Looking back at 2025, technology became the main theme in China's A-share market.

Within technology, the AI sector was by far the brightest spot. AI computing power names such as the so-called “Yi Zhong Tian” trio, Eoptolink, Zhongji Innolight and TFC Communication, and the “Ji Lian Hai” group, Cambricon, Foxconn Industrial Internet and Hygon Information, formed the backbone of tech investing. By contrast, AI applications and edge AI devices lagged AI computing power in share-price gains.

Miaotou believes the core reason AI computing power generated excess returns was certainty, coupled with repeated upside surprises.

Capital and companies were willing to bet on the certainty of current and future AI computing power infrastructure buildouts. Examples include policies promoting domestic GPU substitution and increased capital spending by the four major North American cloud providers as well as Chinese internet companies such as BAT.

By contrast, large-model companies still showed little sign of profitability in 2025, while sales of edge products such as AI phones and AI glasses remained weak. These areas did not give capital markets the same sense of investment certainty.

Still, global tech giants are in a high-intensity, high-growth capital spending cycle, and the market broadly expects this wave of investment in AI and other core technologies to continue over the next few years.

So what will be the main AI investment theme in 2026?

Looking ahead to 2026, Miaotou believes AI computing power still offers strong certainty and is likely to remain the main technology theme. As AI infrastructure is built out and upgraded, subsegments including computing power (GPUs), transmission capacity (optical modules), liquid cooling and PCBs will present investment opportunities.

Miaotou has compiled the market sizes of AI computing power subsegments. Based on compound annual growth rates from 2024 to 2029, the ranking is: China AI data center GPUs > China AI data center liquid cooling > optical chips > AI-related PCBs with 18 or more layers > optical modules.

Next, we look at each in turn.

We start with China AI data center GPUs, the segment with the highest compound growth rate.

From a market-structure perspective, NVIDIA still holds a high share, while domestic GPU makers are rising on policy guidance such as domestic substitution.

Unlike NVIDIA in the U.S. stock market, the valuation logic for Chinese GPU makers cannot be fully quantified and is supported by the broader narrative of domestic substitution.

That said, GPUs are beginning to move into the earnings-delivery phase.

Newly listed GPU makers MetaX and Moore Threads expect to become profitable as early as 2026 and 2027, respectively. Cambricon returned to single-quarter profitability in Q4 2024 and was already profitable in the first three quarters of 2025.

Cambricon's revenue ramp depends on validation and deployment by internet customers, which is why revenue surged in 2025.

In other words, capital spending by Chinese internet companies on AI infrastructure is the main growth driver for domestic GPU makers such as Cambricon.

China's major technology companies are also increasing capital investment. Alibaba, Tencent and Baidu lifted capital expenditure in Q1 to Q3 2025 by 132.46%, 48.24% and 74.49% year on year, respectively.

Against the backdrop of GPU domestic substitution and internet companies stepping up AI investment, some institutions forecast Cambricon's net profit attributable to shareholders at 4.872 billion yuan in 2026 and 7.991 billion yuan in 2027, up 118.71% and 63.99% year on year, respectively.

Even so, capital markets remain concerned about whether valuations for GPU makers such as Cambricon are too high.

On valuation, Cambricon's market capitalization is 571.6 billion yuan. Based on its expected 2027 net profit of 7.991 billion yuan, its price-to-earnings ratio would be 71.48. NVIDIA's trailing P/E is 45.94.

Miaotou therefore believes that although Cambricon has already priced in two years of earnings expectations, its valuation still contains a “bubble” compared with NVIDIA.

To digest that “bubble,” Cambricon and its peers will need higher earnings expectations, which in turn requires stronger capital spending by domestic internet companies on AI infrastructure.

Still, it is indeed possible that China's major internet platforms will step up AI investment and raise capital spending.

On December 23, foreign media reported that ByteDance plans to invest 160 billion yuan in AI in 2026, with half of the budget allocated to AI chip procurement. Based on estimated 2025 profit of $50 billion, ByteDance's 2026 AI investment would amount to nearly half of its full-year 2025 profit.

After the news gained traction, shares of Cambricon and other GPU makers recovered somewhat.

Miaotou believes that in 2026, higher valuations for GPU makers represented by Cambricon will have to be driven by expectations, namely upward revisions to capital spending by major Chinese internet companies. Investors should still watch whether Alibaba, Tencent, Baidu and others raise capital expenditure, as GPUs still have room to beat expectations in 2026.

Liquid cooling is one of the highest-certainty, high-growth tracks in the AI computing power era, and the industry is now in the early stage of a breakout as penetration accelerates.

Liquid cooling replaces air with liquid media for heat dissipation, providing stronger cooling capacity. For example, single-phase cold-plate liquid cooling is far more efficient than air cooling. As the heat flux density of AI chips, such as NVIDIA's GB300, exceeds 500W/cm², traditional air cooling can no longer meet high-power requirements, and liquid cooling is moving from an “optional” technology to a “must-have.”

