In January, Apple announced it would use Google Gemini models in its products, an admission that the iPhone maker had been unable to compete on its own in artificial intelligence over the past few years. But the agreement also shows that Apple is once again relying on outside companies to make up for another shortcoming: cloud computing capabilities.
Now, Apple is likely to further deepen its dependence on Google Cloud. At Apple's request, Google has looked into deploying servers within its data centers to run upcoming versions of Siri, according to two people familiar with the negotiations. The Gemini-powered digital assistant will also comply with Apple's privacy standards. Apple currently relies on Google Cloud for services such as online storage and internal AI model training.
In the past, Apple has stated that it will send complex AI query requests from users to an Apple system called Private Cloud Compute, which runs on servers equipped with Apple's self-developed chips (simple AI queries are processed locally on the device).
For years, Apple has aspired to become self-sufficient in cloud computing. Many company executives have tried to build internal infrastructure to reduce dependence on cloud service providers such as Google and Amazon Cloud Technology.
But those cloud-related efforts have been repeatedly stymied in the past because Apple finance executives viewed cloud computing as a nuisance cost center rather than a strategic priority, according to more than a dozen former Apple executives and engineers. Apple's reluctance to invest heavily in its own infrastructure has led to a steady stream of departures of cloud experts, these people said.
In recent years, there have been growing doubts that Apple has underinvested in data centers compared with its peers. Apple has told investors that its "hybrid" infrastructure strategy - combining public cloud services with Apple's own infrastructure - is working well.
Apple's difficulties in cloud computing parallel its setbacks in the field of AI and are related to each other in many ways. In the past year, Apple has struggled to launch a comprehensive overhaul of Siri powered by AI on time, and the AI features that have been launched have also received mediocre response.
Former employees revealed that the average utilization rate of Apple's private cloud computing is only 10%, and the utilization rate is so low that some servers prepared for Apple's AI cloud are still stored in warehouses and have not yet been installed. However, if Apple claims that the new version of Siri launched this year is popular with users, its demand for AI computing power may surge rapidly, which may also explain why Apple is discussing with Google about hosting assistants.
Apple's reliance on Google for cloud computing is in sharp contrast to the company's insistence on controlling the core components of its products. Apple is known for designing its own hardware, software, chips and other key parts to give its devices an advantage over competing products.
At the same time, Apple is also known for being prudent with money, especially when it comes to projects like building data centers that require huge upfront capital expenditures. In recent years, companies such as Meta, Microsoft, Google, and Amazon have made unprecedented investments in building data centers to adapt to the explosion of AI demand, while Apple has basically stayed out of the way.
On the contrary, Apple's financial executives prefer to rent AI computing power and other services from external cloud service providers. Granted, as OpenAI and Anthropic have proven, owning your own data centers is not a prerequisite for becoming an AI leader. But once cloud service providers choose to raise prices, over-reliance on external companies may make them regret it, and AI companies may be forced to build more servers by themselves.
Former Apple cloud engineers say that perhaps a more serious problem is that Apple's culture still revolves around device sales, which account for the vast majority of its revenue. That remains true even as Apple makes more revenue from music, the App Store and other services that either run on its own cloud infrastructure or rent servers from other vendors.
Apple's messy back-end internal infrastructure further compounded the woes. Multiple departments within the company use their own servers or various cloud services, in contrast to the model of other large technology companies such as Google that use a unified computing resource pool for engineers to call.
Igor Navnyuk, who participated in the development of the next generation of Siri and left in December last year, previously worked in Google's infrastructure department for more than ten years. He said: "The engineering cultures of Apple and Google are completely different. Most of Google's systems are centralized, and everyone uses the same supercomputer; while at Apple, technology choices are separated from each other and fought independently."
Strictly control costs
Apple's cloud woes date back decades.
After the iTunes music store was launched in 2003, Apple began to expand the scale of small data centers to support the booming new business of digital music sales.
As Apple's online services continue to expand, it uses a variety of systems to operate them separately. For example, the personalized playlist function iTunes Genius launched in 2008 is built on technology and servers independent of iTunes because Apple hopes to anonymize user music library analysis data.
Around the same time, Apple began relying increasingly on public cloud services—a then-novel concept pioneered by Amazon Cloud Technologies that rented out large data centers to external enterprise customers. AWS was an early online storage provider for Apple's iCloud storage service launched in 2011.
But when Apple planned to launch the iCloud photo backup function, as a large number of iPhone users took photos, the high cost of public cloud storage overwhelmed the company. To save money, Apple began to build its own servers for iCloud photo backup. The move also had another benefit: Amazon cut the amount it charges Apple for cloud services in half after Apple informed AWS of its plans, according to a former Apple executive involved in the project. Apple has since continued to use both its own servers and cloud service provider servers.
By 2013, Apple's finance department was dissatisfied with the soaring costs of the public cloud and its own servers, and began asking the service team whether it was making full use of the company's existing internal infrastructure.
Former Apple engineers said that these teams apparently lacked coordination on server deployment, resulting in duplication of infrastructure and idle resources. For example, when the iTunes server has idle cloud capacity, other Apple teams cannot use it.
