Apple pushed back mass production of its Baltra AI chip by another year after whispers suggested the thing would hit manufacturing lines sooner. Broadcom got roped into the network tech side while TSMC handles fab work on 3nm nodes, and the delay reasons remain murky. The company already stocked its data centers with M2 Ultra gear and appears ready to slot in M4-based servers for Private Cloud Compute tasks.
The Baltra setup leans toward inference work rather than training models from scratch, with estimates pointing to clusters of roughly 64 chips linked through high-bandwidth LPDDR memory instead of massive GPU farms. Cost efficiency matters more than raw horsepower for what Apple wants to accomplish.
Meanwhile, the tech giant allegedly dropped a billion per year on custom Google Gemini integration for personalized Siri features and web search functions, which suggests training gets outsourced while Baltra handles the lighter lifting once models already exist.
The Baltra setup leans toward inference work rather than training models from scratch, with estimates pointing to clusters of roughly 64 chips linked through high-bandwidth LPDDR memory instead of massive GPU farms. Cost efficiency matters more than raw horsepower for what Apple wants to accomplish.
Meanwhile, the tech giant allegedly dropped a billion per year on custom Google Gemini integration for personalized Siri features and web search functions, which suggests training gets outsourced while Baltra handles the lighter lifting once models already exist.
