Apple AI servers sit mostly idle as company turns to Google for revamped Siri infrastructure

Some Apple AI servers are reportedly sitting unused on warehouse shelves, due to low Apple Intelligence usage - 9to5Mac

Earlier this year, Bloomberg reported that Apple was looking to host the much-delayed new Siri models on Google servers, rather than Apple's Private Cloud Compute. The Information is reporting a similar story this morning, saying that Google has been tasked with running Siri servers inside its data centers, while adhering to Apple's privacy standards. The Information goes on to detail how Private Cloud Compute is not up to task. As well as being underpowered, it is also reportedly underutilized in its current state, with the company only using about 10% of its capacity on average, leading to some already-manufactured Apple servers to be sitting dormant on warehouse shelves. The report explains how Apple's cloud infrastructure is very fragmented. The technologies of different team are run independently, rather than having one centralized pool that any department in the company can draw from. This leads to inefficiencies where parts of the company's stack sit idling, while other parts would use the available server capacity, if they could access it. The Apple finance team has apparently been frustrated about the costs of this duplicate infrastructure. There has apparently been several attempts inside the company to unify everything, but those projects have stalled several times over the last decade. For Private Cloud Compute specifically, the system is described as underpowered and perhaps more trouble than it's worth. Updating the software is apparently trickier and takes time, and more fundamentally the chips (believed to comprise right now of mostly M2 Ultra's) are not powerful enough to run the latest frontier models like Gemini, which the new Siri will be based on. And as the initial crop of Apple Intelligence features hasn't been used as much as Apple expected, the Private Cloud Compute buildout is seen in a negative light. While Apple is expecting much higher demand for the new Siri chatbot features whenever they finally land, the current Private Cloud Compute stack does not seem up to the task of running it. Hence, Apple is in advanced talks with Google to run it inside their data centers instead, a company which already has a lot of experience with mass LLM server buildouts thanks to Gemini. The company already relies on Google's cloud for some iCloud features, like cloud storage. The ever-changing landscape of AI may have forced Apple leadership to change their approach to cloud infrastructure, and may invest more heavily in-house going forward, but implementing those changes is a much longer term trajectory.

Report: Apple Asks Google to Run Siri on Its Servers

Currently, Apple sends its more complex AI queries to Private Cloud Compute, a system that runs on Apple servers using Apple silicon chips. Today, only 10% of Apple's Private Cloud Compute capacity is said to be in use on average. The usage is low enough that some servers intended for Apple's AI cloud system are still in warehouses and have not yet been installed. This could change rapidly upon the launch of the next-generation version of Siri, which could spike Apple's demands for cloud computing. Apple has reportedly suffered from a cultural reluctance to bolster its cloud infrastructure for years, leading to the departure of some key cloud experts from the company, such as Patrick Gates. Gates pioneered the idea of bringing Apple chips to data centers, which later formed the basis of Private Cloud Compute. The company still strongly focuses on hardware devices and consumer features rather than their supporting cloud technologies, despite the growth of services, resulting in a neglect of the need for additional capacity. At the time Apple realized it needed to use the cloud to support its AI efforts, its internal AI infrastructure was "beginning to decay." The company was amid the process of decommissioning old Nvidia-powered servers. Combined with financial pressure, this led the company to increasingly turn to third-party providers like Amazon. For years, Apple banned its AI engineers from Google's cloud technologies because of privacy concerns. Apple software chief Craig Federighi repeatedly vetoed Google Cloud as an option for its AI computing requirements. In 2023, Google made changes to its security systems that satisfied Apple's privacy concerns. Apple then started to adopt Google's cloud infrastructure for artificial intelligence. The issue has been exacerbated by problems with Private Cloud Compute, which takes longer to update than other servers. Moreover, the chips currently used in Private Cloud Compute servers were designed for consumer devices and are not optimized for AI workflows, meaning that they are not well equipped to run large models like Gemini. Apple now wants to be prepared for a potential surge in AI use on its devices when the more powerful, Gemini-based version of Siri debuts later this year, motivating the request for Google to run Siri directly on its servers. See The Information's full report for more.

