More Upward Revisions On AI Infrastructure Spending

Here is a question for you. What is more difficult: predicting the weather thirty days out or what global AI infrastructure spending will be out to the end of the decade?

We don’t know, but it sure looks like someone is gonna spend a lot of money on infrastructure either way.

The Next Platform has always been interested in two things: What technologies are transforming large scale infrastructure, and how much money organizations are investing in particular technologies right now. The prognosticators at IDC put out a pretty skinny forecast for global AI spending between 2025 and 2029, including hardware, software, and services, back in early September, which we picked apart here. And today, the augury wizards at the market researcher put out some data about AI infrastructure spending – covering GPU and XPU accelerated servers, regular servers, and storage – between 2024 and 2029 and drilling down into spending in the second quarter of this year as well.

It takes a long time, apparently, to sort out the numbers because the third quarter is over and IDC is only talking about AI infrastructure spending in the second quarter.

As often happens with these IDC forecasts, we get a chart that we have to reverse engineer into numbers by counting pixels and then flesh out with extra data that IDC gives in the analysis accompanying the chart. So, here is the chart:

This chart above shows the value – revenue to OEMs and ODMs for building machinery for end customers – of AI-accelerated servers (mostly using GPUs but sometimes AI-focused XPUs) versus plain vanilla, non-accelerated servers.

Here is what happens when you count pixels and mix and match the data from Q2 2025 that IDC talked about in its statement:

IDC did not talk specifically about AI storage infrastructure spending in the chart above or in its statement except for what combined server and storage infrastructure would add up to in its forecast for 2029, which was $758 billion. When you back out the server spending, that leaves $10.5 billion in storage spending., a mere 1.4 percent of total infrastructure spending. We think IDC is undercutting storage and we would bet Sam Altman’s last $10 billion of credit that storage will be a much larger part of the AI systems that companies build in the coming years.

As usual, any estimates made by us are shown in bold red italics.

Like other prognosticators, IDC has had to go back to its forecasts and revise them upwards dramatically as the big model builders – mostly OpenAI and Anthropic – and the neoclouds and the regular big cloud builders are all shelling out big bucks to buy all of the GPUs they can to help the world train its models. IDC had been expecting a slowdown in late 2025 and early 2026 and now says that AI server spending will continue ramping in the second half of this year and into next year.

As you can see in the right side of the table that we made above, the vast amount of spending for AI infrastructure in the second quarter was for AI servers accelerated by GPUs and XPUs. To be specific, in Q2 2025, servers of all kinds accounted for 98 percent of the $82 billion in total AI infrastructure spending, which works out to $80.4 billion and which was up more than 2.7X year on year. And AI servers with GPU or XPU accelerators accounted for 91.8 percent of all AI server spending in the quarter, which maths out to $73.8 billion in the second quarter and which increased by nearly 3.1X year on year. If you do the arithmetic, AI storage hardware only drove $1.6 billion, up 16.1 percent, in the second quarter. Again, this seems low to us.

By the way, in Q2, the hyperscalers, cloud builders, and neoclouds (which IDC calls digital service providers) accounted for 86.7 percent of total AI infrastructure spending. (Remember, this is not the revenue they generate from selling capacity to end users, but the money they spent buying AI infrastructure.) This worked out to $71.1 billion.

With the trade war going on between the United States and China, which is also a tech war between the two countries, a breakdown in AI infrastructure spending in Q2 2025 by country is perhaps illustrative:

Remember, this is not where the machinery is built, but where it is deployed. The United States, thanks to the hyperscalers and cloud builders (and mostly thanks to Amazon Web Services, Microsoft Azure, Meta Platforms, Google Cloud, and Oracle Cloud Infrastructure), spent 6.6X more dough on AI infrastructure than did companies in China. If President Trump has his way, the distribution of machinery value by where it is built will look more like the distribution of where the money was spent. This distribution by country does not necessarily reflect where the machines are installed. OpenAI, for instance, is getting some big money from the Middle East, spending it in the United States, and deploying some of it in the Middle East. EMEA might be a lot stronger with AI infrastructure than it looks in this second table above.

Here is the thing: IDC says that spending on AI infrastructure in the United States will grow at a compound annual growth rate (CAGR) of 40.5 percent in the United States and will only grow slightly faster at 41.5 percent CAGR in China, which means the United States will still outspend China on AI infrastructure by a considerable margin by the end of the forecast period of 2029. AI infrastructure spending in EMEA will have a 17.3 percent CAGR and in AP/J will have a 14.3 percent CAGR between 2024 and 2029.

And finally, let’s bring together the overall AI forecast that IDC did two months ago with this AI infrastructure forecast today:

We are not sure where AI networking is in any of this data, but when we subtract out AI infrastructure server and storage spending from overall AI spending, we get a proxy for AI software and services spending. We set the growth of overall AI spending between 2025 and 2029 to the 31.9 percent compound annual growth rate for that period; obviously it will not be that regular, but we do not know what else to do for that in-between time.