In 2013, Apple assigned then-engineering director Patrick Gates to consolidate its disparate server infrastructure into a common pool of resources that all teams could use. He took the lead in establishing a new department "Platform Infrastructure Engineering" to build shared resources based on the modern cloud system models of Amazon and Google.
ACDC project
But according to former Apple engineers, Gates struggled to push various departments in the company to use this centralized platform, and eventually left in 2018. In 2019, Mike Abbott, former vice president of engineering at Twitter and head of early engineering for Microsoft Cloud, took over the department and continued to advance the project.
Abbott is trying to create a more cloud-oriented culture within Apple. In 2021, he launched the "Infrastructure Summit" online event within the company, aiming to promote company-wide collaboration on shared infrastructure.
He also promoted a number of new initiatives, the most famous of which was the ACDC project, which applied Apple chip technology to its own data center servers. The project's full name is "Apple chips in data center" (data centers use Apple chips). Its original intention is to enable data centers running Apple's online services to meet the same strict privacy standards as Apple devices.
According to previous reports, he also proposed that Apple consider eventually leasing servers to external developers, similar to Amazon and Google's public cloud services. A former person in charge of the project said that Apple chip chief Johnny Srouji was a strong supporter of the project and saw it as an opportunity to provide corporate customers with the chips developed by his team.
However, former employees of the team said that many of Abbott's projects were difficult to advance due to opposition from the financial department. The finance team believes that existing servers are underutilized and is unwilling to invest in additional Apple's own cloud services, and believes that relying on external cloud service providers can better control infrastructure costs.
In 2023, Abbott left Apple to join General Motors. His departure soon triggered a brain drain in Apple's cloud field, and many of the employees he recruited followed him to join the car company.
Moving to Google Cloud
The advent of ChatGPT 3.5 at the end of 2022 will be a disruptive moment for the entire technology industry, and it will also completely change Apple’s thinking on AI and cloud.
Before the launch of ChatGPT, Apple's core goal for Siri was to run as efficiently as possible, giving priority to processing user interactions locally on the device rather than in the cloud. Apple believed this could better protect user data privacy. After ChatGPT showed great potential based on cloud-based large models, Apple realized that it could not stick to the old model and needed to rely more on cloud computing power.
But a problem ensued: Apple’s internal AI infrastructure began to age. Relevant sources said that in 2023, the company will begin to decommission a large number of old NVIDIA chips that are gradually failing in the data center. This work has been planned many years ago, but has been shelved. Apple needs to replace these old chips with new models that are more suitable for the latest AI technology.
Driven by the financial department, Apple chose to rely mainly on external cloud service providers to promote its AI plan, just as it does with other services such as storage. Long-term cloud service provider AWS has become Apple's first important partner in AI layout, and Apple is also an early customer of Amazon's self-developed Nvidia AI chip replacements Inferentia and Trainium2.
Cooperation with Google is more tortuous. For years, Apple banned AI engineers from using Google Cloud due to privacy concerns. Siri handles the personally identifiable information of Apple device users, and Apple does not want to take any risk of data being leaked to outside companies. Craig Federighi, the software director who actually serves as Apple's privacy director, has repeatedly rejected plans to use Google Cloud to meet AI computing power needs.
But in 2023, Google upgraded its security system to meet Apple's privacy requirements. Apple quickly began to use Google Cloud to meet its AI needs, including using Google's customized chip tensor processing unit - Apple calculated and found that its operating costs are much lower than similar chips from Nvidia.
private cloud
Apple needs to prove to the outside world that it is taking AI seriously. The explosive growth of ChatGPT has forced Google, Amazon and almost all major technology companies to launch smarter and more conversational AI products. At the 2024 Apple Worldwide Developers Conference, Apple finally took action and announced the launch of Apple Intelligence, an AI toolset based on the generative model popularized by ChatGPT.
Apple assigned the ACDC project team previously led by Abbott to assist in the launch of Apple Intelligence. Although the project was not initially focused on AI, the team urgently built a private cloud computing system codenamed "Project Thimble" to provide more private computing power support for Apple's next generation AI products. Former employees involved in the project said that although Apple announced private cloud computing with Apple Intelligence in June 2024, the system was not actually operational at that time, and the schedule was six months behind (it was finally launched at the end of 2024).
In the following months, Apple successively launched some features of Apple Intelligence, such as AI writing tools, notification summaries, etc., but the public and technology critics were generally disappointed. A more serious problem for Apple is the delay in launching a completely revamped, more conversational version of Siri.
Apple’s talks with Google about Siri hosting could be a sign that Apple wants to prepare for a surge in on-device AI activity when a new version of Siri launches later this year.
Another reason behind the talks could be that private cloud computing doesn't work well within Apple's own data centers. Former employees said that software updates for its AI servers took much longer than other types of servers. In addition, Apple cloud engineers and former employees of the AI department said that Apple chips on private cloud computing servers are not designed for AI and cannot run large models such as Google Gemini well.