The server industry is increasingly anxious.

This statement may sound counterintuitive. After all, AI is still driving the most expensive round of infrastructure construction in human technological history. Market projections indicate that by 2026, the combined AI-related capital expenditures of the world's four largest cloud vendors will reach US$725 billion. At US$725 billion, this equates to nearly US$2 billion invested daily, with a significant portion ultimately flowing into AI servers and their supporting infrastructure.

The server market is also rapidly expanding. In the first quarter of 2026, global server market vendor revenue reached US$122.6 billion, up 30.4% year-on-year. The issue is that more money flowing into AI servers does not mean every dollar will pass through traditional server vendors. Market growth and server vendors' positions in the supply chain are becoming two separate matters.

This year's anxiety among server vendors was evident at WAIC 2026.

Looking upward, chip vendors are no longer satisfied with just selling GPUs or NPUs. Huawei, MetaX, and Moore Threads are now offering interconnects, communication libraries, reference architectures, and entire server racks to customers, aiming to personally address how a single chip can form a thousand-card cluster.

Looking downward, cloud vendors and leading internet companies control models and workloads, enabling 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 global server revenue now bypasses traditional channels, flowing directly between major customers and manufacturers.

Under AI development, the boundaries between chip vendors, ODMs, OEMs, and cloud vendors are blurring. Amid rapid market expansion, server vendors must reanswer a critical question: What makes them truly irreplaceable in the AI supply chain?

The first step server vendors are taking is toward multi-card heterogeneity. This step is especially necessary for Chinese server vendors, given the unique characteristics of the Chinese market, which has seen a growing number of AI accelerator cards.

An industry study in July 2026 surveyed 60 Chinese corporate executives, revealing that respondents expected to allocate 46% of their AI accelerator budgets to domestic chips over the next 12 months, up from 30% currently. Besides NVIDIA, customers can choose from vendors like Huawei Ascend, Cambrian, Kunlunxin, Moore Threads, MetaX Integration, Enflame Technology, and Iluvatar CoreX. These vendors approach the market through different routes, including NPUs, general-purpose GPUs, and dedicated AI processors, each suitable for different models and scenarios.

More choices seem to broaden business opportunities: a single server product line can now accommodate different chips, covering more customers. However, implementation is far from simple as swapping out a hard drive. Replacing an accelerator card may require redesigning the motherboard layout, power supply, cooling, firmware, and signal integrity. After the machine powers on, drivers, compilers, operator libraries, and communication libraries must also be addressed. Migrating the same model to another card often only means it "runs," falling short of optimal performance.

Server vendors are starting with hardware. In March 2026, Inspur Information disclosed that its diversified open computing architecture supports 90% of AI chips in the industry, including x86 and ARM CPU architectures, through open accelerator modules and computing modules. It also shortens the research and development cycle from chip to computing system to 6-8 months.

Next comes the more challenging software adaptation. Supercomputing Converged Systems continues to push into the software layer with B.E.S.T 3.0 and FusionOne AI, integrating operators, template libraries, compilers, model deployment, and inference acceleration into a single solution. It proposes that heterogeneous computing power should be "collaboratively usable, hybrid trainable, and hybrid inferable." H3C's Lingxi Intelligent Computing attempts to unify access, reuse, and scheduling of heterogeneous GPUs. Lenovo's Wanquan Heterogeneous Intelligent Computing Platform aims to manage different chips, models, and clusters ranging from hundreds to tens of thousands of cards within a single system.

A representative from a domestic accelerator chip vendor told the author that internet customers conduct thorough POC testing with their models before purchasing, ultimately calculating total cost of ownership (TCO). Hardware benchmarks are just one factor; software usability, model adaptation time, and continuous chip supply all influence procurement decisions.

This creates renewed value for server vendors. The more domestic chips there are, the less likely customers are to master every hardware and software system individually. Server vendors that merely provide a "compatibility list" remain easily replaceable. Only by reducing chip swap time, performance loss, and operational costs can they offer more than just a box for holding cards.

Another clear trend at WAIC 2026 is that super nodes are becoming the product unit of the AI era. The White Paper on Super Node Technology Systems released by the DeepLink team at the Shanghai Artificial Intelligence Laboratory provided a comparison: Over the past five years, the computing power required for pre-training has grown about 3,000-fold, while single-chip computing power has only increased about 16-fold during the same period.

Super nodes first use Scale-Up to combine more accelerator cards into low-latency computing domains, then connect to larger clusters via Scale-Out. What users purchase may appear to be server racks, but they must be treated as a single "large machine" spanning multiple racks during use.

This raises a question: The number of cards does not equal effective computing power. The white paper calculated in a reference design that expanding the Scale-Up domain from 8 to 32 cards yields a 20% training benefit and a 42% inference benefit for models with hundreds of billions of parameters. Expanding to 512 cards only increases these benefits to 21% and 45%, respectively.

