Selling AI hardware has become the consensus of global investors.
In the past two years, one of the most crowded trades in the global capital market has been buying AI hardware.
NVIDIA's GPUs, Broadcom's custom chips, TSMC's advanced manufacturing processes, as well as optical modules, servers, power supplies and cooling equipment, almost constitute the most definite benefit chain in the AI investment boom.
The logic of investors at that time was not complicated: as long as Microsoft, Google, Amazon and Meta continued to increase capital expenditure, the demand for AI computing power would continue to grow, and hardware companies in the upper reaches of the industrial chain would naturally obtain orders, revenue and profits.
However, by November 2026, the market's trading logic is changing.
In the past, investors worried that they could not buy enough AI chips; now, they are starting to worry that they have bought too many chips. In the past, the market chased order growth and capacity expansion; now, more and more capital is beginning to examine return on capital, customer financing capacity, equipment utilization rate and the final cash flow.
This does not mean that the long-term development logic of the AI industry has ended, but that global investors are re-pricing the AI industrial chain.
In this change, the first to be impacted are precisely the AI hardware companies that previously had the largest increase, the highest valuation and the most consistent market expectations.
01
From buying AI hardware to rushing to reduce exposure
This round of adjustment of AI hardware assets did not start with the sudden deterioration of the performance of a single company, but with changes in multiple market signals at the same time.
On November 5, US semiconductor stocks encountered a sharp sell-off.
The Philadelphia Semiconductor Index fell 12.2% in a single day, and the Nasdaq Index fell 5.5%. This is the largest single-day drop of the Philadelphia Semiconductor Index since April 2025.
Qualcomm's financial report performance fell short of expectations, which became the direct trigger for the market adjustment, but what really made investors uneasy was whether the investment in AI infrastructure could continue to expand at the previous speed.
Reuters then quoted Mark Hackett, chief market strategist at Nationwide, as saying that hedge funds have been actively selling semiconductor and new cloud computing company stocks for several consecutive weeks, and some capital still tends to go short or reduce leverage.
This shows that the market adjustment is not just a short-term profit taking, but some institutions have begun to actively reduce the risk exposure of the AI industrial chain.
This change was quickly transmitted to the Asian market.
On November 16, the AI hardware sector on China's A-share market saw another obvious adjustment. The 5G Communications Index fell by 7.5%, the Science and Technology Innovation 50 Index fell by 5.2%, and Innolight fell by 12.3% in a single day.
At the same time, semiconductor equipment companies in the Japanese market also encountered sell-offs. Tokyo Electron fell by 7.1%, and Advantest fell by 11.1%.
From US chip stocks, to Japanese semiconductor equipment, to Chinese optical module companies, the AI hardware industrial chain has seen obvious cross-market resonance.
More notably, institutional investors are beginning to re-examine how much AI assets they actually hold.
On November 16, Reuters reported that among the world's major enterprises, the scale of investment assets corresponding to companies directly affected by AI has reached at least 25.3 trillion US dollars. For many large investment institutions, AI is no longer an industry theme that can be managed separately, but a common risk source deeply embedded in stocks, bonds and other assets.
PGIM in the United States has even begun to conduct AI risk screening on nearly 1 trillion US dollars of fixed income assets under its management, evaluating both the extent to which the enterprise's business model is affected by AI and the impact that the slowdown in AI capital expenditure may bring to the enterprise's credit status.
The logic behind this is that if there is an adjustment in AI investment, it may not only affect chip companies, but also data center developers, equipment suppliers, power enterprises, and even financial institutions that provide financing for these projects.
Of course, it is not accurate to interpret this change as that global investors have uniformly turned bearish on AI hardware.
For example, an analysis of global customer transaction data by Saxo Bank, which is under Julius Baer, shows that after the sharp drop in semiconductor stocks on November 5, some customers instead actively bought stocks of companies such as NVIDIA, AMD, Micron and SanDisk.
Among them, NVIDIA's average daily buying transaction activity increased significantly compared with before.
This shows that the market has not abandoned AI hardware, but is re-selecting the price and target of purchase.
02
The most dangerous change in AI hardware
If the market adjustment in early November reflects the change in investor sentiment, then the latest AI accelerator prediction model released by SemiAnalysis on November 13 provides a more noteworthy industrial-level basis for this adjustment.
This research institution, which has long tracked global semiconductors and AI infrastructure, has significantly lowered its forecasts for global AI accelerator capital expenditure, chip shipments, HBM demand and power demand in the next few years.
Among them, the most notable one is the forecast of global AI accelerator capital expenditure.
