Accelerated Server Sales Continue to Accelerate

Server Market Revenue Reaches an Unprecedented Level

The global server market reached an extraordinary height in the second quarter ended June 2026. IDC estimates that organizations worldwide consumed $166.32 billion in servers during the quarter, an increase of 52 percent from the same quarter a year earlier and 35.6 percent sequentially. The result is unlike anything previously recorded in the six-decade history of corporate computing.

Approximately half of current systems revenue is now associated with machines that would traditionally be classified as supercomputers. Much of the remaining revenue supports those systems, while the rest serves conventional data processing, analytics, and web infrastructure workloads that represented the leading edge of corporate computing only a decade ago.

GPU-Accelerated Systems Drive Revenue

Of the $162.32 billion spent on server nodes and rack-scale systems in Q2 2026, 52.5 percent came from GPU-accelerated systems, equivalent to $87.4 billion. IDC reported that the average selling price of a GPU-accelerated machine was $170,220, up 43.5 percent year over year from $118,600.

These figures provide several indications about the market's composition.

The number of GPU system units shipped declined as purchasing shifted from conventional server nodes, which place multiple systems in a rack with separate memory domains, toward rack-scale systems. In rack-scale designs, the memory domain extends across the GPUs and typically the CPUs throughout the rack.

Shipments of distinct GPU systems fell 10.7 percent in Q2, to 513,453 units from 575,278 units in Q2 2025. The decline does not mean that all customers are moving to rack-scale systems. Many continue to buy GPU-accelerated machines with four or eight GPUs in a large server node.

A Grace-Blackwell NVL72 rack-scale system with Nvidia B200 GPUs probably costs about $4 million. An inference-optimized NVL72 using B300 GPUs costs approximately $5 million, with the final price varying by roughly $1 million depending on the configuration and supplier.

IDC's average selling price for GPU systems also indicates that a substantial number of lower-cost systems use PCI Express cards such as Nvidia's RTX Pro 6000. A four-GPU node in this category costs approximately $90,000 to $125,000, depending on its CPU, storage, and network configuration. An eight-GPU version costs approximately $180,000 to $250,000, which is above the reported average selling price. An eight-GPU server node using B200 or B300 GPUs could cost roughly $400,000 to $500,000.

This suggests that IDC is counting individual systems rather than rack-scale trays as nodes.

XPU Systems Expand Rapidly

IDC also reported rapid growth in XPU systems. These machines use other types of matrix-math engines for AI workloads, rather than GPUs that can support both AI and general-purpose workloads. XPU system sales increased by a factor of 3.4 to $27.5 billion in Q2.

IDC did not provide an average selling price for these systems, so their shipment volume cannot be calculated precisely. If XPUs cost less than half as much as GPUs, an estimated average selling price would be about $85,000, implying approximately 325,000 XPU systems shipped in Q2 2026.

Shipment volumes are expected to increase as cloud providers, hyperscalers, and at least one AI model developer work on inference chips. Based on estimated Q2 2025 prices, XPU average selling prices appear to have risen by approximately 45 percent, while shipments increased by about 2.4 times. Some of the price increase reflects higher component costs, as with GPUs, while some results from more advanced functions manufactured on newer chip processes.

Accelerated Systems Capture Most Revenue

Combining GPU and XPU systems produces an estimated average selling price of approximately $140,000 and about 840,000 units shipped during the quarter. Together, these systems generated an estimated $137.35 billion in revenue.

Accelerated machines therefore represented approximately 16.1 percent of unit shipments but 69.1 percent of server revenue.

Conventional servers without dedicated acceleration, apart from the vector and occasional matrix engines integrated into their CPUs, generated $51.42 billion in sales, according to IDC. Revenue increased 55.8 percent, supported by the broader server refresh cycle and the growth of AI sandbox systems that execute code and processes created by generative AI inference models.

More than 4 million conventional servers were sold in the quarter, an increase of 16.7 percent. The overall market represented approximately 5.16 million machines based on these figures and estimates.

Original Design Manufacturers Gain Scale

Original design manufacturers generated $89.66 billion in the June quarter, an increase of 35.2 percent as a group. They accounted for 53.9 percent of worldwide server sales.

Dell has begun to benefit substantially from demand associated with generative AI. The company moved ahead of Hewlett Packard Enterprise in server revenue and remained approximately twice the size of Supermicro. Lenovo also had a strong quarter, doubling its server business year over year.

IEIT Systems, the Chinese server manufacturer Inspur's affiliated company, recorded a decline in revenue during the quarter.

The rapid expansion of the generative AI market may reduce server manufacturers' profit margins as they compete for volume while absorbing higher component costs. Nvidia, however, continues to benefit from selling compute engines and networking equipment to organizations building generative AI infrastructure.