Jibang Consulting: Global AI server shipments are expected to increase by more than 20% year on year in 2026, and the penetration rate of AI chip liquid cooling will reach 47%

The Zhitong Finance App learned that in 2026, thanks to the increase in capital expenditure by large CSPs in North America and the rise of sovereign clouds in various countries, there is strong demand for AI data center construction. It is estimated that global AI server shipments will increase by more than 20% over the same period last year. As the computing power of AI chips increases, the single-chip thermal design power consumption (TDP) will rise from 700W of NVIDIA H100 and H200 to 1,000W or more of B200 and B300. Server cabinets must use liquid cooling systems to meet high-density heat flux requirements, boosting the liquid cooling penetration rate of AI chips to 47% in 2026.

On November 27, the “MTS 2026 Storage Industry Trend Seminar” hosted by TrendForce Jibang Consulting, a global high-tech industry research institute, was held in Shenzhen. At the conference, TrendForce Jibang Consulting released the “2026 Top Ten Technology Market Trend Forecast”:

PART: 1 AI chip battle upgrade, liquid cooling penetrates AI data centers on a large scale

In 2026, thanks to increased capital expenditure by large CSPs in North America and the rise of sovereign clouds in various countries, demand for AI data center construction is strong. It is estimated that global AI server shipments will increase by more than 20% per year. NVIDIA (Nvidia), which dominates the AI market, will face more intense competition. First, AMD (Superpower) will follow the NVIDIA GB/VR cabinet solution and launch MI400 full-cabinet products, focusing on CSPS customers; secondly, the strength of CSPS self-developed ASICs in North America will continue to increase; finally, due to the international situation, ByteDance (ByteDance), Baidu (Baidu), Alibaba (Alibaba), Tencent (Tencent) self-developed ASICs, and Huawei (Huawei), Cambricon (Cambricon) and others strengthened independent research and development of AI chips, making competition in the AI market heated up.

As the computing power of AI chips increases, the single-chip thermal design power consumption (TDP) will rise from 700W of NVIDIA H100 and H200 to 1,000W or more of B200 and B300. Server cabinets must use liquid cooling systems to meet high-density heat flux requirements, boosting the liquid cooling penetration rate of AI chips to 47% in 2026. Microsoft (Microsoft) is also proposing next-generation chip package-level microfluidic cooling technology. Overall, the short to medium market is still dominated by water-cooled plate liquid cooling. The CDU architecture will shift from L2A (liquid-to-air) to L2L (liquid-to-liquid) design, and in the long run, evolve towards more refined chip-level cooling.

PART 2: Breaking through bandwidth limitations and achieving high-speed transmission, HBM and optical communication build a new intelligent computing system

The amount of data and memory bandwidth requirements for AI computation from training to inference are growing explosively, causing bottlenecks in transmission speed and energy consumption to surface. In order to solve the problem that AI computation is limited by memory bandwidth and data transmission rate, HBM and optical communication technology are gradually becoming the core breakthroughs in next-generation AI architectures.

Currently, HBM uses 3D stacks and TSV technology to effectively shorten the distance between processor and memory, and is introducing higher channel density and wider I/O bandwidth into the upcoming mass production HBM4 to support large-scale computation of AI GPUs and accelerators. However, when the model parameters broke through the megabyte level and the GPU cluster size expanded exponentially, the memory transmission bottleneck was once again highlighted. Currently, various memory manufacturers have increased the local bandwidth of AI chips through HBM stack structure optimization, packaging and interface innovation, and collaborative design with logic chips.

After solving the memory transmission bottleneck, data transmission across chips and modules has become a new bottleneck limiting system efficiency. To break through this limitation, optoelectronic integration and CPO (Co-Packaged Optics) technology have gradually become R&D priorities for mainstream GPU manufacturers and cloud vendors. At this stage, mass production of 800G/1.6T pluggable optical modules has begun, and it is expected that higher bandwidth SiPh/CPO platforms will be introduced into AI switches (Switches) starting in 2026. New optical communication technology is used to achieve high-bandwidth, low-power data interconnection, and optimize the overall bandwidth density and energy efficiency of the system.

Looking at trends, the memory industry is moving towards “bandwidth efficiency” as its core competitiveness. The new optical communication technology for processing between chips and modules is also the best solution to break through the limitations of electrical interfaces in long-distance and high-density data transmission. Therefore, high-speed transmission technology will become a key direction in the evolution of AI infrastructure.

