AI Trading Undergoes Sharpest Stress Test of the Year.

Over the past week, U.S. AI concept stocks have retreated from their highs, with the semiconductor, storage, and data center chain being the main targets of profit-taking. Samsung's performance failed to ease the market's concerns about the overvaluation of AI, leading to a collective decline in chip stocks; meanwhile, more and more tech stocks have significantly pulled back from recent highs, indicating that investors are no longer willing to pay a premium valuation solely for the phrase "strong AI demand."

Morgan Stanley

In a report on NVIDIA's roadshow, Morgan Stanley stated that the core message conveyed by NVIDIA's management is that growth is still accelerating, and the sources of demand are diversifying. The firm maintains NVIDIA as its top pick in the semiconductor industry, with a target price of $288. The report stated that NVIDIA's data center demand is no longer solely coming from a few large cloud companies; AI labs, emerging cloud firms, enterprise customers, and sovereign AI projects are all expanding their purchases.

The NVIDIA management team divides demand into three main sources. The first is AI labs, currently accounting for about 20% of total demand; the second is traditional hyperscale cloud players, accounting for half of revenue; the third is "AI cloud, industrial, and enterprise" customers, which also represent nearly half of data center revenue. Morgan Stanley stated that factors such as power, land, and geopolitics are driving the expansion of demand from sovereign AI and emerging cloud players.

The report also downplayed the market's concerns about custom ASICs replacing GPUs. Morgan Stanley stated that cloud players will indeed continue to develop in-house chips, but NVIDIA may still retain a large share. The firm believes that in many cases, NVIDIA can still offer the lowest per-token cost, and the growth of custom chips does not necessarily mean a significant decline in NVIDIA's market share.

However, NVIDIA itself acknowledges a more difficult-to-avoid issue: memory shortages may persist for several years. The Morgan Stanley report stated that if the industry requires a 10x token growth each year and memory supply is "essentially static" relative to this growth, system design must adapt. The report mentioned that NVIDIA may alleviate bottlenecks through system optimizations between compute, networking, and memory.

JPMorgan

This is also why TSMC's advanced packaging has become another major theme. In a report on July 10, JPMorgan raised its forecast for TSMC's CoWoS capacity, expecting TSMC's monthly CoWoS capacity to reach approximately 115,000, 190,000, and 225,000 wafers by the end of 2026, 2027, and 2028, respectively. The firm stated that TSMC is accelerating AP7 Phase 2 capacity allocation and selectively outsourcing backend wafer manufacturing to leading foundries such as ASE to unleash more capacity.

CoWoS is a key process for AI chip packaging, used to integrate GPU, HBM, and other chip components together. As projects from Nvidia, AMD, Google TPU, and Amazon Trainium expand, advanced packaging is becoming an increasingly clear constraint in the AI supply chain. J.P. Morgan stated that the overall supply-demand gap has recently widened to about 20%, partly due to agentic AI driving CPU demand.

An important change in this report is that server CPUs are also beginning to be an incremental source of advanced packaging demand. J.P. Morgan stated that by 2027, around 5 million Nvidia Vera CPUs and over 3 million AMD Venice CPUs will be first adopted with CoWoS packaging. As TSMC needs to prioritize serving larger AI accelerator packages, some CPU-related demand may flow to OSAT companies like ASE and Amkor.

The firm also noted that Google TPU, AWS Trainium, and some emerging ASIC projects will continue to consume advanced packaging resources. TPU shipments for 2027 are now forecasted to be raised to 8.9 million units, with a possibility of reaching 11.6 million units in 2028. The 3-lifecycle demand for AWS Trainium is also on the rise, and Trainium 4 may bring more complex packaging requirements in 2028.

Outside of Nvidia, AMD is also seen as a beneficiary of packaging demand. J.P. Morgan stated that AMD MI450 will reach around 1.5 million units in 2027, with MI500 possibly entering production in 2028 and using a large-size CoWoS-L package. AMD server CPUs will also see a significant increase in CoWoS consumption.

Beyond packaging, another significant figure comes from cloud provider capital expenditure. In a report dated July 12, Morgan Stanley stated that due to rising construction costs for each GW of computing power and higher future compute demand expectations, the firm has raised its capital expenditure forecasts for major hyperscale cloud providers in 2027 and 2028 to around $1.2 trillion and $1.4 trillion, respectively.