According to its latest model, the global AI accelerator capital expenditure forecast for 2027 has been lowered from the previous 770 billion US dollars to 627 billion US dollars, a decrease of 19%; the figure for 2028 has been lowered from 1.054 trillion US dollars to 758 billion US dollars, a decrease of 28%; the figures for 2029 and 2030 have been lowered by 41% and 40% respectively.
By 2030, the global AI accelerator capital expenditure forecast will drop from the previous 1.531 trillion US dollars to 923 billion US dollars, a reduction of 608 billion US dollars.
This is not a small forecast revision.
It means that the market's previous expectation of sustained rapid growth in AI computing power investment is being recalculated.
Of course, this does not mean that AI infrastructure investment will stop growing. According to SemiAnalysis' adjusted model, global AI accelerator capital expenditure will still grow from about 392 billion US dollars in 2026 to 923 billion US dollars in 2030.
The question is, at what growth rate did investors value these companies before?
When the future investment scale still grows, but the growth rate is significantly lower than the original expectation, the valuation of hardware companies needs to find a new basis.
This is especially true for those companies whose stock prices have already reflected the rapid growth in the next few years in advance.
SemiAnalysis' forecast adjustment also involves a more important change: the price and profit margin of hardware products.
In the past two years, the supply of AI chips has been tight, and HBM memory, high-end packaging and advanced process production capacity have all been in a relatively scarce state.
In an environment where demand exceeds supply, suppliers can not only get more orders, but also obtain higher profits by raising product prices and optimizing product mix.
However, as more production capacity enters the market, the relationship between supply and demand is changing.
SemiAnalysis lowered the global HBM shipment forecast for 2028 from the previous 9.0EB to 6.4EB, a decrease of 30%, and at the same time believes that the HBM supply-demand relationship is improving, and future price assumptions need to be lowered.
For chip suppliers, this means a new problem: even if shipments continue to grow, the average selling price and profit margin may not be able to maintain the previous level.
The impact on NVIDIA is particularly noteworthy.
SemiAnalysis lowered NVIDIA's shipment forecasts related to computing revenue from 2027 to 2030 by 22%, 39%, 56% and 66% respectively, and lowered the average selling price assumptions for some new generation products.
Behind this is not only the factor of slowing demand growth, but also the change in the competitive landscape.
In the past, NVIDIA GPUs were almost the first choice for large-scale AI model training and inference. But now, Google TPU, Amazon Trainium, and custom ASICs co-developed by Broadcom are becoming another choice for large AI companies.
For giants like Microsoft, Google, Amazon and Meta, they do not necessarily need to rely on the same chip architecture forever.
As long as other solutions can provide better performance-to-cost ratio, lower energy consumption or a more controllable supply chain, they will be motivated to adjust their procurement structure.
For NVIDIA, this means that even if the global computing power demand continues to grow, it may not be able to obtain the same market share and profit margin as in the past.
In contrast, Broadcom performed relatively steadily in this forecast adjustment. SemiAnalysis lowered its cumulative ASIC business revenue forecast from 2026 to 2030 by 11%, which is mainly related to changes in HBM cost and product price assumptions, rather than a large-scale reduction in shipments.
This also explains why the AI hardware industrial chain cannot be simply regarded as a whole.
The competitive advantages, customer structure and product positioning of different companies are determining their performance in the next cycle.
However, what really worries the market is not just the competition among chip suppliers.
The deeper question is, who will continue to pay for this round of AI infrastructure construction?
In the past, the market believed that large cloud computing companies had strong enough cash flow to support years of sustained AI capital expenditure.
But now, the scale of investment is getting larger and larger.
JPMorgan Private Bank's global investment strategy report released in November pointed out that in the third quarter of 2026, the capital expenditure of three large cloud computing companies was lower than market expectations, and the upward adjustment of the capital expenditure forecast for the next 12 months also dropped to the lowest level since the fourth quarter of 2023.
This does not mean that Microsoft, Google, Amazon and Meta will stop investing in AI, but that these companies have begun to pay more attention to investment rhythm and financial discipline.
A study released by the Federal Reserve Bank of St. Louis in November also revealed the change in the capital structure behind this round of AI investment.
The study found that the cash asset ratio of large cloud computing companies dropped from 35% in 2018 to 16% in 2026, while their investment asset ratio nearly doubled.
More notably, there are obvious differences in the financing status between data center construction enterprises and semiconductor companies.
