Apple Secretly Develops AI Server Chip "Baltra," Targeting Nvidia, Aims for 2026 Mass Production

Apple Secretly Develops AI Server Chip "Baltra," Targeting Nvidia, Aims for 2026 Mass ProductionDecember 12, sources familiar with the matter revealed that Apple is intensely developing its first server chip specifically designed for artificial intelligence, codenamed "Baltra." This chip aims to meet the burgeoning computational demands of Apple's next-generation AI capabilities and is expected to enter mass production by 2026

Apple Secretly Develops AI Server Chip "Baltra," Targeting Nvidia, Aims for 2026 Mass Production

December 12, sources familiar with the matter revealed that Apple is intensely developing its first server chip specifically designed for artificial intelligence, codenamed "Baltra." This chip aims to meet the burgeoning computational demands of Apple's next-generation AI capabilities and is expected to enter mass production by 2026. This marks another major milestone for Apple's chip team, following the successful design of iPhone and Mac processors. The team has consistently set industry benchmarks for chip performance and energy efficiency, and the development of the Baltra chip will undoubtedly further solidify its leading position.

This move reflects Apple's consistent commitment to in-house development, aiming to reduce its reliance on external chip suppliers like Nvidia. Nvidia currently dominates the AI chip market, but its high cost, high power consumption, and supply constraints have become bottlenecks for AI development. Apple hopes to lessen its dependence on Nvidia through its own server chip, gaining better control over costs and the supply chain.

This strategic decision is closely tied to Apple's launch last month of its first generative AI feature, "Apple Intelligence." Currently available in newer iPhones and Macs in some countries, Apple Intelligence encompasses various functions, including text generation and correction, image creation, notification summarization, and web content processing. While the iPhone processor can handle some AI tasks, more complex requests are sent to Apple's cloud servers for processing. However, Apple's current chips aren't optimized specifically for AI processing, resulting in a significant speed and efficiency gap compared to chips designed specifically for AI.

To meet the demands of significantly expanding generative AI capabilities in the coming years, and to support its ambitious goal of AI services covering billions of devices, Apple needs more efficient and powerful AI chips. Unlike other tech companies, Apple consistently pursues in-house chip development, not only for stronger technological control but also to better protect user data privacy and security. Over a decade ago, Apple used Nvidia chips for high-performance graphics processing in Macs, but after commercial disagreements, gradually phased them out and invested heavily in its own chip technology.

It's noteworthy that Apple revealed last week it's testing chips designed by Amazon for training large language models. The Baltra chip, however, is likely to focus on inference tasks, receiving new data (e.g., a user-provided image description) and applying it to the AI model to generate output (e.g., generating an image based on the description).

Apple isn't the only major tech company developing AI chips for internal use. Google, Meta, Microsoft, and Amazon are also developing their own AI chips to reduce data center build and operational costs and lessen reliance on Nvidia. However, designing AI chips from scratch is challenging, requiring significant investment and a lengthy development cycle. Many companies still rely on Nvidia chips for model training due to the extremely high computational demands of this process. Therefore, they often choose to collaborate with existing chip companies, leveraging their IP and design services.

Similar to Google, Apple has partnered with Broadcom, utilizing Broadcom's expertise in networking technology, crucial for efficiently handling the massive data volumes required for AI processing. Broadcom typically doesn't license its IP but sells chips directly to customers. However, in its collaboration with Apple, Broadcom seems to have adopted a different strategy, providing limited design services, primarily its networking technology, while Apple manages chip production, with manufacturing handled by TSMC. The specifics of their collaboration remain undisclosed, with spokespeople from both companies declining to comment.

Sources indicate that Apple's design team in Israel is leading the Baltra chip development. This team designed the processor that replaced Intel chips as the core processor for Macs in 2020 and possesses extensive chip design experience. This summer, Apple cancelled a planned high-performance chip for Macs, focusing Israeli engineers' resources on the Baltra chip, reflecting a strategic shift and prioritization of AI chip development.

Apple plans to use TSMC's cutting-edge N3P manufacturing process for the Baltra chip. Compared to the process used for Apple's latest M4 Mac processor, N3P will significantly improve performance. Sources reveal Apple plans to launch at least one iPhone chip using the N3P process next year. Meanwhile, companies like OpenAI and Nvidia also plan to leverage this advanced process to optimize their AI chip performance.

In designing Baltra, Apple will employ the "chiplet" design strategy pioneered by AMD, dividing the large chip into smaller units and then combining them into a complete chip. This approach reduces manufacturing complexity and minimizes potential defects. Dylan Patel, principal analyst at SemiAnalysis, notes, "Broadcom might only be designing certain chiplets related to networking. Using chiplets allows Apple to limit Broadcoms involvement to a small portion of the chip design and ensure the overall designs secrecy."

The core of the Baltra chip will consist of numerous Apple Neural Engines (ANE), accelerating AI tasks. Initially designed for Apple's shelved self-driving car project for inference processing, the ANE later found its way into iPhones.

Apple aims to complete the initial design of the Baltra chip within the next 12 months. Sources say this timeline is aggressive given the chip's large size and complex architecture. After the initial design, Apple expects another year for refinement and testing before mass production. The successful development and production of the Baltra chip will have a profound impact on Apple's competitiveness in the AI field and further cement its leading position in the tech industry.