What determines the direction of chip stocks? How much more money do cloud giants need to invest, and who has received Taiwan Semiconductor's production capacity?

As the chip industry enters the "fast lane", the market's focus shifts to the capital expenditures of American cloud computing giants and the allocation of Taiwan Semiconductor's CoWoS (packaging) capacity, which may lead to more intense market fluctuations in the short term.

On October 23, Morgan Stanley analysts Charlie Chan and Daniel Yen released a report stating that they are currently optimistic about artificial intelligence, but the market's expectations for fourth-quarter capital expenditures may be too high, ultimately leading to less than ideal results. Cloud AI semiconductor customers are still undergoing production transformation, with more purchases expected to take place in the first half of 2025.

Morgan Stanley also stated that Taiwan Semiconductor remains its top choice in the AI semiconductor supply chain. Taiwan Semiconductor ensures the real demand for AI, and this cloud service provider is designing its own AI chip (ASIC custom chip), which boosts market confidence in AI semiconductors.

CoWoS reshuffle, NVIDIA takes over AMD

Overall, Taiwan Semiconductor's CoWoS capacity is expected to reach 3.7 million wafers in 2024, and this number is projected to increase to 7.18 million wafers by 2025.

In the report, Morgan Stanley pointed out that there have been some changes in CoWoS demand dynamics recently. Due to the uncertainty in demand for the new AI chip MI325, AMD seems to have reduced some of its CoWoS wafer orders at Taiwan Semiconductor for 2025, decreasing from 11% in 2024 to 13% in 2025. However, Taiwan Semiconductor's order situation appears to remain robust, with this portion of capacity immediately taken over by NVIDIA, with NVIDIA's demand expected to account for 63% by 2025.

Intel's AI chip manufacturer Habana under Intel has not changed its wafer orders at Taiwan Semiconductor, while Marvell has increased its 2025 orders at Taiwan Semiconductor to three times that of 2024, and Alchip has also started to reserve some CoWoS for certain 3nm projects.

It is worth noting that Taiwan Semiconductor also mentioned in its recent earnings conference call that the company is designing its own AI chip (ASIC) and seems to have reserved 30% of its CoWoS capacity for ASIC.

Furthermore, Morgan Stanley also predicts that if Amazon's Trainium 2 production reaches 200,000 wafers in 2024, based on the CoWoS capacity booking situation, by 2025, Trainium 2 production could reach 400,000 wafers, and Inferentia 2.5 production could reach 500,000 to 600,000 wafers.

Regarding Intel's Gaudi 3, Morgan Stanley believes that the number of AI servers in 2025 will be approximately 20,000 to 25,000 units, which means chip consumption will be 160,000 to 200,000 units, consistent with its 16k CoWoS booking, with chip production of around 200,000 units.

Cloud Computing Capital Expenditure May Peak in the First Half of 2025

In terms of capital expenditure by cloud computing giants, Morgan Stanley pointed out that in past cycles, a 2-3 year upward cycle is usually followed by 2-4 quarters of decline. The decline phase refers to a "slowing year-on-year growth", especially from the slowdown of super-large scale cloud computing companies in the United States.

Data on US cloud capital expenditure in the second quarter of 2024 shows that cloud capital expenditure has rebounded from the lows of the second and third quarters of 2023. Due to investments in GPU servers, year-on-year growth in cloud capital expenditure is accelerating.

Morgan Stanley stated that if this cycle repeats, the upward cycle may continue until the first half of 2025.

Historical data shows that when data center construction is in its initial stages (2005-2015), it is easier to see a significant increase in capital expenditure. After 2015, due to the slowdown of "Moore's Law" and the deceleration of sales growth for cloud service providers, capital expenditure growth has normalized. Morgan Stanley stated in the report that with the rapid development of GPUs and their potential for future monetization, year-on-year capital expenditure growth in the first quarter of 2025 may still be strong, exceeding 30%:

"Therefore, based on historical trends, the peak of the cycle should occur in the first half of 2025."