From the fourth quarter of 2024 to the second quarter of 2026, the average quarterly net debt financing scale of data center construction enterprises increased from 3.23 billion US dollars to 30.17 billion US dollars; while semiconductor enterprises shifted from net borrowing to net debt repayment.
This means that the financing pressure of the AI industrial chain is not evenly distributed.
Chip suppliers can still rely on strong operating cash flow to support business development, but some data center builders and new cloud computing enterprises need to rely more and more on external financing.
Once the financing environment tightens, or customers are no longer willing to commit to procurement in advance, some data center projects may be delayed, which will in turn affect orders for chips, servers, network equipment and power equipment.
This is why the market has begun to re-examine the order quality of the AI industrial chain.
In the past, as long as customers announced the construction of data centers and signed procurement agreements, suppliers could gain market recognition.
Now, investors are more concerned about: Do customers have sufficient funds? When will the project be delivered? Can the equipment be put into use? Can the final revenue cover depreciation, interest and operating costs?
These questions are replacing pure order growth and becoming an important factor determining the valuation of AI hardware companies.
03
The era of buying blindly is over
For global AI hardware companies, the real difficulty may not be the current drop in stock prices, but the change in the way the market evaluates them in the next few quarters.
In the past, the most important competitive advantage of AI hardware companies was whether they could enter the supply chains of NVIDIA, Microsoft, Google, Amazon and Meta.
As long as they entered the supply chain and obtained orders, the market was willing to give a higher valuation.
But now, supply chain status is no longer a sufficient condition.
They are all AI chip companies, but some have long-term procurement commitments from large cloud computing customers, while others rely on new cloud service providers with weak financing capacity; they are all optical module companies, but some are in the cycle of technological upgrading, while others may face pressures such as customer concentration, price competition and slowing capital expenditure.
These differences will gradually be reflected in the company's revenue growth rate, gross profit margin and cash flow.
Chinese AI hardware companies are facing the same problem.
In the past year, the AI hardware trading in the A-share and Hong Kong stock markets has largely benefited from the growth of capital expenditure of overseas cloud computing giants.
Companies such as Innolight, Eoptolink, and TFC Communications are in different links of the industrial chain, but their pricing in the market is closely related to the construction of global AI data centers.
If overseas cloud manufacturers continue to expand capital expenditure, Chinese suppliers will have the opportunity to get more orders; but if the pace of global data center construction slows down, or customers begin to adjust their equipment procurement plans, the Chinese supply chain can hardly stay completely out of the trouble.
The sharp drop in Innolight's stock price on November 16 is a microcosm of this global industrial chain risk transmission.
Of course, China's AI industry also has its own development logic.
Domestic AI chips, domestic servers and data center construction are forming an independent demand market. Overseas chip supply restrictions may also accelerate the localization of domestic computing power infrastructure.
But domestic substitution does not mean that it can fully hedge the changes in the global AI investment cycle.
For hardware suppliers, in the end, they still have to answer the same question: Do customers have real demand, does the project have continuous financial support, and can the products generate stable profits?
From this perspective, what global investors are carrying out is not a simple retreat from the AI industry, but a switch of investment logic.
In the first stage, the market bought the scarcity of computing power.
In the second stage, the market bought order growth.
And next, the market may be more willing to pay a premium for enterprises that can create real cash flow, have stable customers and strong pricing power.
This also means that the differentiation within the AI hardware industrial chain may further increase.
Companies with technical barriers, customer stickiness and cost advantages can still benefit from the next round of industrial expansion; but those companies whose valuations are mainly supported by industry prosperity, customer financing and high capital expenditure expectations need to face stricter market tests.
For investors, what they need to pay attention to in the future is no longer just how many of NVIDIA's next-generation GPUs can be sold, or how much the global demand for optical modules can grow.
More importantly, we need to observe whether the capital expenditure of large cloud computing companies continues to rise, whether AI application revenue can cover computing power costs, whether the utilization rate of data centers can be improved, and whether hardware suppliers can maintain profit margins in price competition.
If these indicators cannot continue to improve, even if the AI industry is still in a growth cycle, the stock prices of some hardware companies may continue to be under pressure.
The AI revolution will not end because of a market adjustment.
But the way the capital market prices AI hardware has changed.
In the past, investors believed that as long as AI continued to develop, people who sold shovels would definitely make money.
Now, they are starting to ask another question: If everyone is buying shovels, how many people can eventually make money by mining?
This may be the real reason why global AI hardware trading changed in November 2026.
This article is from the WeChat official account "Alpha Workshop Research Institute", author: Shi Xia, published with authorization from 36Kr.