The Zhitong Finance App learned that with the intense sell-off of US artificial intelligence (AI) related stocks, the market is increasingly questioning whether the AI investment boom will continue. Bernstein acknowledged in the latest research report released on November 9 that his previous optimism about short-term trends in the AI sector was biased, but emphasized that the long-term growth logic of the AI industry has not fundamentally changed. The bank believes that although the market may experience a technical rebound in the short term, considering that issues such as AI commercialization returns, data center construction restrictions, and financing risks have not been fully resolved, it may take longer for the sector to fully recover.
Bernstein pointed out that although the performance recently announced by AI infrastructure companies is generally stable, and many companies have even raised their performance guidelines, related stocks have declined sharply. The report shows that as of last Thursday, the Philadelphia Semiconductor Index (SOX) has fallen 38% from the high set on June 22, 2026, making it the second most serious round of retracement since the launch of ChatGPT, second only to market adjustments after the announcement of the US “Liberation Day” tariff policy.
Among the network and communication equipment companies covered by Bernstein, Celestica (CLS.US) and Lumentum (LITE.US) shares have both fallen by about 42%, Coherent (COHR.US) by 32%, Ciena (CIEN.US) by 29%, Arista Networks (ANET.US) by 17%, and Cisco (CSCO.US) by 12% since September 28.
Bernstein said that when it began covering related companies at the end of September, it believed that the short-term increase in AI infrastructure spending was highly visible, and that the valuation of the network equipment industry was not expensive compared to other hardware and semiconductor companies. However, recent market performance has proven that this short-term judgment is too optimistic.
It is worth noting that the decline in stock prices was not mainly due to deteriorating corporate profits. Bernstein pointed out that since it began relevant research, the five covered companies that have announced their results have all drastically raised their performance guidelines, exceeding the previously agreed expectations of seller analysts. However, as investors lowered their valuations of AI-related stocks, the revised profit expectations still failed to stop the stock price from falling.
The return on AI investment is being questioned whether huge infrastructure spending can continue to be the focus
Bernstein believes that there are two core concerns behind this round of AI sector sell-off. One of them is whether the commercial revenue of AI model developers can support the continued expansion of infrastructure investment.
In recent years, large technology companies and AI developers have continued to invest huge sums of money to build data centers, procure GPUs, and expand computing power resources, but the market is beginning to worry that the growth rate of terminal AI demand may not keep up with the pace of infrastructure expansion.
If AI models and related applications fail to generate sufficient revenue, it may be difficult for cloud computing companies to obtain ideal returns from huge computing power investments, which in turn affects their subsequent financing capabilities and capital expenditure plans.
This concern was further exacerbated in September. At the time, the annualized recurring revenue (ARR) performance of some AI model developers was weak, causing investors to worry that falling AI service prices might erode the profitability of the industry. However, Bernstein believes that weak revenue in September is more likely to be a short-term phenomenon caused by price adjustments rather than a structural slowdown in AI demand.
The report points out that the revenue performance of AI model developers strengthened again in October, while the amount of data processed by AI models continued to grow rapidly, indicating that falling service prices are driving more users to adopt AI products.
Bernstein believes that falling AI model prices are not necessarily bad for upstream infrastructure vendors. As the cost of using AI decreases, user demand is likely to expand further, thereby increasing the overall demand for computing resources.
The bank also pointed out that even if some leading AI model developers fall behind in competition in the future, it will not necessarily damage the economic benefits of the entire AI infrastructure industry. Previously, when the number of leading model developers was small, AI infrastructure vendors were also able to maintain a healthy business model.
AI infrastructure investment in 2027 will require approximately US$729 billion in revenue support
In order to evaluate whether the scale of AI investment is reasonable, Bernstein established a measurement model based on return on investment. The bank estimates that the cumulative incremental investment in AI infrastructure put into operation by large cloud computing companies will be about 1.2 trillion US dollars by 2027.
Assuming a 15% return on these investments, AI data center operators would need to generate approximately $175 billion in incremental net profit each year.
Further assuming a net profit margin of 24% from related businesses, approximately $729 billion in revenue would be required to support this return target. Bernstein acknowledged that $729 billion is not an easily attainable revenue scale, but believes that this goal is still somewhat feasible.
The bank pointed out that the AI demand outlook is actually weaker at the end of 2024 than it is now. At that time, the market experienced a sharp drop in GPU rental prices due to oversupply of computing power.
However, as AI companies re-invested more computing resources in model training, their reasoning capabilities continued to improve, which subsequently promoted the development of models and AI agent applications with more complex reasoning capabilities, and led to a new round of demand growth.
Bernstein believes that if the supply of computing power tends to be relaxed in the future, AI developers may also adopt similar strategies to drive the next stage of technological breakthroughs by increasing investment in model training. However, at the same time, the bank emphasized that the increase in AI demand is uncertain, and it is still necessary to continue to observe whether AI commercialization revenue can keep up with the scale of infrastructure investment in the future.
The US midterm elections heighten concerns about data center construction, but the impact may be limited in 2027
In addition to the return on AI investment, land, power, and policy restrictions faced by data center construction are also an important source of recent market concerns.
