent, viewing it as an opportunity to offer the chips developed by his team to enterprise customers.
However, former team members said that Abbott's various projects faced significant hurdles due to opposition from the finance department. The financial team, citing low utilization rates of existing servers, was reluctant to approve further investment in Apple's own cloud services and believed that relying on external cloud providers offered better control over infrastructure costs.
Abbott left Apple for General Motors in 2023. His departure quickly triggered a talent drain in Apple's cloud division, with many employees he had recruited subsequently following him to the automaker.
**Shift to Google Cloud** The release of ChatGPT 3.5 in late 2022 was a disruptive moment for the entire tech industry and fundamentally altered Apple's approach to both AI and cloud computing.
Before ChatGPT, Apple's primary goal for Siri was to operate as efficiently as possible, prioritizing on-device processing over cloud-based processing, which Apple believed better protected user data privacy. After ChatGPT demonstrated the vast potential of cloud-based large language models, Apple realized it could not adhere to the old model and needed to leverage more cloud computing power.
A problem emerged: Apple's internal AI infrastructure was aging. Sources indicated that in 2023, the company began decommissioning a large number of aging, failing Nvidia chips in its data centers—a task planned years earlier but repeatedly postponed. Apple needed to replace these older chips with newer models better suited for the latest AI technologies.
Driven by the finance department, Apple opted to rely primarily on external cloud providers for its AI initiatives, mirroring its approach for other services like storage. Long-time cloud provider AWS became Apple's first major partner for its AI push, with Apple also becoming an early customer for Amazon's custom AI chips, Inferentia and Trainium2, designed as alternatives to Nvidia's.
The partnership with Google was more complex. For years, Apple had prohibited its AI engineers from using Google Cloud due to privacy concerns. Since Siri handles personally identifiable information from Apple device users, Apple was unwilling to assume any risk of data exposure to an external company. Craig Federighi, Apple's head of software who effectively oversees privacy, repeatedly vetoed proposals to use Google Cloud for AI computing needs.
However, in 2023, Google upgraded its security systems to meet Apple's privacy requirements. Apple quickly began adopting Google Cloud for its AI needs, including the use of Google's custom Tensor Processing Units—which Apple calculated offered significantly lower operating costs compared to equivalent Nvidia chips.
**Private Cloud** Apple needed to demonstrate to the world that it was taking AI seriously. The explosive growth of ChatGPT pressured Google, Amazon, and nearly every other major tech company to launch smarter, more conversational AI products. At Apple's Worldwide Developers Conference in 2024, the company finally responded by announcing Apple Intelligence—a suite of AI tools based on the generative models popularized by ChatGPT.
Apple tasked the team behind the ACDC project, previously led by Abbott, with assisting the rollout of Apple Intelligence. Although the project was not initially focused on AI, the team urgently developed a system called Private Cloud Compute, internally codenamed "Project Thimble," to provide more private computational support for Apple's next-generation AI products. Participants in the project revealed that although Apple announced Private Cloud Compute alongside Apple Intelligence in June 2024, the system was not yet operational at that time, being approximately six months behind schedule. It eventually launched by the end of 2024.
In the following months, Apple gradually released some Apple Intelligence features, such as AI writing tools and notification summaries, but the public and tech critics were largely disappointed. A more significant problem for Apple was the continued delay in launching a comprehensively revamped, more conversational version of Siri.
The discussions between Apple and Google regarding hosting Siri may indicate that Apple is preparing for a potential surge in device-side AI activity following the release of the new Siri version later this year.
Another reason behind the talks could be that the Private Cloud Compute system has not been performing optimally within Apple's own data centers. Former employees reported that software updates for its AI servers take considerably longer than for other server types. Furthermore, former Apple cloud engineers and AI team members stated that the Apple chips used in Private Cloud Compute servers were not specifically designed for AI and struggle to run large language models like Google's Gemini effectively.
Comments