The second quarter ended four weeks ago, but the market researchers and analysts of the world – present company included – are still dicing, slicing, and mulling the numbers from the first quarter and putting together forecasts for all of the different facets of the IT sector of the global economy.
We tore apart and embellished the AI accelerator and CPU forecasts that AMD chief executive officer Lisa Su presented at the recent Advancing AI 2026 event in Silicon Valley, which you can see here, and poking around the IDC site, we found this gem giving its assessment of the current AI market and a forecast out to 2030. And with a little spreadsheet magic, that let me build a model that compares what AMD thinks the future likes with what IDC thinks might happen.
This latest report on AI accelerated server and storage spending from IDC was interesting for a couple of reasons. The first important thing is to get a new forecast as the price inflation as well as increased demand are driving spending levels up. IDC has, in fact, raised its forecast for 2026 and beyond.
IDC reckons that the world spent $89.7 billion on AI infrastructure, including machines with GPUs and other kinds of XPUs powering them, up 33.1 percent from Q1 2025. And just to make it interesting, IDC separated out AI server hosts that were based on X86 processors from Intel and AMD from AI server hosts powered by Arm-based CPU chips. In the AI accelerated segment of the market in Q1 2026, X86 AI hosts took in $34.6 billion in revenues, but thanks to the wide adoption of Nvidia’s “Grace” CG100 server processors in its rackscale NVL72 platforms using its “Hopper” and “Blackwell” GPUs, Arm-based system revenues surpassed X86-based system revenues in the AI segment, brining in
Now, you have to remember that is the value of a server node or a rackscale system with all of its GPUs installed, and the processors represent a relatively small portion of the overall cost of the system. This is especially true as high-speed networking and more complex and dense racks are representing higher portions of the overall cost of the rackscale machines even as GPUs are getting more expensive.
The Arm share of AI hosts was 60.5 percent, while X86 represented only 39.5 percent. Google is also finally using its Axion Arm processors in its TPU clusters, and Amazon Web Services has been using its Gravitons as AI host CPUs for a number of years.
Just for fun, I pulled together the overall Q1 2026 server spending data from IDC and did my best to extract systems revenues for machines using IBM’s Power and z processors from the rest of the non-X86 portion of the market that IDC tracks. As you can see, my best guess puts IBM’s homegrown Power and z machines at around $2.3 billion in sales, a 1.9 percent share of the $122.6 billion in servers sold in Q1 2026. X86 machines represented $63.9 in sales, for a 52.1 percent share, and Arm machines accounted for $56.4 billion, a 46 percent share. It won’t be long before Arm passes by X86, no matter what Intel and AMD do. The hyperscalers and cloud builders and some AI model builders are shifting to cheaper, homegrown CPUs, and Arm was how they got there. (Some day, they will move to RISC-V for even lower costs.) If you do the math, there was about $3.4 billion in Q1 2026 spent on Arm servers that were not AI hosts, compared to $29.3 billion for non-AI X86 systems.
By the way, if you manage to get your hands on IDC’s server shipment data, be careful. IDC counts the chasses that have a particular kind of CPU in them, but it is not counting distinct NUMA domains as it should. In other words, a box with eight Arm server sleds, which are servers in their own right even if they share DPUs for network access, counts as one shipment, not eight. This is stupid. So when you see Arm has only 13 percent shipment share in IDC data, they are counting the server chassis or the rack, not unique, standalone processing machines that just happen to share a common skin. As far as we know, when you talk about distinct compute NUMA domains, Arm has slightly more shipments than X86 in the overall datacenters of the world, thanks to the huge buying by the hyperscalers and cloud builders.
Anyway, here is a chart that shows the Q1 2026 data from IDC as well as its forecast from 2025 through 2030. I pulled back data from 2023 and 2024 to complete the dataset:
On the right side of this table is the Q1 data for 2025 and 2026, including a distribution of AI infrastructure sales in the Q1s of 2025 and 2026 by country of purchase and their growth rates. The United States dominates, but only grew by 30.3 percent $67.9 billion. China is smaller at $7.8 billion and growing slower at 9.3 percent. The Asia Pacific/Japan region is growing Ai infrastructure spending twice as fast as the US, at 62 percent, but only accounted for $5.8 billion in revenues. Western Europe is growing its AI infrastructure spending more modestly, with sale up 20.4 percent to $5.1 billion. The Middle East and Africa had a 3.3X bump, but from a very small base and only just now broke through $1 billion in sales of AI infrastructure.
It is the trend data that we were most interested in, as we said. The compound annual growth rate between 2023 and 2030 for infrastructure spending comes to 49.2 percent, which is similar to what both Nvidia and AMD have told us independently to expect. All growth rates have their limits because a global economy can only do so many transactions, and this does not insult my intelligence at all even if the AI infrastructure spending per year is going above $1 billion in 2029 and growing more reasonably into 2030.
If anything, the IDC AI infrastructure spending data – which includes servers and switches as well as a smidgen for storage – is conservative. Comparing this with the recent AMD forecasts illustrates this. Here is a chart that shows the IDC AI infrastructure spending over time compared to the AMD forecast for AI accelerator spending (compute engines plus HBM memory) hat Su presented last week:
They track nicely, even if they start diverging in 2029 and 2030. But alas, you have to figure out how much server and switch spending those accelerators will drive to do a proper comparison. So I did that in another table shown below and that line is shown in the chart above as the dashed orange line. Take a look:
Just for fun, I made a guess about how much the full datacenter spending would be to support those implied AI system revenues, and then reckoned how much juice these machines in the aggregate might burn.
What you can see is that the AMD forecast is much rosier than the IDC forecast starting in 2027 and beyond. First of all, the multiplier I used to figure out AI system spending from AI accelerator spending is growing over time because networking, storage, and racks are representing a larger piece of the overall system costs. So are CPU hosts and their DRAM and flash. So even as the costs of GPUs and XPUs go up, the costs of these other elements of the AI systems are going up faster.
And thus, there is a big gap between the $1.21 trillion in AI infrastructure spending that IDC expects in 2030 and the $2.11 trillion that I calculated using the AI accelerator TAM figures from AMD.
Do with this what you will. The only way to accurately predict the future is to live it.