AI server machine chain reshuffles, shipments from 3 manufacturers squeeze into the top five in the world

According to the latest research report from DIGITIMES, shipments of high-end AI servers using high-bandwidth memory are expected to reach 2.37 million units in 2026, an annual increase of 56%. (File photo)

[Reporter Fang Weijie/Taipei Report] Global AI server demand continues to rise, and the shipment ranking of the entire machine chain has changed. According to the latest research report of DIGITIMES, the shipment volume of high-end AI servers using high-bandwidth memory (HBM) is estimated to reach 237 in 2026. million units, an annual increase of 56%. Among them, Quanta (2382) and Hon Hai (2317) ranked second and third among global suppliers. Wiwynn (6669) benefited from increased orders from cloud customers, and its shipment ranking jumped to fifth in the world, surpassing Supermicro.

DIGITIMES pointed out that as large-scale language model capabilities continue to improve, and agent-based artificial intelligence (Agent AI) applications drive demand for business automation, large-scale cloud service providers (CSPs), emerging cloud operators (Neo Cloud) and leading AI laboratories in North America are accelerating the construction of data centers. It is estimated that global AI server shipments will exceed 2.5 million units in 2026, with high-end AI servers accounting for 15% of total server shipments for the first time.

Observed from the supplier rankings, Google and Celestica's cooperation system maintains the first place in the world, Quanta and Hon Hai occupy the second and third places respectively. The rankings of the top four suppliers are expected to remain unchanged. It is worth noting that Dell, ranked fourth, will ship nearly 300,000 high-end AI servers this year, an annual increase of more than 162%. Wiwynn’s ranking rose to fifth due to the growth of cloud customer orders, and Supermicro dropped to sixth.

DIGITIMES analyst Xiao Shenglun pointed out that as the procurement scale of Tier 2 cloud companies gradually expands and they prefer to purchase AI servers through branded companies, it is estimated that original equipment manufacturer (OEM) shipments will account for more than 20% of the total high-end AI servers in 2026, and the proportion of direct purchases from original design manufacturers (ODM-direct) will drop to 79%. The global AI server procurement model is undergoing structural changes.

In terms of product architecture, NVIDIA's GB300 NVL72 has shipped more smoothly than its predecessor, driving the proportion of high-end AI servers using GPUs to increase to nearly 60% in 2026. Among them, the shipment proportion of NVL72 in different AI cabinet types is expected to rise sharply from 24% in 2025 to 42%, surpassing the Fish rack used by Google TPU and becoming the largest model. Traditional architectures such as NVL8 and PCIe continue to shrink.

On the other hand, the procurement share of large North American cloud players is expected to drop from 78% to 68%, while Neo Cloud will increase to 17%, with leading AI laboratories contributing nearly 6% for the first time. Among them, Google and Amazon are still the top two customers. Oracle and SpaceXAI rose to third and fourth respectively, surpassing Microsoft. Anthropic entered the list for the first time and ranked eighth with 135,000 units. Nebius's shipment growth rate exceeded 4 times.

Looking forward to the market outlook, Xiao Shenglun believes that market concerns about the safety of AI models may cause some industries to re-examine their computing power deployment plans and slow down the pace of model training in the short term. However, relevant computing resources are expected to gradually shift to long-term inference and harness engineering needs driven by Agentic AI commercial applications. As the transformation of next-generation inference AI servers progresses, although short-term shipment growth may temporarily slow down, the overall high-end AI server market is still expected to return to a high-speed growth track.