How Sustainable Is This Crazy Server Spending?

We can all talk until we are blue in the face about how weird it is for so much money to be spent on servers during the GenAI boom, but after reviewing the latest market report from IDC – which is one again but sporadically giving out some stats to the public – we thought that to feel the full impact of this change, we should draw you a picture of the past 26 years of server revenues by quarter so you can take it all in.

So, here it is:

The blue line in the chart above is the nine quarters between Q3 2021 and Q4 2023 when IDC decided to stop giving out server data on a quarterly basis as it has been doing since 1996. It is actually four quarters more of not giving any data to the public because the Q4 2023 data is inferred from growth rates when data was put out in Q4 2024. We think withholding this information was foolish, obviously, and for clearly self-serving reasons because we love to interpret the data put out by IDC, Gartner, and others who try to count boxes and licenses and money and such.

There has been much talk of comparing server spending in the Dot Com Boom with the GenAI Boom, and as this chart makes obvious, they are not even close to comparable in magnitude even if they are somewhat alike in kind. (These numbers are not inflation adjusted, which you should do over long time horizons and which would tend to magnify the Y axis on the left side of the chart and shrink it on the right side of the chart, with the fulcrum point at the year you pick to normalize against. We just ran out of time.)

You cannot see the long buildup to even reach $12 billion or $13 billion a quarter in server sales per quarter in 1999, and that was when RISC/Unix systems were the hot commodity, X86 servers were on the rise, and IBM and a few others were still making very profitable billions of dollars selling proprietary systems. This was very much an enterprise systems market with the hyperscalers on the rise during the Dot Com Boom.

The collapse in server spending was absolutely awful, and you can see the sawtoothing in the market on a slight upwards trend as proprietary systems had their predictable Q4 spending spikes. But server spending did not get back to the levels of the peak of the Dot Com Boom for many, many years, as shown by the green horizontal line.

When the Great Recession hit in 2008 and was tough through 2009 and into 2010, server sales dropped down to the level of the pit of the Dot Com Bust, and it was hyperscaler and now cloud builder spending that started to drive the cycle more and more. The deep trough in 2012 was caused by companies awaiting the delayed “Sandy Bridge” Xeon E5 processors as well as the launch of Windows Server 2012, plus the collapse of the RISC/Unix except as a database server among enterprises.

There was another transitional trough in server revenues in 2017 as AMD re-entered the server market and companies were also awaiting the launch of the “Skylake” Xeon SPs by Intel. In the following year, the hyperscalers and cloud builders bought tons of new gear, finally driving revenues well above the Dot Com Boom peaks, but then the first wave of the first Trump Administration’s trade war tariffs started, causing a bit of tepidness in spending and a pulling back in IT investment, and in 2019 there was a server spending drought as the hyperscalers and clouds realized they went a little crazy buying capacity in 2018 and pulled back. The coronavirus pandemic causes that jagged spending pattern that cycles around $20 billion in quarterly server spending, and then the IDC data is missing where the blue line is and where the GenAI Boom is building momentum.

And look where we are now. Quarterly server spending, thanks to the huge number of GPU-accelerated and XPU-accelerated systems, is an order of magnitude larger than it was back in 1999. And this enormous change happened in a relatively short period of time. The question everyone wants to know is whether or not this new normal level of $100 billion or more in system sales per quarter is sustainable, or will be subject to the laws of gravity and come right back down again the second that GenAI return on investment doesn’t pan out.

We don’t think anyone truly knows, but we see a lot of infrastructure players doing well, but have yet to see a software powerhouse in GenAI have the kind of revenues that will sustain this server capacity consumption. There are a lot of companies hoping, and a lot of them building, but there seems to be a lot more hope and building then collecting revenues that are larger than the investments.

But, as you know full well by now, we can’t just leave that question hanging in the air on a skyhook. So we grabbed the annual forecast that IDC put out for server spending earlier this summer – which had data points for 2024 and 2029 and which we filled the gaps in with our best estimates between those data points – and then we converted it into a quarterly estimate by using the wiggly pattern that you see at the tail end of the real IDC server data from 2020 and 2021 with its dominant Q2 and Q4 and relatively weaker Q1 and Q3 cycles. We don’t think the dive in Q4 2025 will be as bad as it looks, but it has been a killer year so far as companies may be now awaiting the next Vera-Rubin products from Nvidia and Venice-Altair products from AMD, so maybe it will slow a bit now and then zigzag its way up to the heavens.

If this indeed something like what the shape of the future looks like – provided enough HBM memory can be made for accelerators, which is not a sure bet, or that models continue to need an increasing amount of compute, which is also not a sure bet – then the server market three decades after the Dot Com Boom was just building up steam will be an order of magnitude larger. Those red dashed lines are our estimates, informed by the IDC endpoints from 2024 and 2029 in its forecast.

Those same companies that busted your hump when IT spending was 1 percent of revenues are going to apparently be fine if it is 10X that amount. (Again, inflation is about 2X of that over 30 years.)

If you assume there is a basic undercurrent of about $55 billion a year in general purpose datacenter compute – meaning application, database, and web serving infrastructure and non-AI data analytics – then between 2014 and 2029, there has been/will be somewhere around $3 trillion in overall server spending, with AI-related server spending accounting for $2.18 billion of that and that general server spending being the remaining $825 billion.

We are not saying we believe this, but rather just that this is a logical way to look at the IDC data that we have seen. As we have been saying all along, we have trouble believing that the chip makers and packagers of the world can create enough chippery to chase such revenues, and that the world will finance such large sums without proof of the payback that is commensurate with such investments.

We will also remind you that the forecasts we talked about back in July from IDC and AMD had AMD being even more aggressive about the AI forecast.

With that all in your mind, let’s drill down into the third quarter of 2025, the most recent data that IDC has put out for the server market:

Sales of X86 servers are still important, with $76.3 billion in machinery sold in the third quarter, up 32.8 percent as companies upgrade their aging server fleets and many still use X86 hosts in their HPC and AI clusters. But that non-X86 part of the server market grew at 192.7 percent to $36.2 billion in the quarter, which just goes to show you how the hyperscalers and clouds are deploying their own Arm servers and Nvidia is finding great success with its “Grace” Arm server CPU in its AI and HPC infrastructure. IBM is also seeing an uptick in with new Power11 and System z17 proprietary enterprise systems rolling out this year.

IDC said that sales of servers with embedded GPUs grew by 49.4 year in year and represented more than half of the server revenues in Q3 – probably somewhere just shy of $70 billion is our guess. IDC said further that in the first three quarters of 2025 there was a total of $314.2 billion in GPU-accelerated machinery sold. We presume that this does not include proprietary XPU systems acquired by the clouds and hyperscalers or sold to model builders, such as Google TPU or Amazon Trainium XPUs.

The other interesting thing to note is how the ODMs are now pushing almost 60 percent of worldwide server revenues, up from 45 percent last year. Dell might be the biggest OEM by far at this point thanks to its AI revenues, but we strongly suspect that a few ODMs are just as large if not larger in terms of server revenues worldwide.