PART 3: NAND Flash Vendors Strengthen AI Solutions to Accelerate Inference Work and Reduce Storage Costs

AI training and inference work requires high-speed access to large data sets with unpredictable I/O patterns, creating a performance gap with existing technology. To this end, NAND Flash vendors are accelerating the promotion of specialized solutions, which include two key products: storage level memory (SCM) SSD/KV Cache SSD/HBF technology, which is positioned between DRAM and traditional NAND, providing ultra-low latency and high bandwidth characteristics, making it an ideal choice for accelerating real-time AI inference workloads.

Another is the Near Line QLC SSD. QLC technology is being applied at an unprecedented speed to the hot/cold data storage layer of AI, such as model checkpoints and data set archiving. The storage capacity per grain of QLC will be 33% higher than that of TLC, greatly reducing the unit cost of storing huge AI data sets. It is estimated that by 2026, the market penetration rate of QLC SSDs in Enterprise SSDs will reach 30%.

PART 4: Energy storage systems will rise to the energy core of AI data centers, and demand will usher in explosive growth

AI data centers are moving towards large-scale clustering. Their load fluctuates greatly, and power stability is strictly required, causing the energy storage system to change from “emergency backup” to “the energy core of AI data centers.” It is estimated that in the next five years, in addition to existing short-term UPS backup and improvements in power quality, the proportion of medium- to long-term energy storage systems of 2 to 4 hours will rapidly increase to meet the needs of backup, arbitrage, and power grid services at the same time. The deployment method will also gradually infiltrate from a centralized battery energy storage system (BESS) at the data center level to distributed BESS at the cabinet level or cluster level, such as battery backup units, to provide faster instantaneous response.

North America is expected to become the world's largest AI data center energy storage market, dominated by hyperscale cloud vendors. China's “East Digital Western Computing” strategy will drive the migration of data centers to the western region where green power is abundant, and AI data centers+energy storage will become standard equipment for large bases in the west. The new energy storage capacity of global AI data centers is expected to surge from 15.7 GWh in 2024 to 216.8 GWh in 2030, with a compound annual average growth rate of 46.1%.

PART 5: AI Data Centers Move Towards 800V HVDC Architectures, Boosting Third-Generation Semiconductor Market Demand

Data centers are undergoing a complete transformation in power infrastructure. Server cabinet power is rapidly rising from kilowatts (kW) to megawatts (MW), and the power supply mode is shifting to 800V HVDC (high voltage DC) architectures to maximize efficiency and reliability, drastically reduce copper cable usage, and support more compact system designs. Third-generation semiconductor SiC/GaN is the key to achieving this transformation, and several semiconductor vendors have announced their participation in NVIDIA's 800V HVDC program.

SiC is mainly used in the front-end and mid-end links of data center power supply architectures, and is responsible for processing the highest voltage and maximum power conversion operations. Although SiC power semiconductors currently lag behind traditional Si in terms of maximum voltage ratings, they have excellent thermal performance and switching characteristics, which are essential for next-generation solid state transformer (SST) technology.

With its high frequency and high efficiency advantages, GaN plays an important role in the middle and end of the power supply chain, and pursues the ultimate power density and dynamic response. It is estimated that the penetration rate of third-generation semiconductor SiC/GaN in data center power supply will rise to 17% in 2026, and is expected to exceed 30% by 2030.

PART 6:2nm GAAFET innovation, 2.5D/3D packaging breakthrough

As 2nm enters mass production, a trend of pursuing higher transistor density inward and larger package sizes outward has formed in the commercial competition for advanced manufacturing processes. At the same time, it emphasizes heterogeneous integration (heterogeneous integration) capabilities. Through the combination of multi-chip stacks with different functions and different technology nodes, it meets the needs of high-efficiency computing and artificial intelligence applications.

In pursuit of higher transistor density, semiconductor wafer manufacturing officially switched from FinFET to GaAFET, completely covering the silicon channel through Gate-Oxide to achieve more efficient current control while chasing high-intensity computing power. For external components, 2.5D and 3D packaging technology provide high-density packaging solutions with multiple chip stacks, making inter-chip interconnection faster and lower power consumption, bringing breakthroughs to next-generation data centers and high-performance computing.

As various 2nm GAAFETs entered mass production, TSMC (TSMC), Intel (Intel), and Samsung (Samsung) launched 2.5D/3D packaging technologies such as CoOS/SoIC, EMIB/FOVEROS, and I-Cube/X-Cube, respectively, to provide integrated foundry services for front and rear segments. How to achieve a balance and commercial advantage between capacity utilization, reliability, cost and yield will be a core challenge for major foundry and packaging plants.

