Apple’s Balta wafer orders rising
Morgan Stanley says Apple is ramping system‑in‑package (SoIC) capacity at TSMC for its Balta custom ASIC, with orders of about 36,000 wafers in 2026 and 60,000 in 2027. Those committed wafer volumes signal Apple is scaling a cloud/hybrid AI compute strategy that will affect packaging, test and supply allocation. The reported cadence matters because large wafer commitments can shift domestic manufacturing and procurement planning for Apple programs. (x.com)
Apple is reportedly reserving far more of Taiwan Semiconductor Manufacturing Company’s most advanced stacking capacity for a chip called Baltra than most people expected, with Morgan Stanley pointing to about 36,000 wafers in 2026 and 60,000 in 2027. Those are factory-slot numbers, not product-launch rumors, which is why supply-chain people pay attention to them first. (wccftech.com) A wafer is the big round silicon disc that gets cut into many chips, and reserving tens of thousands of them years ahead is like booking a huge block of seats before the stadium is built. Apple has done that for iPhone and Mac processors for years, but this time the reservation is tied to server hardware for artificial intelligence work. (wccftech.com) (apple.com) The manufacturing step here is called System on Integrated Chips, which is Taiwan Semiconductor Manufacturing Company’s method for stacking separate pieces of silicon on top of each other so they behave more like one chip. The company says this 3D stacking raises connection density and improves performance, which is exactly what matters when a server is moving huge amounts of model data back and forth. (tsmc.com 1) (tsmc.com 2) Apple already told the public in June 2024 that some Apple Intelligence requests run in the cloud on Apple silicon servers through a system it calls Private Cloud Compute. Apple also said those servers are built for larger foundation models than the ones that fit on an iPhone, which means the company has already committed to owning at least part of the server layer itself. (apple.com 1) (apple.com 2) (apple.com 3) That is why Baltra matters. Reports in late 2024 said Apple was working with Broadcom on its first server chip for artificial intelligence, and later reporting said Apple’s chip team was targeting 2027 for dedicated server processors alongside future Mac chips. (9to5mac.com) (macrumors.com) The jump from 36,000 wafers to 60,000 wafers in one year suggests Apple is planning for scale, not a lab experiment. A reservation that large usually reaches beyond chip design into packaging houses, test capacity, server boards, memory supply, and data-center planning, because all of those pieces have to line up at the same time. (wccftech.com) (tsmc.com) There is also a competitive angle inside Taiwan Semiconductor Manufacturing Company’s factories. Reports in January 2026 said the artificial intelligence boom was already reducing Apple’s old ability to dominate every leading-edge queue at Taiwan Semiconductor Manufacturing Company, so locking in advanced stacking now looks like a way to secure room before graphics-chip demand eats more of it. (9to5mac.com) (macrumors.com) If Apple keeps more inference work on its own silicon, it gains tighter control over privacy, cost, and feature timing than it would get by renting everything from outside cloud providers. Apple’s own security documents describe Private Cloud Compute as an extension of the iPhone security model into the cloud, and that promise gets easier to defend when the server chip, server operating system, and model stack are all designed in-house. (apple.com 1) (apple.com 2) (apple.com 3) So the real story in those wafer numbers is not a single chip launch in 2027. It is Apple quietly building the factory footprint for a hybrid artificial intelligence system where small jobs stay on your device, larger jobs move to Apple-run servers, and the bottleneck shifts from app demos to who controls enough advanced packaging to ship at scale. (apple.com) (apple.com) (tsmc.com)
