With the market size exceeding 120 billion U.S. dollars, the rankings of Chinese server vendors are reshuffled.

The server industry is growing increasingly anxious.

This statement may sound counterintuitive. After all, AI is still driving the most expensive round of infrastructure construction in the history of human technology. According to market forecasts, the total capital expenditure related to AI of the world's four major cloud vendors in 2026 will reach 725 billion US dollars. Calculated at 725 billion US dollars, that is equivalent to an investment of nearly 2 billion US dollars per day, a large part of which will eventually flow to AI servers and their supporting infrastructure.

The server market is also expanding rapidly. In the first quarter of 2026, the revenue of global server vendors reached 122.6 billion US dollars, a year-on-year increase of 30.4%. The problem is that the more money flows to AI servers, it does not mean that every sum of money will go through traditional server vendors. The growth of the market and the position of server vendors in the industrial chain are becoming two separate matters.

01

The standard answer for a single card no longer exists

The anxiety of server vendors this year is fully reflected at WAIC 2026.

Looking upstream, chip vendors are no longer satisfied with selling GPUs or NPUs. Huawei, MetaX, and Moore Threads have launched interconnection, communication libraries, reference architectures and full racks to customers, hoping to personally figure out how a single chip can form a thousand-card cluster.

Looking downstream, cloud vendors and leading internet companies master models and workloads, can define servers by themselves, and then place orders directly to ODMs. In the first quarter of 2026, ODM Direct revenue still accounted for 50.2% of the global server market. In other words, about half of the global server revenue can now be completed directly between large customers and manufacturing plants.

With the development of AI, the boundaries between chip vendors, ODM vendors, OEM vendors, and cloud vendors have become increasingly blurred. Under the rapid market expansion, server vendors must re-answer a question: what exactly makes them irreplaceable in the AI industrial chain?

We see that the first step that server vendors are taking is to move towards multi-card heterogeneous computing. Especially for Chinese server vendors, this step has become extremely necessary. After all, the particularity of the Chinese market has led to the emergence of more and more AI acceleration cards in the market.

An industry survey of 60 Chinese corporate executives conducted in July this year shows that respondents expect to allocate 46% of their AI accelerator budget to domestic chips in the next 12 months, higher than the current 30%. In addition to NVIDIA, customers can also choose from vendors such as Huawei Ascend, Cambricon, Kunlunxin, Moore Threads, MetaX, Enflame, and Iluvatar. They enter the market through different paths such as NPUs, general-purpose GPUs, and dedicated AI processors, with different applicable models and scenarios.

With more choices, the business seems to be broader: the same server product line can be equipped with different chips to cover more customers. However, in actual practice, it is not as simple as replacing a hard drive. Replacing an acceleration card may require redoing the motherboard layout, power supply, heat dissipation, firmware and signal integrity; after the machine is powered on, it is also necessary to handle drivers, compilers, operator libraries and communication libraries. When the same model is migrated to another card, it often only "can run", far from reaching reasonable performance.

Server vendors start with hardware. In March this year, Inspur Information disclosed that its diverse and open computing power architecture, by opening up accelerator modules and computing power modules, allows the same system platform to support 90% of the AI chips in the industry, as well as CPU architectures such as x86 and ARM, and shortens the R&D cycle from chips to computing power systems to 6 to 8 months.

The next step is more difficult software adaptation. Super Fusion continues to promote B.E.S.T 3.0 and FusionOne AI to the software layer, putting operators, template libraries, compilers, model deployment and inference acceleration into one set of solutions, and proposes that heterogeneous computing power should be "cooperatively available, hybrid trainable, and hybrid inferable". H3C's Lingxi Intelligent Computing strives to uniformly access, reuse and schedule heterogeneous GPUs; Lenovo's Wanquan Heterogeneous Intelligent Computing Platform aims to put different chips, models and clusters ranging from 100 cards to 10,000 cards into the same management system.

A person from a domestic accelerated chip vendor told the author that internet customers will repeatedly conduct POC tests with their own models before procurement, and finally calculate the TCO. Hardware benchmarking is only one of the items. Whether the software is easy to use, how long it takes to complete model adaptation, and whether the chips can be supplied continuously will all affect the procurement result.

This gives server vendors new value space. The more domestic chips there are, the less likely customers are to fully master each set of software and hardware system one by one. If a server vendor can only list a "compatibility list", it is still easy to be replaced; if it can reduce the chip replacement time, performance loss and operation and maintenance costs at the same time, what it sells is no longer just a box for installing cards.

02

Servers are being swallowed up by racks

Looking at WAIC 2026 as a whole, another obvious trend is that super nodes have become the product unit in the AI era. The "White Paper on Super Node Technology System" released by the DeepLink team of Shanghai AI Laboratory gave a set of comparisons: in the past 5 years, the computing power required for pre-training has increased by about 3000 times, while the computing power of a single chip has only increased by about 16 times in the same period.

Super nodes first use Scale-Up to form more acceleration cards into a low-latency computing domain, and then access larger clusters through Scale-Out. What users buy is ostensibly still a rack, but when in use, they need to treat it as a "large machine" spanning multiple racks.

This also brings a problem: the number of cards does not equal effective computing power. The above white paper calculates in a set of reference designs that when the Scale-Up domain is expanded from 8 cards to 32 cards, the training benefit of the hundred-billion-parameter model is 20%, and the inference benefit is 42%; after expanding to 512 cards, the two benefits only increase to 21% and 45%.

