Has the computing power sector finally staged a rebound from its lows? What has been happening in this sector lately?
Since October, semiconductor materials and equipment, PCB, optical modules, and computing power leasing all suffered sharp declines across the board, but today's opening finally brought a rebound. Beyond shifts in market risk appetite, the competitive landscape of the industry has also changed — recently there has been extensive media coverage of NVIDIA (NVDA) at its Berlin GTC regarding physical AI, robotics platforms, and its "open attitude" toward cooperation with a leading mainland memory manufacturer, yet there has been very little analysis of another piece of news from the same period: NVIDIA announced it had signed a new round of large-scale capacity expansion agreements with six suppliers — Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL, and Unimicron — covering four important segments: substrates, advanced packaging, testing, and optical engines.
But none of these six companies comes from mainland China. Geographically, they precisely cover Japan, South Korea, Taiwan region of China, the United States, and Singapore, like a deliberately drawn circle. Who is this circle meant to keep out? And what is it trying to lock in? What impact does it have on A-share computing power chain companies?
The six companies: which segments and what expansion pace
These six suppliers mainly involve four major segments.
Substrates, three companies: IBIDEN (Japan) is expanding its Ono plant to three times its current size, with ramp-up in the first quarter of 2028, and NVIDIA providing billions of dollars in "capacity reservation prepayments"; Samsung Electro-Mechanics (South Korea) is newly joining this round, building new and expanding facilities in Sejong and Busan; Unimicron (Taiwan region of China) will begin production of KF substrates in the second quarter of 2027, reaching full capacity by mid-2028. The overall goal is to raise total substrate capacity to 2.5 times the current level within three years.
Advanced packaging, one company: Amkor's interposer wafer capacity in South Korea will increase by more than 300% by the end of 2028.
Testing, one company: NVIDIA is prepaying US$1.1 billion to KYEC to build a new testing plant in the United States, expected to enter mass production in April 2027; its Singapore plant is expected to enter mass production in February 2027, doubling total testing capacity within two years.
Optical engines, one company: SPIL's CPO optical engine capacity is expected to reach 15 times the current level by mid-2027, and double again by the end of 2028.
This list is a rare case of NVIDIA giving an expansion plan with "specific timing plus multiples," and the money for expansion is advanced by NVIDIA itself to drive it forward. This is the same playbook as previously locking in more than half of TSMC's CoWoS capacity for 2026 (about 800,000 to 850,000 wafers) and prepaying US$1.5 billion to Amkor in July this year.
Among the four major segments involved — ABF substrates: mainland manufacturers currently hold a combined market share of about 4%, with Xingsen Technology accounting for 45% to 50% of the Ascend chain and Shennan Circuits ranking second, yet both are basically invisible on NVIDIA's chain; outsourced advanced packaging: TSMC alone accounts for more than 85% of CoWoS; the top five mainland players combined hold 32.6%, but the domestic localization rate for high-end 2.5D/3D packaging is less than 10%; independent testing: KYEC alone takes 32% of global independent testing; the top four mainland players had combined revenue of 4.57 billion yuan in 2025; optical engines: China dominates pluggable optical modules, with seven of the top ten on the LightCounting list being Chinese manufacturers, and Zhongji Innolight Co.,Ltd. (300308) holding more than 20% revenue share in 2025 and more than 40% in 800G. But the official division of labor for NVIDIA's CPO is a different table: silicon photonics wafers go to TSMC, packaging and testing go to SPIL, laser chips go to Lumentum, laser module subcomponents go to Tianfu Communication, and complete units go to Foxconn.
To sum it up in one sentence: of the four segments locked in this time, three are areas where mainland manufacturers have weak share, and this cold arrow is aimed at where new players want to go; while in the strongest optical module segment, new rivals are being cultivated for mainland leaders.
Two further questions
Question one: What is NVIDIA's real intention?
Using prepayments to reserve capacity is nothing new: TSMC used a "customer deposit" mechanism back in the 1990s, and during the 2021 chip shortage Tesla also prepaid to secure capacity. NVIDIA itself has been doing this for the past three years — it took more than half of CoWoS, leaving AMD able to grab only the remaining reservations in the first half of 2027, while Broadcom was allocated 240,000 wafers in 2026. What is different this round is geography: capacity is specifically locked into Japan, South Korea, Taiwan region of China, the United States, and Singapore, with mainland China temporarily not on the expansion list. The intention has at least three layers.
First, securing supply: seven Vera Rubin chips have already been taped out, with mass production moved up to the third quarter of 2026, and the first batch shipping to Microsoft, Google, Amazon, Meta, and Oracle in July. Second, squeezing competitors: the CoWoS era already proved that whoever holds capacity reservations holds the shipment rhythm, and this round extends the same logic to substrates, testing, and optical engines. Third, using economic means to achieve geopolitical hedging: buyers pay, sellers build factories — seemingly a routine economic decision.
Question two: Will new CPO players suppress the bargaining power of Zhongji Innolight Co.,Ltd. (300308) and its peers?