This spending will bring about a leap in computing capacity. Morgan Stanley predicts that hyperscale cloud providers' available computing power will increase from around 30GW in 2025 to nearly 120GW in 2028, a fourfold increase. Among them, AWS may have the largest available computing power by 2028, around 35GW; Google around 31GW; Meta reaching approximately 14GW and 21GW in 2027 and 2028, respectively.

However, the larger the capital expenditure, the more investors will demand a return. Morgan Stanley stated that the market will be closely watching these companies to see if they can translate the additional hash rate into "sustained, incremental, and profitable revenue." This means that cloud revenue, AI APIs, enterprise software, advertising tools, search monetization, and private model hosting will all become key valuation drivers.

Meta is Morgan Stanley's top pick among internet stocks. The report stated that Meta has five profit "options" that have not yet been fully priced in by the market, including neocloud, Meta AI Search, API revenue, subscriptions, and upward-trending advertising revenue. The bank estimated that if Meta were to use 100MW of GB300 hash power for the Muse Spark 1.1 API, it could generate around $8.6 billion in revenue and contribute approximately $1.91 to the 2028 EPS. If 25% of Meta's 15 million advertisers were to pay around $200 monthly to use related products, this could also translate to approximately $8 billion in revenue and about $2 in EPS.

Amazon's focus, on the other hand, is on AWS. Morgan Stanley expects AWS revenue to grow by around 35% and 40% in 2026 and 2027, respectively, and believes that this forecast may still be conservative. The bank anticipates that AWS's second-quarter backlog will increase by about $110 billion sequentially, reaching approximately $475 billion, mainly driven by transactions from private AI labs. The report suggests that AWS has the capability to access nearly all mainstream, small, and custom models, giving it an advantage in an environment that optimizes the token cost for each task.

Google's highlights are in Cloud, TPU, and Search. Morgan Stanley predicts that Google Cloud will continue to grow rapidly in 2026 and views Google as the "full-stack AI winner." However, the bank also warns that Google faces short-term risks related to hash power limitations, which could affect revenue growth or product launch cadence.

Bernstein

These assessments collectively explain why semiconductor equipment stocks are still favored. In a report dated July 13, Bernstein discussed a market concern: if memory stocks retreat, can semiconductor equipment stocks still rise? The bank's conclusion is affirmative.

Bernstein stated that the historical correlation between memory stocks and wafer fabrication equipment (WFE) stocks is not as high as investors might think. From 2012 to 2018, the correlation between memory stocks and WFE company stock prices was only about 0.4; since 2019, the correlation has increased slightly, but overall remains low. The report argues that the factors driving semiconductor equipment stocks are not just memory prices but also include logic chips, wafer fabrication, advanced packaging, HBM expansion, and regional industry policies.

The bank pointed out that in multiple past cycles, memory and equipment stocks could move in different directions. For example, from 2015 to 2016, memory stocks declined, but WFE stocks performed significantly better; from 2021 to 2022, amidst the chip shortage and interest rate changes, WFE similarly outperformed. Bernstein believes that even if memory prices normalize in 2027, logic and foundry investments along with additional memory capacity expansion will continue to support equipment demand.

Bernstein remains bullish on WFE in the long term, noting that SK Hynix recently announced approximately 100 trillion KRW investment in Cheongju, South Korea, and the South Korean government is also considering supporting Samsung and SK Hynix in building wafer fabs in the southwest of Korea. The bank believes that there is still room for upward revisions to WFE companies' 2028 EPS.

These latest reports still cover the hottest targets in AI trading—Nvidia, TSMC, cloud providers, semiconductor equipment, and the storage chain. The change lies in the analysis no longer solely focusing on demand growth and order visibility but starting to further dissect whether these companies can deliver on growth: Can Nvidia continue to reduce the cost per token? Can TSMC and OSAT expand enough packaging capacity? Will HBM and server CPUs create new supply constraints? Can the power and data center investments of cloud providers translate into revenue?

This has led the AI market into a more discerning stage. Strong demand remains a prerequisite, but high valuations require more evidence. For popular targets, the market will now look beyond just "is there AI demand" to how much resources they can obtain in bottlenecks, how much cost they can bear, and ultimately how much computing power can be turned into sustainable profits.