Apple planning Baltra ASIC ramp
Analyst Ben Bajarin reports that Apple is ramping SoIC production at TSMC for a Baltra ASIC, with an expected 36k–60k wafer target across 2026–27 aimed at AI server and Siri workloads. That level of wafer demand suggests Apple is investing beyond handset silicon into bespoke parts for server or edge AI functions, signalling a broader hardware footprint. The move would tighten Apple’s supply interactions with TSMC and raise the stakes on foundry sequencing and capacity planning. (x.com)
Apple is reportedly reserving a surprisingly large block of Taiwan Semiconductor Manufacturing Company packaging capacity for a chip called Baltra, with 36,000 wafer-equivalents in 2026 and 60,000 in 2027 tied to Apple’s artificial intelligence server plans. That is not the kind of order you place for a lab experiment. (x.com) (edgen.tech) Apple already runs some artificial intelligence jobs on its own servers through Private Cloud Compute, which is the cloud system behind Apple Intelligence requests that are too large to stay on an iPhone, iPad, or Mac. Apple says those servers use custom Apple silicon and are designed so user data sent to the cloud is not accessible even to Apple. (security.apple.com) (apple.com) That makes Baltra easier to understand. Apple is not suddenly trying to become a public cloud company like Amazon Web Services; it is trying to build the back-room machines that answer Siri and Apple Intelligence requests when the phone cannot do the job alone. (security.apple.com) (apple.com) The manufacturing clue here is SoIC, short for System on Integrated Chips, which is Taiwan Semiconductor Manufacturing Company’s method for stacking and bonding chips in three dimensions. Instead of laying every function side by side on one slab of silicon, SoIC lets designers pile pieces on top of each other like floors in a building to shorten the distance data has to travel. (tsmc.com) (investor.tsmc.com) Taiwan Semiconductor Manufacturing Company says SoIC is built for cloud, network, and edge workloads that need more bandwidth and lower delay between chip blocks. Those are exactly the kinds of jobs that show up in artificial intelligence servers, where moving data between memory and compute can waste huge amounts of time and power. (tsmc.com) (ieeexplore.ieee.org) The report around Baltra points to inference, not training. In plain English, training is the expensive process of teaching a model from giant data sets, while inference is the cheaper but nonstop work of answering real user prompts one by one after the model already exists. (kad8.com) (techpowerup.com) That fits Apple’s current artificial intelligence strategy. Apple has been explicit that Private Cloud Compute exists to run larger models for user-facing features, and recent reporting says Apple is also preparing a 250,000-square-foot server factory in Houston scheduled to open in 2026 for the machines that support Apple Intelligence. (security.apple.com) (apple.com) (cnbc.com) If Baltra replaces some current servers built around M-series chips, Apple gets a part designed for one job instead of a general-purpose chip adapted for many jobs. Morgan Stanley’s view, as cited in follow-on reporting, is that Baltra would replace M-series Ultra processors in Private Cloud Compute to improve inference performance and power efficiency. (edgen.tech) (9to5mac.com) The bigger shift is organizational. Apple spent the last decade extending custom silicon from phones to watches to tablets to laptops, and Baltra suggests the same playbook is now moving into data centers, where the customer is not a shopper holding an iPhone but Apple’s own Siri and Apple Intelligence stack. (apple.com) (security.apple.com) (edgen.tech) That also gives Taiwan Semiconductor Manufacturing Company a larger role in Apple’s artificial intelligence timetable. When one company is reserving tens of thousands of advanced packaging slots for 2026 and 2027, launch timing stops being only a design question and becomes a factory scheduling question too. (x.com) (tsmc.com) (investor.tsmc.com)
