The server industry is becoming increasingly anxious. This statement sounds somewhat 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 combined AI-related capital expenditures of the world's top four cloud providers are projected to reach $725 billion in 2026. Calculated at $725 billion, this equates to an investment of nearly $2 billion per day, a significant portion of which ultimately flows into AI servers and their supporting infrastructure.
The server market is also expanding rapidly. In the first quarter of 2026, vendor revenue in the global server market reached $122.6 billion, a year-on-year increase of 30.4%. The issue is that The fact that more capital flows into AI servers does not mean every dollar passes through traditional server vendors. Market growth and the position of server vendors in the industry chain are becoming two separate issues.
01. The Standard Answer for a Single Card Has Disappeared The anxiety of server vendors this year is evident at WAIC 2026.
Looking upstream, chip manufacturers are no longer satisfied with merely selling GPUs or NPUs. Huawei, MetaX, and Moore Threads have pushed interconnects, communication libraries, reference architectures, and entire racks to customers, hoping to personally answer how a single chip can form a thousand-card cluster.
Looking downstream, cloud providers and top internet companies master the models and workloads, allowing them to define servers themselves and place orders directly with 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 manufacturers.
Given the growth of AI, the boundaries among chip manufacturers, ODMs, OEMs, and cloud providers are becoming increasingly blurred. Amid rapid market expansion, server vendors must re-answer a question: where exactly is their irreplaceable value in the AI industry chain?
We have observed that the first step server vendors are taking is to move towards multi-card heterogeneous architectures. This step has become particularly necessary for Chinese server vendors. After all, the uniqueness of the Chinese market has led to an increasing number of AI accelerator cards appearing in the market.
An industry survey of 60 Chinese corporate executives in July this year showed that respondents expect to allocate 46% of their AI accelerator budgets to domestic chips over the next 12 months, up from the current 30%. Besides NVIDIA, customers can also choose among vendors such as Huawei Ascend, Cambricon, Kunlunxin, Moore Threads, MetaX, Enflame Technology, and Iluvatar CoreX. They are entering the market through different routes such as NPU, general-purpose GPU, and dedicated AI processors, each suited for different models and scenarios.
With more choices, the business appears to have broadened: the same server product line can be equipped with different chips to cover more customers. However, actually doing it is not as simple as swapping a hard drive. Changing an accelerator card may require redoing the motherboard layout, power supply, cooling, firmware, and signal integrity. After the machine is powered on, drivers, compilers, operator libraries, and communication libraries still need to be handled. Migrating the same model to another accelerator card often only means it “can run” — a far cry from delivering optimal performance.
Server vendors are starting with hardware. In March this year, Inspur disclosed that its diverse and open computing power architecture, through open accelerator modules and computing modules, enables the same system platform to support 90% of the industry's AI chips, as well as CPU architectures such as x86 and ARM, reducing the R&D cycle from chip to computing system to 6 to 8 months.
Next comes the more challenging software adaptation. xFusion continues to push B.E.S.T 3.0 and FusionOne AI towards the software layer, integrating operators, template libraries, compilers, model deployment, and inference acceleration into a single solution, and proposing that heterogeneous computing power should be "collaboratively usable, hybridly trainable, and hybridly inferable." H3C's Lingxi Intelligent Computing attempts to unify the access, reuse, and scheduling of heterogeneous GPUs; Lenovo's Wanquan Heterogeneous Intelligent Computing Platform aims to place different chips, models, and clusters ranging from hundreds to tens of thousands of cards into a single management system.
A representative from a domestic accelerator chip manufacturer told the author that internet customers will repeatedly run POCs with their own models before procurement, ultimately calculating the TCO. Hardware benchmarking is just one aspect; whether the software is user-friendly, how long it takes to adapt the model, and whether the chip can ensure continuous supply will all affect the procurement outcome.
This gives server vendors renewed room for value creation. The more domestic chips there are, the less likely customers are to fully master every software and hardware system one by one. If server vendors can only provide a "compatibility list," they remain easily replaceable; only if they can compress chip-swapping time, performance loss, and O&M costs simultaneously, are they selling more than just a box for mounting cards.
02. Servers Are Being Swallowed by a Single Rack Looking across WAIC 2026, another obvious trend is that the super node has become the product unit in the AI era. The "Super Node Technology System White Paper" released by the DeepLink team of the Shanghai AI Laboratory provided a set of comparisons: over the past five years, the computing power required for pre-training has grown by about 3,000 times, while single-chip computing power has only grown by about 16 times during the same period.
Super nodes first leverage Scale-Up to build a low-latency computing domain populated with more accelerator cards, then use Scale-Out to link into larger clusters. What users buy is superficially still a rack, but when using it, they must treat it as a "large machine" spanning multiple racks.
This also brings up a problem: the number of cards does not equal effective computing power. The aforementioned white paper calculated in a reference design that when the Scale-Up domain expands from 8 cards to 32 cards, the training gain for a 100-billion-parameter model stands at 20%, and the inference gain at 42%; after expanding to 512 cards, the two gains only increase to 21% and 45%.