According to institutional forecasts, global data center liquid-cooling penetration will rise from 14% in 2024 to 31% in 2026.

Liquid-cooling service providers still follow a business model built around major customers, similar to Apple's supply chain.

NVIDIA began large-scale adoption of liquid cooling with GB200 NVL72, and Google's TPU v7p and other products are also expected to deploy liquid cooling. As a result, capital markets are paying more attention to liquid-cooling service providers that have entered the supply chains of NVIDIA, Google and others.

However, NVIDIA and Google use different liquid-cooling supplier models.

Liquid-cooling systems for NVIDIA's GB200/300 and other products are mainly led by long-term high-end suppliers such as Vertiv and Delta. Domestic companies typically enter as second-tier component suppliers or full-chain solution providers. Companies such as Envicool and BYD Electronics have already been included in RVL/AVL recommendation lists.

This “recommended list plus ODM free decision-making” model lets domestic second-tier suppliers see large orders but makes it hard for them to lock in long-term large orders, leaving order stability relatively weak.

For second-tier component suppliers, Google's liquid-cooling orders are much “steadier” than NVIDIA's and offer greater certainty. Google directly connects with and designates liquid-cooling component suppliers.

Because vendors will all push liquid-cooling solutions in 2026, Google may shift to developing new liquid-cooling suppliers with lower shares in the NVIDIA system in order to avoid conflicts in supply and capacity.

In other words, Google could bring an unexpected upside to domestic liquid-cooling suppliers. For example, Envicool has already won a bid for a Google data center project, and Suzhou Shiquan New Materials' ultra-thin VC heat spreader has passed Google certification.

Against the backdrop of expected liquid-cooling volume growth in 2026, Miaotou believes domestic liquid-cooling component suppliers that can enter Google's supply chain may gain opportunities for valuation expansion.

Investors also need to watch whether supply-chain expectations around Google orders are realized. For example, Envicool attracted capital after entering Google's supply chain. On valuation, based on Envicool's 2027 earnings expectations, its P/E has reached 74.56, above the sector average.

Miaotou believes other liquid-cooling suppliers with lower valuations could see greater valuation elasticity once they enter the supply chains of Google and other major customers. Investors should keep tracking changes in the liquid-cooling supply chains of Google and NVIDIA.

If GPUs handle calculation and represent computing power, optical modules handle transmission and represent “transport capacity.”

The performance of optical modules directly determines the efficiency and stability of data transmission. If data transmission cannot keep up, it is like a traffic jam on a highway: no matter how strong the computing power is, it goes to waste.

There are two investment logics for optical modules:

Product Technology Iteration, such as the increase in transmission speed from upgrading 800G optical modules to 1.6T optical modules;

Downstream Demand Growth, namely capital spending by cloud providers such as Google, Amazon, Meta, Microsoft and BAT on data center construction.

Among A-share optical module companies, the “Yi Zhong Tian” trio, Eoptolink, Zhongji Innolight and TFC Communication, are the most representative. Since the start of 2025, they have risen 450%, 422% and 231%, respectively, with their share prices repeatedly hitting new highs.

This has also made investors worry about a “bubble.”

Based on Wind consensus expectations, Zhongji Innolight, TFC Communication and Eoptolink are expected to deliver net profits attributable to shareholders of 25.297 billion yuan, 3.873 billion yuan and 20.670 billion yuan in 2027, respectively, corresponding to P/E ratios of 27.54, 42.08 and 21.26.

Miaotou previously argued in “‘Yi Zhong Tian’s’ Bubble Has Been Priced to 2027” that the “high valuations” supporting optical modules come from expectation gaps, such as upward revisions to cloud providers' capital spending, Google's TPU chips, and increases in optical module demand driven by network pricing, or scale-up.

These “better-than-expected” logics are now being, or have already been, realized.

In other words, earnings expectations for optical modules have already been priced out to 2027. Still, institutions such as Goldman Sachs remain bullish on the segment. Goldman, for example, raised its target price for Zhongji Innolight to 762 yuan.

Miaotou believes that in 2026, higher valuations for optical module companies will still require a new “narrative” or positive catalyst.

In 2026, investors should focus on whether North American cloud providers again revise capital expenditure upward, and on new opportunities from optical module speeds moving toward 1.6T as well as silicon photonics, OCS, CPO and other emerging technology trends. All of these could become catalysts.

Optical chips are the core components of optical assemblies and optical modules, and their performance directly determines information transmission speed and network reliability.

The investment logic for optical chips is broadly similar to that for optical modules, but with an added domestic-substitution angle.

In terms of market structure, European and U.S. companies such as Broadcom, Lumentum and Coherent control the optical chip market. According to ICC data, overseas vendors account for about 75% of the 25G optical chip market and about 95% of the market for optical chips above 25G.

