We live in the future, and sometimes it is just amazing how different things are today from two decades ago when the age of accelerated computing got its start in the HPC ModSim arena, eventually being adopted as the architecture of choice for AI applications.

The market is utterly transformed, or rather, a new set of infrastructure suppliers have taken over AI system manufacturing and distribution and the venerable companies peddling systems of record – IBM, the various parts of Hewlett Packard Enterprise that started out as Tandem and Digital Equipment, Sun Microsystems – are either gone or have businesses that are so much smaller relative to a much embiggened systems market that they do not even rise to the top vendor rankings.

I have spent some time building this chart below, which shows the sweep of changes in datacenter compute leading up to the Dot Com boom, when Internet technologies were first broadly commercialized and when the tremendous amounts of data to train AI models was first being digitized and amassed. The data is quarterly server revenues as reckoned by the box counters at IDC, which is the dataset that has been most broadly available to the public and arguably the most useful over the decades that IDC gave out summary data to help peddle its deeper market research and keep its name out there as an IT expert.

During the rise of GenAI, both IDC and Gartner stopped giving out public data, and the blue line in the chart shows that gap. IDC recently started giving out quarterly data, and we have added this to the data set and filled in some gaps with estimates. We have not adjusted this data for inflation, but probably should. All it would do is rotate the data around whatever year you picked as the normalized US dollar. The data to the left of the normalized year would lift further as you moved to the left into the past, and the data to the right of this normalization point would be pushed down with increasing vigor as you move to the right – both due to the effects of inflation.

The green line shows the peak of server revenues during the Dot Com boom, a level that in absolute dollars was not reached consistently for two decades. Assuming the market for systems of record is relatively stable at around somewhere between $15 billion and $18 billion a quarter– for most back office systems, Moore’s Law advances plus packaging and networking enhancements offer more incremental capacity than they could ever dream of using – then the remaining more than $100 billion in sales of systems are for AI capacity or the front office and back office systems that feed into them. In that sense, AI is driving far more than half of the revenues in the server market.

What is immediately obvious is that if the GenAI bubble bursts, as many worry it might, there is going to be enough soap for many, many companies to drown in. A GenAI bust, should it come to pass, will make the Dot Com bust, the Great Recession, and other collapses in the server space look like a joke, as you can plainly see. Nvidia will be hit the hardest, but many OEMs and ODMs, countless chip suppliers, all the clouds big and small, up and down the value chain, will take their own hits.

The magnitude of such a collapse is hard to grasp. This GenAI boom in its totality – datacenters and equipment, but also power, cooling, land, and financial vig – is a significant driver of the global economy, and is largely a phenomenon of the United States, with China to a less degree (which is odd, but more on that in a moment) and sovereign nations starting to catch the GenAI religion.

The good news, if you want to global economy to somehow hold itself together, is that it looks like demand is well exceeding supply, and the gargantuan AI projects that have been outlined are still moving ahead.

In the fourth quarter of 2025, the most recent data just released by IDC, server sales grew by 52.4 percent to $125.3 billion, which also represented an incredible 11.4 percent sequential increase from Q3 2025. Based on what IDC said in its report, we calculate that of this vast and expensive pile of machinery, $70.65 billion was for GPU-accelerated systems, which represented 56.4 percent of all server revenues. We do not know how many XPU-accelerated system sales are in there, but it would have to be a substantial amount. We do not think that IDC is mixing revenues of systems with GPUs and XPUs, or it would have said as much. We wish IDC would clarify this because we have a hard time believing there was $54.65 billion in back office and front office system sales in the quarter.

What IDC did say – and which is a distinction that is not all that useful any more, but you never throw away data – is that X86 CPUs were used in $69.8 billion of machinery in Q4 2025 (up 16.9 percent year on year) compared to $55.5 billion in non-X86 servers. While IBM’s Power Systems and System z mainframes are some of that non-X86 iron, the vast majority by far is now Arm-based servers made by the hyperscalers, the cloud builders, and now the AI model builders. That was 2.5X growth year on year for the non-X86 camp, and it will not be too long before the value of machines based on Arm will exceed the value of machines based on X86.

Mostly because of the value of the GPUs attached to the Arm systems, but some of it is indeed Arm-based infrastructure servers that have no accelerators.

Dell is by far the largest of the server OEMs now, bringing in $12.65 billion in sales, up by a factor of 2.3X year on year thanks to some large AI system deals. Supermicro, despite its current difficulties with the US government over alleged smuggling of sanctioned GPU systems into China, had a very good Q4, with server sales of $11.7 billion, also up 2.3X year on year. IEIT Systems, an upstart server maker out of China that is now the number three OEM in the world, had $5.19 billion in sales, up 33.7 percent. Sino-American system maker Lenovo, which is ranked number four, had $5.07 billion in server revenues, up 34 percent, while HPE, which is getting more and more picky about deals because it doesn’t want to take lower margins on AI deals, had $3.87 billion in sales, down 8.6 percent year on year but up 14 percent sequentially.

Other OEMs comprised $20.29 billion in server revenues in Q4 2025, up 11 percent as a group, while the ODMs – Foxconn, Inventec, Quanta, WiWynn, Jabil, and others – as a group shipped $66.62 billion in server gear, down a smidgen sequentially but up 60.5 percent year on year.

Which brings us to the geopolitical part of servers.

“The United States is the fastest growing region in the server market, with an increase of 72.4 percent compared to the fourth quarter of 2024, fueled by 80.1 percent growth in the accelerated server segment,” IDC wrote in its report. “Canada grew 70.7 percent, pushed by the same reason. EMEA and APeJC also showed double digit growth, with 43.6 percent and 27.9 percent, respectively. PRC and Latin America showed smoother but healthy growth of 17.7 percent and 12.8 percent each while Japan declined by 4.7 percent as it couldn’t match an important investment a year ago.”

What IDC doesn’t tell us is what the server spending amounts are by country, but clearly China is now growing its server budgets anywhere near the rate that we are seeing among the tech titans in the United States. This seems perplexing to us. If China is so much smarter about AI, as many contend it is and

as the DeepSeek model first demonstrated back in January 2025

, then maybe the total addressable market in China is not as big as we might be led to believe.