The competition of AI servers this year is more like a "card counting competition": 64 cards, 128 cards, 256 cards, 1024 cards, and clusters expanding to 10,000 cards and 100,000 cards. But "super node is an extreme engineering problem, not the more cards the better, the one that fits the application scenario is the best." said a person from a domestic GPU vendor in an interview. Super nodes expand the scope of low-latency communication, and also expand the impact of failures: if one card drops out, one network congestion occurs, or one liquid cooling connector fails, what may be slowed down is no longer just one server.

At this time, the advantages of different companies are placed on the same table.

The first category is full-stack vendors that master chips, interconnection and software. Huawei displayed the Atlas 950 SuperPoD, a 1024-card cluster, at WAIC 2026. According to the data disclosed by Huawei, it has 256TB of global unified memory with a round-trip delay of 3 microseconds; the previous Ascend 384 super node has been deployed in more than 750 sets. Huawei can design NPUs, Lingqu interconnection, complete machines and CANN software stacks at the same time, and use vertical integration to improve system efficiency. Alibaba follows a similar path with self-developed chips, ALink interconnection and Panjiu servers, and can use real cloud workloads to test the system.

The second category is GPU vendors that extend from chips to systems. MetaX Xijing S600 fully interconnects 64 GPUs in a single rack; Moore Threads MTT C256 connects 256 GPUs with two standard racks, and claims to reduce the delay between cards to the sub-microsecond level. They build super nodes to prove that their own GPUs can not only run on a single card, but also form a commercial system. The advantage is that they know their own chips and communication stacks better, while the short board is that commercial delivery also involves networks, storage, power supply, liquid cooling, fault management and service systems.

The third category is traditional server and ICT vendors, for this type of vendors, the more important thing is to emphasize compatibility and delivery. We see that ZTE's OEX super node focuses on openness and fast chip replacement. It orthogonally and directly connects the computing tray and the switching tray, cancels the internal high-speed cables, and opens up the mechanical and electrical interfaces. According to the data disclosed by ZTE, this design can shorten fault maintenance from hour level to minute level, and compress the complete machine development cycle to 3 to 6 months; by replacing key modules, it can adapt to different GPU platforms. ZTE has also jointly launched the Matrix solution with vendors such as Biren, MetaX, Enflame and Iluvatar.

H3C's UniPoD S80000 series can deploy 32 to 64 acceleration cards in a single rack, can be scaled up to 1024 cards via Scale-Up, and extended to 16384 cards via Scale-Out. At the same time, H3C has also launched 102.4T intelligent computing switches, all-flash storage, liquid cooling, power supply and operation and maintenance systems, demonstrating the full-stack collaborative capability of "computing-network-storage-cloud-security-operation".

YuanTu has made forward-looking layouts of high-density computing, high-speed interconnection, high-efficiency power supply and liquid cooling heat dissipation technologies, and provides multi-level super node complete machine solutions ranging from 32 cards, 64 cards to 128 cards. While most domestic computing power vendors only focus on the single server track, YuanTu chooses to independently develop and deliver in batches on both the "server + high-speed switch" tracks, connecting "computing, network, storage and management", and turning the joint development, supply chain and manufacturing capabilities accumulated from long-term services for leading customers into standardized self-owned products.

03

Hardware is sold only once, Tokens are produced every day

The competition of AI servers now also needs to focus on Tokens. As of March this year, the daily average Token call volume in China has exceeded 140 trillion, a 1400-fold increase compared with 100 billion at the beginning of 2024. After large models move from training to application, customers no longer only look at the number of GPUs and peak computing power, but also ask how many Tokens the same machine and power can generate in one day, how long users need to wait, and how much it costs per million Tokens.

Since the beginning of this year, "Token Factory" has emerged intensively. Different AI server vendors have different views on Token production.

H3C has built the Turing Token Factory, an integrated operation and management platform oriented to computing power, models, tasks and Tokens, which takes over model tasks and application requests from the top, and uniformly schedules GPU resources and computing power allocation from the bottom. Earlier, Yu Yingtao, President and CEO of H3C Group, disclosed that "Turing Town" is not just a computing power park in the traditional sense. It is more like a "Token Factory" centered on Token production, scheduling and usage.

Super Fusion has put the Token Factory into actual operation: it has a daily demand of more than 50 billion Tokens, more than 30 agent scenarios, and has achieved 100% R&D coverage for all employees through AI-enabled software engineering in two years. In the latest Agentic practice, one team produced 500,000 lines of code in parallel by multiple agents within 3 months.

04

Conclusion

The global server market is still expanding. The Allen Chang team of Goldman Sachs published a research report, raising the total scale of the global server market, and predicting that the global server market revenue will reach 1.1 trillion US dollars in 2028. Among them, the revenue of AI server racks will expand at a compound annual growth rate of 118% to 561.4 billion US dollars, accounting for 51% of the total revenue of the global server market.

AI servers have not become a more standardized business. Chip paths, interconnection protocols, liquid cooling methods and model workloads are all differentiating, and instead, it has changed from a standard product to a complex engineering project. This will become the next threshold for AI server vendors.

This article is from WeChat Official Account "Semiconductor Industry Panorama" (ID: ICViews), author: Jiu Lin, published with authorization from 36Kr.