If you knew that you could book $50 billion of revenue, for which you would have to do some of the most complex manufacturing, testing, and delivery in your company’s history, but you would only have operating margins of maybe 5 percent to 7 percent when all was said and done, would you do it? Would you do it if you thought it would give you the street cred to expand from a handful of very big customers to thousands of enterprises buying much smaller – but perhaps more profitable – chunks of infrastructure?

If you were Michael Dell facing a new fiscal year in the GenAI Boom, you would.

That, in a nutshell, is the AI server strategy of the IT hardware company that bears his name. 

Well, almost. Dell says it already has 4,000 enterprise and sovereign AI customers, and it is still only making mid-single digit operating income off its AI business. The reason is simple: Almost of the margin goes to Nvidia, and frankly because Nvidia is doing most of the engineering work to build complex AI server nodes, some of which are rackscale machinery. Sometimes AMD or Intel get a piece of the action with high-end CPUs for AI server hosts when customers want air-cooled, four-way or eight-way GPU nodes, but with the rackscale GB200 and GB300 NVL72 systems from Nvidia, there isn’t much for Dell to do but assemble the parts in a rack and be responsible for making it work.

It is a metal-bending, delivery, and tech support job, and therefore, even five, six, or seven points is pretty good in a world where the big original design manufacturers (ODMs) like Quanta Computer and Foxconn will do it for three points. Buy American is worth a few points, but probably not more than that. And besides, none of the hyperscalers and major cloud builders do buy American, but some of the AI model builders – xAI comes immediately to mind – do.

Now, to be fair, with air-cooled machines with four or eight GPUs, there is a little more engineering work – Dell gets to choose the CPU and the CPU-GPU interconnect and make the motherboard and deal with airflow and power, and therefore it can get more profits. Nvidia doesn’t really want this, but some HPC centers do, as evidenced by the “Horizon” supercomputer being installed at the Texas Advanced Computing Center.

Even with its AI server business growing fast but counterbalancing the profits from sales of traditional, general purpose PowerEdge servers and PowerMax, PowerStore, and PowerScale storage and even with crazy price increases for DRAM main memory and enterprise flash storage for servers, Dell was able to keep the operating income for systems on an even keel for Q4 F2026 ended in January.

Let’s go through the numbers.

In the quarter, Dell had $27.62 billion in product sales, up 53 percent year on year. Services revenues were $5.76 billion, off 2 percent. Add it up, and total revenues were $33.38 billion, up 39.5 percent.

Dell ended the quarter with $13.26 billion in cash and investments and $31.5 billion in debt.

Dell’s operating income was $3.09 billion, up 43.2 percent, and net income was up $2.26 billion, up 47.6. Operating income was 9 percent of revenue for the whole company, to give you some perspective. The PC business, which has been smaller than the Infrastructure Solutions Group datacenter business for three quarters now, typically brings in around 6 points of revenue as operating income; it was 4.7 percent in Q4.

Infrastructure Solutions Group, including all sales of servers and switching and support, had $19.6 billion in sales in Q4 F2026, up 72.7 percent, and operating income was $2.9 percent, up 41.4 percent and representing 14.8 percent of that revenue. Dell is making some coin on traditional server and storage upgrades, which is what it looks like to me, given that AI servers were about half of ISG revenues and are supposed to have margins in the “mid-single digits.”

By the way, if you take AI servers out of the picture, then the rest of Dell, including sales of traditional servers and storage and PCs all together, only rose by 11.5 percent to $24.43 billion.

That is not a very dramatic business story, so as I said at the top of this story, you can see why Dell was so eager to have $8.95 billion in AI server sales, up by a factor of 4.5X over the year ago period.

This is quite a contrast with a few years ago, when the GenAI boom started and the hyperscalers and big cloud builders were getting the cast majority of Nvidia GPU allocations. The traditional OEMs like Dell, Hewlett Packard Enterprise, Lenovo, and Cisco Systems were not invited to the party, and IBM had stopped going to parties years before this.

In fiscal 2023, my model says that Dell’s AI business was maybe $120 million – that’s it. Noise in the data, and not very much noise at that. And then Dell started getting some GPU allocations in fiscal 2024 as Nvidia started getting worried that Dell might cozy up to AMD and also realized that if it wanted GenAI to mainstream, it needed OEMs to actually care about selling its products. Suddenly, AI servers hit $1.81 billion.

Fast forward, and Dell ended Q4 F2026 with a $43 billion AI server backlog and said further that it would make at least $50 billion in AI server sales in fiscal 2027, which started in February. That is double the $24.56 billion in AI server sales Dell did in fiscal 2026, which was 2.5X the $9.73 billion it did in fiscal 2025.

Drilling down a little deeper into the systems business at Dell, here is a financial table that carves out systems and storage and then breaks AI servers and non-AI servers from each other over the past two years by quarter:

Just for fun, I calculated how many eight-way GPU node equivalents that Dell AI server revenue might represent and then guess at the total number of GPUs (from both Nvidia and AMD) that revenue might have driven. This is mostly for fun, and to show how the allocations and the number of systems had ramped. Of course, Dell is selling some GB200 NVL72 and GB300 NVL72 rackscale systems as well.

With 1 gigawatt of datacenter capacity hosting around 550,000 GPUs, then the estimated 420,000 GPUs that Dell sold for all of fiscal 2026 might have driven around 765 megawatts. That’s it.

Here is the other thing. If Dell has over 4.000 enterprise, neocloud, and sovereign AI customers, and it has sold somewhere on the order of 616,000 GPUs total over the three years where it had an AI server business, then after you take out the 50,000 GPUs that went to xAI for its first “Colossus” AI supercomputer (Supermicro got the other half of phase one, and appears to have gotten all of the 100,000 GPUs in phase 2 of the machine), then the average customer Dell AI server customer has bought 142 GPUs, give or take. That is around $5.7 million for the GPUs, on average. We suspect there are a lot of customers with a rack or two of four-way or eight-way GPU servers, dozens with thousands of GPUs, and a few with tens of thousands of GPUs.

One more thing: Dell’s traditional server business is showing signs of slowing growth. In fiscal 2025, the traditional PowerEdge systems business accounted for $17.41 billion in sales, up a healthy 10.1 percent from the $15.82 billion in fiscal 2024. In the fiscal 2026 year just ended, the traditional PowerEdge systems business only grew by 5.8 percent to $18.41 billion, and only represented 44.4 percent of total systems sales.

Dell is finally reflecting what has been happening in the overall market as the hyperscalers and cloud builders have skewed server spending toward AI systems, and this is probably a permanent shift as we look to the years ahead. There may be a quarter or two where they are close, from time to time, but the overall trend will not reverse.