Stock Markets April 10, 2026 08:38 AM

Why Apple Is Reserving SoIC Capacity at TSMC and What It Means

Massive wafer commitments point to in-house AI server chips for Apple Intelligence rather than incremental Mac production

By Leila Farooq
Why Apple Is Reserving SoIC Capacity at TSMC and What It Means

Apple has secured large allocations of TSMC’s System on Integrated Circuit (SoIC) capacity for 2026 and 2027, commitments that industry analysts say are likely destined for server-side AI accelerators rather than the company’s Mac lineup. The scale of the wafer reservations exceeds the needs of current Ultra-powered Mac shipments, supporting the view that Apple is building bespoke AI ASICs for its Private Cloud Compute infrastructure to support Apple Intelligence.

Key Points

  • Apple has booked 36,000 SoIC-equivalent wafers for 2026 and 60,000 for 2027 at TSMC.
  • The wafer volumes far exceed the SoIC demand implied by 2025 shipments of M2 and M3 Ultra-powered Mac Pro and Studio units, suggesting the capacity is aimed at server-side AI infrastructure.
  • Analysts believe the primary use is a 3nm AI ASIC for Private Cloud Compute, internally named "Baltra," intended to improve AI inference performance and efficiency for Apple Intelligence.

Overview

Apple is committing significant production volume to TSMC’s advanced 3D chip packaging technology, System on Integrated Circuit (SoIC), as it scales infrastructure for its artificial intelligence initiatives. SoIC is TSMC’s platform for high-performance chiplet designs that enables densely integrated, 3D stacked packages.

What SoIC is and who has used it

SoIC is an advanced 3D packaging approach from TSMC designed to support high-performance chiplet architectures. Until recently, a major user of this technology was AMD, which deployed SoIC across its MI-series AI GPUs and in some of its premium gaming processors. That prior adoption established SoIC as a production-proven option for demanding compute workloads.

Apple’s wafer commitments and scale

According to analysis from Morgan Stanley, Apple has placed orders with TSMC equivalent to 36,000 wafers for 2026 and 60,000 wafers for 2027. Those totals are sizable by industry standards and suggest these allocations are intended for a substantial program rather than modest product upgrades. The magnitude of these reservations has prompted market observers to look beyond Apple’s current Mac product line for the primary end use.

Why the Mac lineup is unlikely to consume that capacity

Apple shipped roughly 80,000 Mac Pro and Studio units powered by M2 and M3 Ultra processors in 2025. Using an estimated yield of 50 Ultra processors per wafer, those shipments would correspond to about 1,600 wafers of SoIC demand. That figure represents only a small fraction of Apple’s reserved SoIC wafer allocations, indicating the bulk of capacity is likely committed to applications other than these Mac models.

Analyst interpretation and the Baltra ASIC

Morgan Stanley analyst Erik Woodring has interpreted the wafer bookings as evidence that most of the SoIC capacity will be used for Apple’s 3nm AI ASIC for Private Cloud Compute. Internally, Apple has reportedly referred to this in-house accelerator as "Baltra." Woodring suggests the chip is intended to replace the M-series Ultra processors currently deployed in Private Cloud Compute to improve AI inference performance and efficiency. He also noted reports of a multi-year collaboration between Apple and Broadcom, which align with the direction of this internal development.

Private Cloud Compute and Apple Intelligence

Private Cloud Compute acts as the server-side backbone for Apple Intelligence, the suite of AI features Apple has introduced for its devices. By developing a proprietary AI ASIC for these servers, Apple would move away from dependence on external cloud providers for certain inferencing workloads, aiming for tighter vertical integration and potentially greater control over costs and performance.

Scale comparison and remaining unknowns

To provide context, Morgan Stanley notes that AMD’s allocation of 42,000 SoIC wafers in 2026 is expected to produce nearly 800,000 MI-series AI GPUs. Apple’s 36,000 wafer allotment in 2026, growing to 60,000 in 2027, is therefore substantial if applied primarily to Private Cloud Compute infrastructure. Despite the apparent strategic clarity, Woodring cautions that the performance and efficiency of Apple’s AI ASIC at scale remain key unknowns.

Capitalization and potential cost effects

The analyst further observes that Apple’s expenditures for this infrastructure appear more likely to be recorded as capital expenditure rather than ongoing operating costs. That accounting treatment could be more cost-efficient over time, though the full financial and operational implications will depend on chip performance, deployment scale, and integration with Apple’s server environment.


Note: The article summarizes reported wafer commitments and analyst commentary without asserting outcomes beyond the evidence presented.

Risks

  • Uncertainty about the actual performance and efficiency of Apple’s AI ASICs at scale could affect the realized benefits for Private Cloud Compute - impacts cloud infrastructure and enterprise IT markets.
  • Large capital expenditure timing and scale could pose financial allocation risks if the chips do not deliver expected returns - impacts corporate capex planning in technology and semiconductor sectors.
  • Shift to proprietary server hardware may face integration and operational challenges that are not yet quantified - impacts cloud services and data center operations.

More from Stock Markets

Tel Aviv Benchmark Reaches Record as Insurance, Financials Drive Gains Apr 10, 2026 Oslo market closes lower as Media, Transport and Financials weigh; OBX down 0.93% Apr 10, 2026 Palantir Shares Partly Recover After Trump Endorsement as Anthropic Sparks Selling Pressure Apr 10, 2026 Erayak Power stock jumps as company retools for North American market Apr 10, 2026 HawkEye 360 Files for NYSE Listing as Defense Tech IPO Activity Accelerates Apr 10, 2026