Look at the numbers first. In 2025, the global optical communication device market was about US$24.1 billion, while CPO switches are forecast at only US$3.2 billion in 2026; CPO is only being deployed on the switch side (the Spectrum-X silicon photonics switch only entered full mass production in August), and GPU-side links still rely on pluggable modules. Zhongji Innolight Co.,Ltd. (300308) had revenue of 38.2 billion yuan and net profit of 10.8 billion yuan in 2025, with a 50% to 70% share in 1.6T. Its interim and third-quarter reports this year charged ahead strongly. The same is true for Eoptolink. Over the next two years, bargaining power is most likely safe.
Historical lessons
In 1987, Japan's semiconductor share of the global market once exceeded 50%, but the peak came in that very year. The 1986 U.S.-Japan Semiconductor Agreement forced Japan to open its domestic market, 1987 brought 100% punitive tariffs, and the second agreement in 1991 directly required foreign products to exceed 20% of Japan's domestic share. By 1993, the United States had reclaimed the throne. Then came the financial crisis, Micron's sale, and the Japanese semiconductor industry was wiped out. Back then, trade rules were used; now, prepayments are used: large buyers redraw the map of production origins with their wallets and market access. NVIDIA's cold arrow is not aimed at the heart, but at trying to block the path of upgrading: squeezing away optical module bargaining power layer by layer, locking substrates, advanced packaging, testing, and other segments into overseas chains, leaving domestic manufacturers able to retreat only to domestic customers.
But this both matters and does not matter. Being besieged and counter-besieging has almost become routine, and it has long been "you fight your war, I fight mine." In the past two years, more and more people should understand a principle: a country or an industry does not get the opportunity to develop just because it wants to. Or perhaps, you may first be allowed to develop a bit, and then comes the "seizure." The competition of alliances and counter-alliances, of bayonets drawn and blood spilled, is just that cruel. What matters is that we must not lose the sense of crisis of "having reached the most dangerous moment" and the spirit to strive upward.
Related ETF
ChinaAMC ChiNext Computing Power ETF (158041) tracks the ChiNext Computing Power Infrastructure Index, anchoring the infrastructure segment of the AI industry chain. Among its exposures, the "data center (AIDC)" concept accounts for 91.03%, the highest in the entire market; the chip concept accounts for 62.04%; the co-packaged optics (CPO) concept accounts for 53.52%; the computing power leasing concept accounts for 38.84%; and the liquid cooling server concept accounts for 30.95%.
References: 1. NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout, NVIDIA official press release. 2. Jensen Huang's GTC Berlin European debut: all seven Vera Rubin chips taped out, 2026-10-23. 3. Locking in 850,000 wafers: NVIDIA swallows more than half of TSMC's 2026 CoWoS packaging capacity, 2025-12. 4. Ibiden secures advance payments from Google, Amazon, and Intel for packaging substrate capacity. 5. 2025 Global Outsourced Semiconductor Assembly and Test (OSAT) TOP10 list, Core Thought Research Institute. 6. Packaging substrate market landscape: ABF substrates dominated by Taiwan region of China/Japan, Minmetals Securities. 7. Weice Technology, A-share prospectus (including domestic and international independent testing market data). Note: This article represents the author's views and is for reference only.
Scroll down to view product information and risk disclosures:
Data sources: Wind, iFind, listed company announcements, ChinaAMC, the above individual stocks are not recommendations, as of October 23, 2026. Performance of the ChiNext Computing Power Infrastructure Index: the index's full-year returns for 2023-2025 were 48.46%, 34.97%, and 79.34%. Historical index performance does not predict future performance of the fund product. Funds involve risks, and investment requires caution. An ETF is an exchange-traded index fund; secondary market prices are subject to premium/discount and liquidity risks, and fund performance may deviate from the tracked index due to tracking error; past performance does not represent future returns. The above fund is classified as R4 (medium-high risk). It is an equity fund, with risk and return higher than hybrid funds, bond funds, and money market funds. Individual stocks are not recommendations. Before investing in a fund, investors should carefully read the fund's legal documents such as the Fund Contract, Prospectus, and Product Information Summary, fully understand the fund's risk-return characteristics and product features, and fully consider their own risk tolerance based on investment objectives, investment horizon, investment experience, and asset status. On the basis of understanding the product and suitability opinions, investors should make rational judgments and prudent investment decisions, and independently bear investment risks. Index performance does not represent product performance, and secondary market price performance does not represent net asset value performance. This fund is an ETF fund, and investors investing in this fund face potential risks such as failure to meet the agreed tracking error control target, the index compiler ceasing services, and constituent securities being suspended from trading; risk of deviation between the benchmark index return and the average stock market return; risk of benchmark index volatility; risk of deviation between the fund portfolio return and the benchmark index return; risk of benchmark index changes; risk of premium/discount in the secondary market trading price of fund shares; risk of errors in the subscription and redemption list; risk of errors in reference IOPV decisions and IOPV calculations; delisting risk; risk of investor subscription and redemption failure; liquidity risk in converting redemption consideration for fund shares; and derivatives investment risk, among others. MACD golden cross signals have formed, and these stocks are rising well!
Comments