AppleInsider.com

Years of under-investment in data centers may mean that Apple will increase its reliance on Google when the revamped Siri launches. Or it may not, it's hard to tell from the inconsistent report. Even back in 2021, Apple was Google's largest corporate cloud customer , as the company preferred leasing data centers rather than build up its own network of servers. That's reportedly worked well, even now when Apple Intelligence requires more servers. According to The Information, what servers Apple does own are proving to be vastly underused. Citing unspecified former Apple employees, the report claims that on average, only 10% of Apple's Private Cloud Compute capacity is in use. That's low enough that the report also claims that Apple has servers still waiting in warehouses to be installed. However, Apple is said to be talking about expanding its footprint in Google data centers, and so expanding the relationship it has with that firm about AI. As previously announced, Google Gemini will provide technology for the forthcoming improved Siri. When that updated Siri is released, Apple expects to see a radical increase in Apple Intelligence usage. Consequently, it is reportedly discussing leasing more servers from Google. Not a privacy issue Apple servers in a Google data center are still Apple-controlled. The report says that previously Apple has forbidden the use of Google servers because of privacy concerns. Now, however, improvements to that issue have meant Apple is more willing to use them, perhaps instead of spreading its services across multiple companies. "Apple and Google have two very different engineering cultures," said Igor Naverniouk, a now ex-Apple engineer said to have previously worked on the new Siri. "At Google, most things are centralized. Everybody is using the same supercomputer. At Apple, technology choices are siloed." For the new Siri, Tim Cook has already said that the deal with Google to use Gemini will not see Apple relaxing its privacy systems. "We're not changing our privacy rules," said Cook. "We still have the same architecture that we announced before, which is on device plus Private Cloud Compute." When first announced in 2024, Apple said that Private Cloud Compute ran on Apple Silicon-based servers. Those servers were secure, did not store user data, and used encryption to maintain security. Apple offered, and continues to offer, to allow security experts to test its Private Cloud Compute privacy systems. None of the privacy and security technologies will change with the use of Google servers. The difference will presumably be that the servers do not run on Apple Silicon, and are based in Google data centers. Apple-controlled servers are still controlled by Apple, even if they are housed in a Google facility. Apple's dependence on Google This new report doesn't make any claims about Apple's privacy being weakened by the deal. But it does argue that Apple is weakened by being dependent on third-party servers. It also accuses Apple of being tight-fisted over investment in its own servers. It's highly critical over what was called Project ACDC, a plan to develop processors for data centres. Reportedly, Apple has been working since at least 2013 to get its use of cloud servers into one pool for all of its teams. Between then and 2023, two different people were tasked with centralizing server usage, and then left Apple after their projects failed. Consequently, Apple was left with internal disarray with its teams spending individually on data servers. If one team had excess capacity, still no other team could use it. Plus reportedly, what servers Apple did have were aging and the company had to replace them. It's not clear how the devices produced in the Houston effort factor into all this, and the report does not discuss that at all. So the argument is that Apple should spend more on data centers and do so in a more unified way. But at the same time, the report criticizes Apple for becoming reliant on Google. There's also a presumption is that Apple was too frugal to buy servers and instead rented them, as if renting did not cost money as well. Apple always prefers to do everything itself, to own the whole stack, from design through to manufacturing. That does mean that it renting servers goes against its usual practices. But then the reason to rent is not to save money per se, it is to provide capacity. Apple can increase its server usage as soon as demand requires it, instead of having to wait months or years to build up that capacity itself.

