Accelerated Server Sales Are Themselves Accelerating

AI & Compute

IDC: GPU-Accelerated Servers Hit $87.4B in Q2 2026, Up 52% YoY

GPU-accelerated servers hit $87.4 billion in Q2 2026 at a $170,220 average price, IDC reports. XPU sales tripled to $27.5 billion. Accelerated machines hold 16% of units but 69% of revenue.

By
Sophie Lindqvist
Filed
Channel
AI & Compute
Read
3 min read

The global server market generated $166.32 billion in revenue during the quarter ended June 2026, according to IDC's quarterly tracker report. That figure rose 52 percent year on year. Sequential growth reached 35.6 percent. "We have never seen anything like this before in six decades of corporate computing," the analyst wrote.

How concentrated is the spending in accelerated systems?

GPU-accelerated machines accounted for $87.4 billion of the $162.32 billion spent on server nodes and rackscale systems. IDC put the average selling price of a GPU system at $170,220. That ASP climbed 43.5 percent from $118,600 a year earlier. Unit shipments fell 10.7 percent to 513,453 from 575,278 in Q2 2025.

The decline reflects a structural shift. Customers are moving from standalone server nodes to rackscale systems where memory spans every GPU and CPU in the rack. A single Nvidia Grace-Blackwell NVL72 rack using B200 GPUs costs roughly $4 million; the inference-tuned B300 variant approaches $5 million depending on configuration and supply chain variables.

What does the ASP reveal about system configurations?

The $170,220 average price implies heavy sales of lower-end machines. A four-way node built around Nvidia's RTX Pro 6000 lists between $90,000 and $125,000 depending on the CPU, storage and network configuration chosen. Eight-way configurations climb to $180,000 to $250,000. Eight-way nodes equipped with B200 or B300 GPUs command $400,000 to $500,000.

IDC is not double-counting rackscale trays as nodes. The reported composition stands as filed by the analyst.

How large is the XPU segment?

XPU systems, defined as machines running non-GPU matrix engines dedicated to AI workloads, posted sales of $27.5 billion in Q2. That figure represents a 3.4x increase year on year. IDC has not published average selling prices for XPU machines. Assuming unit ASPs near $85,000, shipments land around 325,000 units.

XPU average prices may have risen 45 percent year on year, the source estimates. Shipments likely grew 2.4x. The shipment numbers here will continue expanding as custom inference designs from hyperscalers and AI model builders reach volume production in 2026 and 2027. Component cost increases explain part of the price rise.

What share do accelerated machines hold?

Combined GPU and XPU revenue reached $137.35 billion across roughly 840,000 units. That gives accelerated hardware a 16.1 percent share of units and a 69.1 percent share of server revenue. Plain CPU servers recorded $51.42 billion, up 55.8 percent on refresh cycles and growing GenAI sandbox demand from enterprises. Unit volume exceeded 4 million, up 16.7 percent. Totals for the market reach 5.16 million machines.

Who is selling the most?

Original design manufacturers collectively booked $89.66 billion, up 35.2 percent. ODMs represent 53.9 percent of worldwide server sales. Among branded vendors, Dell has overtaken Hewlett Packard Enterprise. Dell now runs at roughly twice Supermicro's server revenue. Lenovo doubled its server business year on year. IEIT Systems, the Inspur-linked Chinese supplier, contracted against the trend, which the source flags as odd.

Profit margins across branded server makers are likely compressing. Vendors "eat the GenAI bear meat with too little fat," the analyst wrote. Volume offsets thin margins today. Sustained margin pressure on system builders will not be absorbed indefinitely through volume gains as component costs remain elevated across the accelerator supply chain. Nvidia continues to capture revenue above the system builders. "Nvidia can sure get rich selling everybody compute engines and networking for GenAI," the analyst noted.

What happens next?

GPU unit shipments declined while revenue rose sharply, indicating rackscale architectures will absorb more incremental dollars. XPU penetration will rise as custom inference silicon enters volume production. Profit recovery at system builders depends on component cost normalization, a condition the source does not confirm. Pricing power among accelerator suppliers remains the swing factor for the next two quarters.

Source: The Next Platform

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

347 articles

Related articles

« Previous articleNext article »