Morgan Stanley: AI chip demand drives up wafer and packaging use, while supply expansion also brings cost pressure

How large a market can AI chip demand support?

Morgan Stanley expects the global semiconductor market to reach 2.5 trillion US dollars in 2027, including an AI chip market of approximately 809 billion US dollars, or about one-third of the total.
In its optimistic scenario based on supply-chain data, the bank assumes cloud capital expenditure will reach 1.422 trillion US dollars in 2027, including approximately 800 billion US dollars in AI server capital expenditure, corresponding to 641 billion US dollars in revenue from cloud AI GPUs and custom chips.
Cloud providers' spending continues to expand, but growth may slow: the bank's US technology team estimates from the top down that combined capital expenditure by 14 listed cloud service providers worldwide will approach 1.6 trillion US dollars in 2028, up 12% year over year.

Morgan Stanley expects spending on wafers consumed by AI computing chips to exceed 55 billion US dollars in 2027; TSMC's advanced-node wafer prices may rise 5%–10% in 2027. The bank also raised its forecasts for TSMC's capital expenditure in 2027 and 2028 to 83 billion and 92 billion US dollars, respectively.
CoWoS is an advanced packaging method that connects computing chips with components such as high-bandwidth memory. The bank expects global demand for CoWoS wafers to rise from 1.394 million wafers in 2026 to 2.509 million wafers in 2027; Nvidia's demand is expected to rise from 780 thousand to 1.222 million wafers.
To meet AI demand, Morgan Stanley expects TSMC could expand CoWoS capacity to 260 thousand wafers per month by 2028. The scale of expansion also depends on whether downstream chip shipments can support it.

Based on the bank's model of chip shipments and memory configuration per chip, demand for high-bandwidth memory will be approximately 44.923 billion Gb in 2027, with Nvidia remaining the main user. Both chip shipments and memory capacity per chip affect this demand.
In addition to buying GPUs, cloud service providers continue to develop custom chips. Morgan Stanley expects total Google TPU shipments to rise from 3.3 million chips in 2026 to 7.5 million chips in 2027; however, the upside for shipments of AI chips such as TPUs may be constrained by the supply of ABF substrates.
Memory supply is also expanding. The bank expects ChangXin Technology's DRAM capacity to reach approximately 500 thousand wafers per month in 2028, or about 15% of global capacity; its share of the high-bandwidth memory market, measured by bit shipments, is expected to rise from 3.8% in 2026 to 4.4% in 2027.

Morgan Stanley believes the supply chain is prioritizing AI chips, resulting in shortages in areas such as memory; rising wafer, packaging and testing, and memory costs will also increase margin pressure on chip design companies in 2026.
The bank expects AI chip growth to slow in 2027, while still expecting AI chip revenue to rise quarter over quarter. This means slower growth and continued revenue increases can occur at the same time.
Besides funding budgets, energy, chip capacity and regulation are also listed in the report as constraints on AI growth; the United States faces energy constraints, while China faces chip capacity constraints.