Apple's Baltra ASIC Can't Come Soon Enough As The Vast Majority Of Its Current AI Servers Are Reportedly Rotting On The Shelves

With the hardware side of Apple's AI strategy reportedly in an equal, if not greater, disarray than the software side, it is hardly a surprise that the Cupertino giant has been forced to rely on Google's servers along with its very capable Gemini model to ship the revamped Siri and the AI features that come with it. Against this backdrop, the bespoke Baltra ASIC might be Apple's only viable avenue of escaping Google's ensnaring clutches. The Information is out with a scoop today, reporting that the current use of AI servers for Apple Intelligence features has fallen far short of Apple's original expectations, leading to an embarrassing situation where around 90 percent of the Cupertino giant's Private Cloud Compute capacity is sitting idle on warehouse shelves. As a refresher, under Apple's Private Cloud Compute framework, relatively simple AI tasks are performed by using on-device models and the computational resources of the device itself, while the more complex tasks are offloaded to Apple's private cloud servers using encrypted and stateless data for subsequent inference. However, according to The Information's sources, the current uptake of Apple Intelligence features has fallen far short of Apple's original expectations, leading to a situation where just 10 percent of Apple's Private Cloud Compute capacity has been actively deployed within data centers. What's more, the prevailing situation has been aggravated by the fact that Apple's current AI infrastructure is very fragmented, with different teams utilizing disparate technology stacks rather than leveraging a single, unified server tech stack. The current paradigm is chronically inefficient and regularly leads to cost overruns from duplicate infrastructure. It is for this reason that Apple has picked Google's servers to power the next-gen Chatbot Siri. As a refresher, Apple intends to launch a revamped version of Siri this year, bringing the much-delayed in-app actions, personal context awareness, and on-screen awareness to its bespoke voice assistant, thereby, enabling a wide variety of agentic actions across apps, based on personal data and on-screen content. This revamped Siri would be powered by a 1.2-trillion-parameter custom Gemini AI model, dubbed the Foundation Models version 10. Even so, with next year's iOS 27 update, Apple is planning to launch a dedicated Siri chatbot that will run on Google's own TPUs and cloud infrastructure, but owned by Apple. The iPhone manufacturer insists that the arrangement would not result in a change in Apple's stringent privacy-related safeguards. According to Bloomberg's Mark Gurman, the Siri chatbot will be baked into Apple's software rather than debuting as a standalone app, allowing it to search the web, generate content, including images, provide coding assistance, summarize and analyze information, as well as upload files. It will be able to use personal data to complete tasks and sport a substantially improved search feature. Apple is also designing a feature that will let the Siri chatbot view open windows and on-screen content, as well as adjust device features and settings. The chatbot Siri will reportedly leverage a much more advanced version of Google's Gemini model, known internally as Apple Foundation Models version 11. According to Gurman, "the model is expected to be competitive with Gemini 3 and significantly more capable" than the one supporting the revamped Siri. Back in Spring 2024, multiple reports emerged that Apple was working with Broadcom on its first AI server chip, bearing the internal codename "Baltra." Some reports at the time also suggested that the chip would leverage TSMC's 3nm 'N3E' process, and that the design process would conclude over the coming 12-month period. According to the Information, The chip itself could sport various chiplets, with each of them designed for a specific function. Apple could later combine each of these chiplets into a single unit, with Broadcom possibly included to help with how each of these processors communicates with each other when simultaneously running in Apple Intelligence servers. This siloed approach would allow Apple to keep the overall design of the AI ASIC hidden even from its partners like Broadcom. As for the actual servers, we previously reported that Foxconn has been tasked to produce them, with Apple's assembling partner slated to receive some assistance from Lenovo and its subsidiary when it comes to the overall design. Given the long-standing server-related disarray at Apple, the Baltra-based servers might be the only avenue left for the Cupertino giant to rectify its chronic inefficiencies in this arena and to escape Google's clutches. After all, Apple has previously indicated that its utilization of Google's Gemini models is likely to be a temporary arrangement. Do note that Baltra-based servers are currently expected to undergo mass deployment in 2027 or 2028.