Against the backdrop of surging shipments of 800G and 1.6T high-speed optical modules, demand for 100G and above optical chips is strong, and growth will far exceed that of medium- and low-speed optical chips.

Lumentum said on its Q3 2025 earnings call that the optical chip supply-demand imbalance had worsened. Even after adding capacity, the company was still in a position where it had to make allocation decisions almost every day. The gap relative to total customer demand has now risen from 20% in the previous quarter to 25% to 30%. Looking to 2026, high-end optical chip prices are expected to rise given the supply-demand imbalance.

Miaotou believes that if domestic optical chip companies achieve breakthroughs in high-end optical chips in 2026, their valuation elasticity will be greater than that of optical module companies.

The phones, computers, routers and air-conditioner remotes we use every day all contain one or more green boards, though they may be other colors. Those boards are PCBs, or printed circuit boards.

Compared with traditional servers, AI servers require significantly higher PCB performance in transmission speed, layer count and density, pushing technology roadmaps toward high-end categories such as high-multilayer boards and HDI.

This has created an investment logic for PCBs in which both volume and price rise.

On volume, Prismark forecasts that from 2024 to 2029, AI-related PCB boards with 18 or more layers will grow at a compound annual rate of 20.6%, far above the average growth rate of the PCB industry.

On price, the PCB value per AI server can reach 5,000 yuan, more than three times that of a traditional server.

As AI servers iterate, upgrades in PCB/CCL materials and processes are creating new value increments.

For example, mainstream general-purpose servers have already adopted M6-grade copper-clad laminate, while M7 and M8-grade CCL are the main materials used in AI servers and 400G/800G switches. Each increase in layer count and material upgrade raises the PCB's price per square meter.

NVIDIA expects Rubin GPUs to enter mass production in the second half of 2026.

NVIDIA's Vera Rubin architecture will use an orthogonal backplane, a specially designed PCB, to replace traditional copper cables. The backplane is made by laminating three 26-layer PCBs or four 26-layer PCBs, combined with a high-end material stack including M9 resin substrates, high-grade HVLP4 copper foil and Q cloth, or quartz fiber cloth.

As a result, orthogonal backplanes are expected to ramp in 2026, driving growth in PCBs, CCL and other raw material segments.

In addition, NVIDIA plans to launch its second-generation Rubin Ultra NVL576 platform in the second half of 2027. Its ultra-high-layer PCB backplane is expected to be designed with M9+Q cloth or M9.5+Q cloth materials. If adopted, this board material would open another new growth avenue for demand for PCBs used in NVIDIA AI servers.

Notably, high-end PCB capacity supply is expected to remain tight in 2026. According to China Merchants Securities estimates, effective PCB capacity among listed Chinese companies that can meet AI demand is about 120 billion yuan, while demand is expected to be about 150 billion yuan.

Miaotou therefore believes that as GPU architectures continue to iterate, the PCB industry has strong growth expectations for 2026. However, PCB companies also saw large share-price gains in 2025.

Based on current earnings forecasts and valuations for some PCB companies, many PCB names have P/E ratios close to or even above 50 times. They are currently in an overvalued state and will need high earnings growth over the next two years to digest those valuations.

Miaotou believes that if relevant PCB companies can also deliver high growth in line with 2026 expectations, their valuations may still have room to rise. NVIDIA Rubin GPU mass production in 2026 will be an important factor affecting earnings growth and valuation expansion for related PCB companies, and investors should watch it closely.

After all that, how should investors rank the thematic opportunities across AI sectors?

Miaotou believes AI sectors can be ranked from strongest to weakest according to investment certainty.

The liquid-cooling industry has a high growth rate and could produce “dark horse” companies that enter Google's or NVIDIA's supply chains, so it can be ranked as “solid.” GPUs, optical modules and PCBs have high growth certainty, but valuations are also high, so they need upside surprises to drive valuation expansion and can be ranked as “elite.” Because of expectations for price increases and technology breakthroughs, optical chips have somewhat higher elasticity than optical modules and can be ranked as “top tier.”

In 2026, the certainty around edge AI devices and AI applications is not high, and new catalysts are needed, such as the launch of a new hit edge AI product or a new hit AI application. Once catalysts emerge, edge AI and AI applications will have large expectation gaps and elasticity. For now, however, they can only be ranked as “NPC.”

In addition, expectations for a looser macro environment will bring a liquidity premium to technology sectors represented by AI computing power. The Federal Reserve is likely to cut rates twice in 2026, which could improve liquidity for A-shares, and high-growth areas such as AI computing power will receive higher valuations from the market.

Miaotou believes that in 2026, companies across AI computing power segments such as liquid cooling, optical chips, GPUs, PCBs and optical modules are likely to deliver varying degrees of earnings growth. If growth beats expectations, related companies could achieve a Davis double play under expectations for looser liquidity.

Overall, AI computing power offers high certainty and still has the potential to exceed expectations. Compared with other tracks, it is more likely to deliver excess returns and remains worth watching in 2026.