The competition for AI servers this year is more like a "card-counting contest": 64 cards, 128 cards, 256 cards, 1,024 cards, and clusters expanding to tens of thousands or hundreds of thousands of cards. But "a super node is an extreme engineering problem; it is not the case that the more cards, the better. It is only good if it fits the application scenario," said a representative from a domestic GPU manufacturer in an interview. Super nodes expand the range of low-latency communication, but also expand the impact of failures: a single card going offline, a network congestion, or a problem with a liquid cooling connector could slow down more than just a single server.
At this point, the advantages of different companies are laid out on the same table.
The first category is full-stack vendors that master chips, interconnects, and software. Huawei showcased the Atlas 950 SuperPoD at WAIC 2026, which is a 1,024-card cluster. According to data disclosed by Huawei, it features 256TB of global unified memory with a round-trip latency of 3 microseconds; the previous Ascend 384 super node has been deployed in over 750 sets. Huawei can simultaneously design NPUs, Lingqu interconnects, entire machines, and the CANN software stack, trading vertical integration for system efficiency. Alibaba follows a similar route with its self-developed chips, ALink interconnects, and Panjiu servers, and can verify the system with real workloads on the cloud.
The second category is GPU manufacturers extending from chips to systems. MetaX's Xijing S600 fully interconnects 64 GPUs in a single rack; Moore Threads' MTT C256 connects 256 GPUs using two standard racks, claiming to reduce inter-card latency to the sub-microsecond level. They build super nodes to prove that their GPUs can not only run on a single card but also form commercial systems. The advantage is a deeper understanding of their own chips and communication stacks, while the shortcoming is that commercial delivery also involves networking, storage, power supply, liquid cooling, fault management, and service systems.
The third category is traditional server and ICT vendors. For these vendors, it is more important to emphasize compatibility and delivery. We have seen that ZTE's OEX super node focuses on openness and fast chip swapping. It orthogonally and directly connects computing trays with switching trays, eliminates internal high-speed cables, and opens up mechanical and electrical interfaces. According to data disclosed by ZTE, this design can reduce fault maintenance from hours to minutes and compress the server 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 CoreX.
H3C's UniPoD S80000 series can deploy 32 to 64 accelerator cards in a single rack, with Scale-Up expandable to 1,024 cards and Scale-Out extendable to 16,384 cards. At the same time, H3C also introduced 102.4T intelligent computing switches, all-flash storage, liquid cooling, power supply, and O&M systems, demonstrating the full-stack collaborative capability of "computing-network-storage-cloud-security-O&M".
Yuantu has forward-lookingly laid out capabilities in high-density computing, high-speed interconnects, high-efficiency power supply, and liquid cooling technologies, and provides multi-level super node server solutions ranging from 32, 64 to 128 cards. While most domestic computing power vendors only focus on the single track of servers, Yuantu has chosen the dual tracks of "servers + high-speed switches" for simultaneous independent R&D and batch delivery, integrating "computing, networking, storage, and management," and turning the joint development, supply chain, and manufacturing capabilities accumulated from long-term service to top customers into standardized proprietary products.
03. Hardware is Sold Only Once, While Tokens are Produced Every Day The competition for AI servers now also comes down to tokens. As of March this year, China's average daily token invocation volume has exceeded 140 trillion, a 1,400-fold increase from 100 billion at the beginning of 2024. After large models move from training to application, customers look not only at the number of GPUs and peak computing power but also ask how many tokens the same machines and power can generate in a day, how long users have to wait, and how much it costs per million tokens.
Since the beginning of this year, terms like "Token Factory" have emerged frequently. Different AI server vendors hold divergent views on token generation.
H3C has created the Turing Token Factory, an integrated operation and management platform for computing power, models, tasks, and tokens, which takes on model tasks and application requests upstream and uniformly schedules GPU resources and computing power allocation downstream. Previously, H3C revealed that "Turing Town" is not just a computing power park in the traditional sense. It is more like a "token factory" centered around token production, scheduling, and usage.
xFusion has already "practiced what it preaches" and truly put the Token Factory into operation: with a daily demand of over 50 billion tokens and 30+ agent scenarios, AI-empowered software engineering achieved 100% R&D coverage across all employees in two years. In the latest agentic practice, a single team of multiple agents working in parallel produced 500,000 lines of code in three months.
04. Conclusion The global server market is still expanding. A research report by Goldman Sachs' Allen Chang team raised the total size of the global server market, forecasting that global server market revenue will reach $1.1 trillion in 2028. Among them, AI server rack revenue will grow at a compound annual growth rate of 118% to hit $561.4 billion, accounting for 51% of the total global server market revenue.
AI servers have not turned into a more standardized business. Chip routes, interconnect protocols, liquid cooling methods, and model workloads are all diverging; instead, it has reverted from a standard product to a complex engineering challenge. This will become the next threshold for AI server vendors.