This year's AI server competition resembles a "card-counting race": 64-card, 128-card, 256-card, 1,024-card clusters, and expansions toward 10,000-card and 100,000-card setups. However, "super nodes represent an extreme engineering challenge; more cards are not necessarily better—only those suitable for application scenarios matter," said a representative from a domestic GPU vendor in an interview. While super nodes expand low-latency communication ranges, they also amplify fault impacts: a single card going offline, network congestion, or a liquid cooling joint issue may no longer slow down just one server.

At this point, the strengths of different companies are placed on the same table.

The first category consists of full-stack vendors that control chips, interconnects, and software. Huawei's Atlas 950 SuperPoD, showcased at WAIC 2026, is a 1,024-card cluster. According to Huawei's disclosures, it features 256TB of global unified memory and a round-trip latency of 3 microseconds; its previous Ascend 384 super node has been deployed in over 750 instances. Huawei can simultaneously design NPUs, Lingqu interconnects, entire servers, and the CANN software stack, trading vertical integration for system efficiency. Alibaba's approach, using self-developed chips, ALink interconnects, and Panjiu servers, follows a similar path and validates the system with real cloud workloads.

The second category includes GPU vendors extending from chips to systems. MetaX Xijing S600 fully interconnects 64 GPUs in a single cabinet; Moore Threads' MTT C256 connects 256 GPUs across two standard cabinets, claiming to reduce inter-card latency to sub-microsecond levels. By building super nodes, these vendors aim to prove their GPUs can function not just as individual cards but also as commercial systems. Their advantage lies in deeper understanding of their chips and communication stacks, while their weakness is that commercial delivery also involves networking, storage, power supply, liquid cooling, fault management, and service systems.

The third category comprises traditional server and ICT vendors, for whom compatibility and delivery are more critical. We see that ZTE's OEX super node emphasizes openness and rapid chip swapping. It directly connects computing trays with switching trays orthogonally, eliminating internal high-speed cables and opening mechanical and electrical interfaces. According to ZTE's disclosures, this design reduces fault maintenance from hours to minutes and shortens entire machine development cycles to 3-6 months; by replacing key modules, it can adapt to different GPU platforms. ZTE also launched the Matrix solution in collaboration with vendors like Biren, MetaX, Enflame, and Iluvatar CoreX.

H3C's UniPoD S80000 series can deploy 32 to 64 accelerator cards in a single cabinet, with Scale-Up expandable to 1,024 cards and Scale-Out extendable to 16,384 cards. Meanwhile, H3C has introduced 102.4T intelligent computing switches, all-flash storage, liquid cooling, power supply, and operational systems, demonstrating full-stack collaboration (collaborative) capabilities across "computing, networking, storage, cloud, security, and maintenance."

Yuantu has proactively developed technical capabilities in high-density computing, high-speed interconnects, efficient power supply, and liquid cooling. It offers multi-tier super node solutions ranging from 32-card to 128-card configurations. While most domestic computing power vendors focus solely on servers, Yuantu has chosen to independently research, develop, and mass-deliver both "servers + high-speed switches" simultaneously, integrating "computing, networking, storage, and management." It transforms the joint development, supply chain, and manufacturing capabilities accumulated from long-term service to top clients into standardized proprietary products.

The competition for AI servers now also hinges on tokens. As of March this year, China's daily token usage has surpassed 140 trillion, a 1,400-fold increase from 100 billion in early 2024. As large models shift from training to application, customers no longer focus solely on GPU counts and peak computing power; they also ask how many tokens a machine can generate per day with the same power, how long users must wait, and the cost per million tokens.

Since this year, "Token Factories" and "word-element factories" have emerged dense (frequently). Different AI server vendors hold varying views on token production.

H3C has built the Turing Token Factory, an integrated operations management platform for computing power, models, tasks, and tokens. It undertake (connects) model tasks and application requests upward while uniformly scheduling GPU resources and computing power allocation downward. Previously, Yu Yingtao, President and CEO of H3C Group, revealed that "Turing Town" is not just a traditional computing power park but more akin to a "Token Factory" centered around token production, scheduling, and usage.

Supercomputing Converged Systems has already "put the Token Factory into practice," handling over 50 billion tokens daily across 30+ intelligent agent scenarios. AI-empowered software engineering has achieved 100% research and development coverage within two years. In its latest Agentic practices, a three-month effort by a small team produced 500,000 lines of code in parallel using multiple agents.

The global server market continues to expand. Goldman Sachs' Allen Chang team published a research report raising the total global server market size projection, expecting revenue to reach US$1.1 trillion by 2028. AI server rack revenue will expand at a 118% CAGR to US$561.4 billion, accounting for 51% of total global server market revenue.

AI servers have not become a more standardized business. Diverging chip routes, interconnect protocols, liquid cooling methods, and model workloads are transforming them from standard products back into complex engineering projects. This will become the next threshold for AI server vendors.