The AI supply chain is being dismantled again: TSM.US (TSM.US) CoWOS production capacity surged 33%! Demand for HBM doubles

The Zhitong Finance App learned that Damo released a research report saying that the global AI supply chain is showing a strong demand trend, and TSMC (2330.TW, TSM.US) CoWOS advanced packaging production capacity continues to expand, and the total production capacity is expected to reach 93 kwpm in 2026. China's AI market is facing hardware supply bottlenecks, and some developers may switch to Huawei's chip solutions. Also, I am optimistic about the performance of Asian ASIC design service providers such as Alchip in the Trainium 3 project. Investors remain optimistic about the AI industry, but are concerned about sustainable demand and investment opportunities outside of NVIDIA (NDVA.US).

Report mind map:

The main contents of the report are as follows:

I. Core background and overall market sentiment

Morgan Stanley's global semiconductor team met with about 20 companies, AI industry experts, and investors in Taiwan and Beijing to learn about the AI supply chain.

Investors are very optimistic about AI, but the key issues focus on possible issues with AI demand and investment targets other than NVIDIA.

II. Situation related to TSMC's CoWoS production capacity

2026 production capacity forecast: TSMC's total COWos production capacity in 2026 is about 90-95k, close to the previous 90k forecast, up 33% from 70k at the end of 2025.

Segmented capacity changes: CoWoS-L is likely to expand to 68k (previously estimated 55k), indicating strong demand for Blackwell or Rubin, so TSMC's CoWos capacity forecast for the end of 2026 was raised from 90kwpm to 93kwpm, while non-TSMC CoWoS production capacity remained at 12kwpm.

Customer demand adjustment

NVIDIA: Total CoWOS consumption will remain at 580k units in 2026, CowOS-L will be raised from 540k to 550k, and CowOS-s will be lowered.

Broadcom: CoWoS consumption was raised to 110k units in 2026 due to increased demand for Meta's mTiAv3 chips.

AWS+Alchip: CoWoS bookings increased from 30k to 40k due to strong demand for Trainium3.

III. Status of China's AI Market

Demand and supply: Demand for AI applications/inference in China is strong, but hardware supply is a bottleneck.

NVIDIA B30 chip: Chinese AI developers are aware of the chip, but have not confirmed an official purchase order; Taiwan's supply chain shows that 2H25 orders for this chip total 2 million units of wafers, with an annual production capacity of about 5 million units.

Alternative: If the B30 cannot be shipped to China, some developers plan to switch to a Huawei chip, but have yet to see the Huawei 910C sold; some developers wait for the B30 supply to increase capital expenses, and their own ASIC can only run reasoning.

LLM progress: China's next-generation big language models are being delayed due to GPU capacity limitations.

IV. ASIC Design Service Provider in Asia

AWS ASICs: Alchip is the sole source of Trainium3 XPU, and revenue is expected to grow strongly in 2026, depending on TSMC's 3nm wafers and CoOS allocations; Marvell focuses on “XPU-Attach” chips, and its Trainium lifecycle and related revenue will determine 2026 growth.

Trainium4: Design service providers are expected to be identified in July, and Alchip has a better chance.

5. AI chip testing and other supply chain conditions

The AI chip testing supply chain shows that the test time for Blackwell chips may rebound from the current 600-700 seconds to 800-900 seconds, and KYEC is still optimistic about KYEC's market share in 2026 due to the addition of new test equipment and KYEC participation.

6. Industry data and forecast

Growth of cloud semiconductors in 2025: Preliminary indications are that cloud semiconductors grew strongly in 2025, as confirmed by aspects such as the AI chip testing supply chain.

CoOS demand breakdown: NVIDIA will account for 66% of CoOS demand in 2026, Broadcom will account for 13%, AMD will account for 7%, etc. (see Exhibit 4-6 below for details).

Key Q&A:

Q1: How will TSMC's CoWoS production capacity increase in 2026, and what impact will it have on related customer demand?

A: TSMC's total CoWos production capacity in 2026 is about 90-95k, up 33% from 70k at the end of 2025, of which CoWS-L may expand to 68k, reflecting strong demand from Blackwell or Rubin. Impacts on customer demand include: NVIDIA's CoWoS-L consumption raised to 550k; Broadcom raised CoOS consumption to 110k due to increased demand for Meta-related chips; and AWS+AlChip increased bookings from 30k to 40k due to Trainium3 demand.

Q2: What are the major issues facing the Chinese AI market and how to deal with them?

A: The main problem in the Chinese AI market is hardware supply bottlenecks, despite strong application/inference demand. Chinese developers are eyeing the NVIDIA B30 chip, and the Taiwanese supply chain says it has ordered 2 million units of 2H25 wafers, but no formal purchase order has been confirmed. In terms of countermeasures, if the B30 cannot be shipped, some developers plan to switch to a Huawei chip, but the Huawei 910C is not yet on sale; some developers wait for the B30 supply before deciding to increase capital expenses, and their own ASIC can only run reasoning.

Q3: What is the competitive landscape for ASIC design service providers in Asia?

A: In terms of AWS ASICs, Alchip is the sole source for Trainium3 XPUs, and revenue is expected to grow strongly in 2026 due to the project, which depends on TSMC's 3nm wafers and CoOS allocations; Marvell focuses on “XPU-Attach” chips. Furthermore, it is expected that the Trainium4 design service provider will be determined in July, and Alchip will have a better chance of winning.

Other charts: 1. TSMC expects to produce 5.1 million chips in 2025, and the annual shipment volume of GB200 and NVL72 is expected to reach 30,000 sets.

Left axis (LHS): quarterly shipments (in millions of units), representing Blackwell chip production

Right axis (RHS): (thousand sets), represents the number of NVL72 racks

2. Artificial intelligence semiconductor foundry wafer revenue and high-bandwidth memory (HBM) demand in 2025 calculate the estimated advanced process wafer consumption in 2025 (by customer)

Projected high bandwidth memory consumption in 2025 — Nvidia is considered the largest customer

3. High-bandwidth memory (HBM) silicon through-hole (TSV) production capacity is expected to double by 2025

4. High-bandwidth memory (HBM) demand is expected to nearly double that of 2024 in 2025