Bernstein pointed out that the Democratic Party's overall victory in the US midterm elections has heightened market concerns about data center project approval and construction progress. Among them, Texas received particular attention. The state is one of the most important regions for AI data center construction in the US, and newly elected governor Gina Hinojosa proposed suspending approval of some data center projects during the election campaign.
Furthermore, the market is also concerned about signs that Google is delaying part of its capital expenditure due to land and electricity supply issues. However, Bernstein believes that the impact of relevant policy changes on AI infrastructure delivery in 2027 may be relatively limited.
The bank pointed out that the relevant measures currently adopted by various states generally allow projects that have already started to advance, and projects that have not yet started will be difficult to complete in 2027, considering the construction cycle.
Take Texas as an example. Among the highly credible data center projects scheduled to be delivered in 2027, the capacity that has not yet started is only about 0.3 gigawatts, and these projects may have been postponed.
In contrast, the 2028 project faced greater policy uncertainty. However, Bernstein believes that local governments generally do not want to completely block data center construction, but rather require relevant companies to bear more infrastructure costs and provide more financial benefits to local communities. Therefore, AI data center operators still have the opportunity to promote project implementation by adjusting construction plans, bearing additional costs, or negotiating with the government.
The bank also pointed out that AI companies currently still tend to extend the usage time of existing GPUs as much as possible, rather than shutting down old devices and using the freed up power to deploy new generation chips that are more efficient. This may indicate that the industry is still more concerned about insufficient chip computing power, rather than thinking that electricity supply has become an insurmountable obstacle.
The macroeconomic environment and capital flows have intensified, and sell-offs may rebound in the short term, but full recovery still needs to be awaited
Bernstein believes that in addition to concerns about industry fundamentals, macroeconomic and market trading factors have also intensified this round of AI stock sell-off. According to the report, some investors may reduce their holdings of stocks in the technology, media and telecommunications sectors in order to prepare for the upcoming IPOs of large technology companies to release liquidity.
Meanwhile, rising US interest rates, market concerns that the US and Iran may break out again, and some short deals push highly leveraged portfolios to close positions, which may also increase downward pressure on the market.
Bernstein believes that the market performance on Thursday was characterized by a certain panic sell-off, so there is a possibility of a technical rebound in AI-related stocks in the short term.
However, the bank remains cautious about the timing of a full recovery.
Bernstein pointed out that the AI sector will lack a sufficiently strong fundamental catalyst for some time to come. The market may need to wait for major cloud computing companies to update their capital expenditure plans in January 2027 and the Nvidia GTC conference in March to obtain more evidence on AI demand and infrastructure investment prospects.
Furthermore, the Philadelphia Semiconductor Index is also entering a period of relatively weak seasonal trading in history. The report notes that in 2024 and 2025, the overall performance of the index from November to December was relatively lackluster.
As a result, Bernstein believes that although the current market may be too pessimistic about the long-term future of AI, investors will still need time to rebuild their confidence.
Maintain the “outperform the market” ratings of many AI industry chain companies and lower the target prices of three companies
In terms of specific investment proposals, Bernstein maintained profit forecasts for the AI network and communications equipment industry, but lowered the target prices for Lumentum, Celestica, and Corning (GLW.US) in view of declining market valuation levels. Among them, Lumentum's target price was lowered from 900 US dollars to 635 US dollars, mainly reflecting changes in consideration valuation of the Maywell Technology (MRVL.US) acquisition transaction.
Celestica's target price was lowered from $495 to $340. Bernstein believes that the company's current valuation is lower than that of some traditional electronics manufacturing service companies, and it is difficult to fully explain it using fundamental factors, but considering the lack of a clear catalyst for growth in the short term, it still adopts a more conservative valuation method.
Corning's target price was lowered from $140 to $120. The bank pointed out that although Corning is expected to benefit from increased demand for AI-related products, non-AI businesses such as telecommunications, smartphones, TVs, and automobiles are under pressure from weak demand in the terminal market, and the market may also be unwilling to continue to give it a high AI valuation premium.
Meanwhile, Bernstein maintained the “outperforming the market” ratings of Arista Networks, Ciena, Celestica, Coherent, and Lumentum, and maintained the “in sync with the market” ratings of Cisco and Corning. The bank gave Arista Networks a target price of $270, Ciena $440, Coherent $350, and Cisco $110.
Bernstein pointed out that most of Arista and Cisco's revenue and profits still come from enterprise customers, so compared to some companies that are highly dependent on AI capital expenditure, they are relatively less affected by concerns about the slowdown in AI investment.
Overall, Bernstein believes that the main problem facing the AI sector is not a sudden deterioration in corporate performance, but rather that investors are beginning to re-evaluate the return prospects of huge AI investments, data center construction restrictions, and reasonable valuations of related companies.
Although these risks are difficult to completely eliminate in the short term, the bank still believes that the continued spread of AI technology and the growth in computing demand will provide long-term support for the industrial chain. However, for investors who expect AI stocks to quickly recover lost ground, Bernstein's judgment is relatively cautious: a short-term rebound may occur, but to actually drive the full recovery of the sector, more clear evidence of demand growth, capital expenditure, and profit returns is still needed.