Apple’s Baltra ASIC Delayed as Most AI Servers Gather Dust on Shelves

The turmoil surrounding Apple’s AI hardware strategy appears to be as significant as the challenges faced on the software side. This has compelled the tech giant from Cupertino to increasingly rely on Google’s servers and its potent Gemini model to introduce the much-anticipated updates to Siri. In light of these circumstances, the development of Apple’s proprietary Baltra ASIC could provide a crucial pathway for the company to mitigate its dependence on Google.

Internal Disarray: Apple’s AI Servers Largely Underutilized

Recent reports from The Information reveal unsettling insights into Apple’s AI infrastructure: a staggering 90 percent of the company’s Private Cloud Compute capacity remains dormant on warehouse shelves. This underperformance is vastly below initial projections and reflects a significant misalignment between Apple’s ambitions and actual progress in its AI offerings.

Under the framework of Apple’s Private Cloud Compute, straightforward AI processes are managed on-device, utilizing the device’s own computational capabilities. In contrast, more intricate tasks rely on Apple’s private cloud servers, employing encrypted and stateless data for these operations. Unfortunately, this model has not translated into efficient deployment. Sources indicate that only 10 percent of Apple’s cloud compute resources have been actively utilized.

Compounding this issue is the fragmented nature of Apple’s current AI architecture. Different teams across the company are using varying technology stacks rather than a cohesive, unified system. This disorganization not only hampers efficiency but also leads to significant cost overruns due to redundant structures. Consequently, Apple has turned to Google’s servers to support the development of an advanced version of Siri.

As part of its ongoing enhancements, Apple plans to launch a revamped Siri featuring new in-app functionalities, context-awareness, and the ability to interact with on-screen content. This upcoming iteration aims to empower Siri to execute a broader range of tasks, enhanced by a custom AI model known as Foundation Models version 10, which boasts an impressive capability of 1.2 trillion parameters.

In conjunction with the iOS 27 update, Apple intends to unveil a dedicated Siri chatbot, which will operate on Google’s TPUs and cloud infrastructure while maintaining ownership by Apple. The company assures users that this partnership will not compromise its robust privacy protocols.

Mark Gurman from Bloomberg indicates that this chatbot will integrate seamlessly into Apple’s software ecosystem rather than appearing as a standalone application. Users can expect advanced capabilities, including web searches, content generation, coding assistance, summarization, analysis, and file uploads. The chatbot will harness personal data to enhance functionality, complemented by an upgraded search feature. It will also be equipped to access users’ open windows and adjust settings on devices.

The forthcoming chatbot will utilize an advanced version of Google’s Gemini model, known internally as Apple Foundation Models version 11. According to Gurman, this model is anticipated to compete with Gemini 3 and offers enhanced capabilities compared to the iteration supporting the revamped Siri.

Introducing Apple’s Baltra ASIC

In Spring 2024, reports surfaced indicating that Apple is collaborating with Broadcom to develop its inaugural AI server chip, codenamed Baltra. This chip is speculated to utilize TSMC’s advanced 3nm ‘N3E’ manufacturing process, with its design phase expected to conclude within the next twelve months.

According to insights from The Information, the Baltra chip may feature multiple chiplets, each designed for specific functions. This modularity would allow Apple to consolidate these components into a single unit while ensuring optimized intercommunication between processors, even keeping the overall architecture confidential from collaborative partners like Broadcom.

As reported earlier, Foxconn has been chosen as the partner for manufacturing the associated server units, with design support also coming from Lenovo and its subsidiaries.

Given the long-standing challenges related to servers at Apple, the successful rollout of Baltra-based servers may be essential for the company to address its operational inefficiencies and reduce its reliance on Google. Apple has previously hinted that its collaboration with Google for AI solutions will be a temporary measure. Currently, these Baltra-based servers are projected to commence mass deployment around 2027 or 2028.

Source & Images