Accelerated Server Sales Are Themselves Accelerating

retained text

Accelerated Server Sales Are Themselves Accelerating
Do you know what the difference is between an F-18A taking off from the deck of an aircraft carrier to attack long range bombers at 50,000 feet and the revenue trend in the server market?
Me neither.
Either way, you get a stiff neck craning to see the heights.
In the second quarter ended in June, IDC reckons that the world consumed $166.32 billion in servers, an increase of 52 percent year on year and up an amazing 35.6 percent sequentially. We have never seen anything like this before in six decade history of corporate computing. And here is chart showing quarterly sales of servers worldwide from the peak of the Dot Com boom until today to prove it.
We live in a world where half of the revenue being spent on systems are for what we would normally call supercomputers, and two-thirds of the other half is being spent to be the butler and secretary for those supercomputers, with the remaining one-third of that other half being spent for traditional data processing, data analytics, and web infrastructure workloads that use to represent the state of the art in corporate computing only a decade ago.
Of that $162.32 billion spent on server nodes and rackscale systems in Q2 2026, some 52.5 percent of these machines were for GPU-accelerated systems, which works out to $87.4 billion. IDC said in its public quarterly server market tracker report that the average selling price of a GPU-accelerated machine was $170,220, up 43.5 percent year on year from $118,600.
This number can be used to calculate a few insights.
First, we can calculate the number of units sold from this, and it turns out that unit sales of GPU systems were down because of the shift from server nodes that fit many to a rack and have distinct and separate memory domains to rackscale systems where the memory domain stretches across all of the GPUs and usually the CPUs in the rack. In Q2 the shipments of distinct GPU systems fell by 10.7 percent to 513,453 unit, down from 575,278 units in Q2 2025.
What this also tells you is that there are still many, many customers who are buying GPU-accelerated machines with only four or eight GPUs in a fat server node and that all of the buying, despite the talk, is not for rackscale systems with memory coherence across all machines in the rack. You can know this because a Grace-Blackwell NVL72 rackscale system probably costs somewhere around $4 million or so with Nvidia B200 GPUs, and the inference-tuned NVL72 using B300 GPUs costs $5 million or so, give or take a million dollars depending on who you ask.
That ASP for GPU systems given by IDC implies a very large number of low-end systems using cheaper PCI-Express cards, like Nvidia’s RTX Pro 6000. A four-way node using this class of GPU costs on the order of $90,000 to $125,000, depending on the CPU, storage, and network configurations. An eight-way version doubles that to somewhere between $180,000 to $250,000 – which is higher than the ASP. If you want to use B200 or B300 GPUs in an eight-way server node, you are talking maybe $400,000 to $500,000.
This is an inescapable conclusion, unless IDC is doing something really dumb like counting trays in a rackscale system as a node. I do not think this is the case.
Here is the other neat thing that IDC said is that sales of XPU systems – those using other kinds of matrix math engines to run AI workloads and only AI workloads, not general purpose stuff that GPUs can also do – rose by a factor of 3.4X to $27.5 billion in the second quarter. We do not have ASPs from IDC and therefore cannot compute shipments of these devices accurately, but if you assume XPUs cost less than half of a GPU, that puts XPU systems at an ASP of around $85,000 and shipments in Q2 2026 at around 325,000. The shipment numbers here will only grow overt time as all of the clouds and all of the hyperscalers and one of the AI model builders are all working on inference chips. . . . By the way, using these numbers and fiddling up some Q2 2025 XPU ASPs, I think XPU ASPs are up around 45 percent and shipments are up 2.4X. Some of that price increase is due to increasing component costs – as is the case with GPUs, too – and some of it is for more advanced functions on more advanced chip processes.
If you add it all up, you get an average GPU/XPU system costing around $140,000 and around 840,000 units shipped in the second quarter of this year, for a total of $137.35 billion. That gives accelerated machines a 16.1 percent share of units but at 69.1 percent share of revenues.
Plain vanilla servers that do not have acceleration augmentation excepting the vector and (rarely) matrix engines inside of their CPUs accounted for $51.42 billion in sales, IDC says, up 55.8 precent because of the general server refresh cycle and also due to the rise of AI sandbox systems that execute the code and processes created by GenAI inference models. There were just north of 4 million such servers sold in the second quarter, up 16.7 percent. The whole market represented 5.16 million machines, by our math and estimation.
The original design manufacturers are making money hand over first in this exploding market, hauling in $89.66 billion in the June quarter, up 35.2 percent as a group and representing 53.9 percent of all server sales worldwide.
Here is the breakdown by vendor:
Dell has clearly started benefitting from the GenAI boom, as we recently talked about tearing apart its most recent financial results. Dell has blown by Hewlett Packard Enterprise in servers, and has maintained its position as being twice as large as Supermicro. Lenovo had a very good quarter in terms of server revenues, doubling its business year on year. IEIT Systems, which is an arm of Chinese server maker Inspur, actually shrunk, which is odd.
I strongly suspect profits as a share of revenue are down for all of these server makers as they eat the GenAI bear meat with too little fat. The difference is that with servers, you can make it up in volume, and with lean bear meat, eating more and more to sate your hunger will kill you.
Nvidia can sure get rich selling everybody compute engines and networking for GenAI, though.

document fetched 2026-09-22 02:53 · canonical https://nextplatform.com/compute/2026/09/21/accelerated-server-sales-are-themselves-accelerating/5297993