PART 7: Humanoid robot shipments grew by more than 700% in 2026, focusing on AI adaptation and scenario applicability

2026 will be a key year for commercialization of humanoid robots. Global shipments are estimated to increase by more than seven times per year and exceed 50,000 units. Market momentum is focused on two main axes: AI adaptivity (AI adaptivity) technology and scenario application orientation. AI adaptive technology combines efficient AI chips, sensing fusion, and the evolution of large-scale language models (LLM) to enable robots to learn and make dynamic decisions in real time in an unstructured environment, and demonstrate the ability to “plan and then act”.

In this context, new humanoid robot products in 2026 will no longer use specifications or flexibility as the only selling point, but will lock in specific scenario values from the design stage, and can support complete tasks in the field, from manufacturing and handling, warehousing and sorting to inspection assistance, etc. that are expected to enter as early as possible. In 2026, humanoid robots will officially enter a new stage in the AI-driven, application-centered industry.

PART 8: Advanced laptop displays are being accelerated, and the folding machine mainstream process welcomes key points

OLED displays have ushered in a cross-generational turning point. High-generation (8.6th generation) AMOLED production lines of Chinese and Korean panel manufacturers continue to expand production. As the cost structure and yield continue to improve, OLED display technology is accelerating the coverage of full-size products from small to large, simultaneously driving the average unit price (ASP) and supplier bargaining power of high-end components in related supply chains such as driver ICs, TCON, touch modules, and cooling design.

OLEDs break through the physical bottlenecks of LCD in thickness and energy consumption with characteristics such as self-luminescence, high contrast, and variable refresh rate, and meet Apple (Apple)'s dual requirements for image accuracy and energy efficiency. Apple expects to officially introduce OLED panels to MacBook Pro in 2026, which will drive high-end laptop display specifications to OLED. It is estimated that the penetration rate of OLED laptops is expected to reach 5% in 2025. After 2026, driven by Apple, it is expected to increase to 9-12% from 2027 to 2028.

Furthermore, as Apple has the opportunity to officially enter the folding phone market from the second half of 2026 to 2027, it will redefine the value of folding phones with software and hardware integration, brand trust, and supply chain synergy, driving the market focus from “dazzling appearance” to “deepening productivity and experience”. It is estimated that it will drive global folding phone shipments to exceed 30 million units in 2027. Currently, folding phones still face the last hurdle to mainstream — hinge reliability, flexible panel packaging, yield and cost control. Apple's prudence in product verification and quality reflects the importance it attaches to the timing of entry and user experience. It also highlights that folding phones still need time and strength to cross the gap before they truly reach maturity.

PART 9: Meta drives global near-vision display leap forward, LEDos technology accumulates growth energy

Along with the deepening of AI applications, Meta launched Meta Ray-Ban Display AR glasses with display functions to lock down “information provision” applications, bring AI closer to everyday life, reshape user usage behavior, and enhance the two-way interaction experience between AI and users through first-perspective data collection and feedback. The display technology uses LCoS, which has stable performance in full color and maturity, not only to secure technical development time for LEDos that are not yet fully mature, but also to accumulate market volume through good user experience.

As market expectations and Meta iterative product planning advance, the trend is towards LEDoS technology with higher brightness and contrast to expand application scenarios. Coupled with the continuous layout of manufacturers such as Apple, Google (Google), RayNeo (Thunderbird Innovation), INMO (Film Technology), Rokid (Rocky), and Vuzix, the cost is expected to accelerate to the sweet point expected by the public, which is beneficial to the development of LEDOS. It is estimated that more mature full-color LEDOS solutions will appear in 2027-2028, and Meta is also expected to launch a new generation of AR glasses equipped with LEDOS.

PART 10: The penetration rate of assisted driving will increase in 2026, and Robotaxi will start expanding in multiple regions around the world

It is estimated that the penetration rate of L2 (inclusive) and above assisted driving will exceed 40% in 2026, and intelligent electric vehicles will continue to be the driving force for the growth of the automobile industry. L2 driver assistance technology has matured, and the key transition cost is being popularized. The integrated chip and controller for cabin driving will enter large-scale mass production in 2026, initially focusing on the Chinese mid-tier automobile market. Traditional car manufacturers are also actively promoting the intelligent transformation of fuel vehicles, which is also the driving force for assisted driving to become a standard vehicle.

On the other hand, Robotaxi, which targets the L4 class, is ushering in a wave of global expansion. In addition to the loosening of local regulations, fleet platforms and service providers are becoming more active in adopting Robotaxi, and developers exploring more generalized AI models such as End-to-End (E2E) and VLA (Vision Language Action), all of which have helped the Robotaxi market expand. It is expected that by 2026, Robotaxi will accelerate its coverage of markets such as Europe, the Middle East, Japan, and Australia, and will no longer be limited to China and the United States.