I continue to be impressed by Dell.
Dell’s Infrastructure Solutions Group (ISG) is a good proxy for enterprise AI demand, and that demand is accelerating. It took Dell eight quarters to sign its first 3,200 AI customers, but only three for the next 3,300.
Dell also makes, believe it or not, the best agentic PC I’ve used. And it has NOTHING to do with NPUs, “AI PC” branding, or running inference locally!
The best agentic PC experience will be on machines that agents can easily customize to fit the user’s bespoke needs. This implies an OS the agent can easily control and reconfigure. Sure, a Mac lets an agent do a lot from the terminal, but macOS is pretty locked down. Windows is a bit better, but still scatters settings across the registry and the Settings app and isn’t totally open. But Linux! On Linux, nearly every setting is a text file or a command, so an agent can read it, change it, and reload it in seconds.
Dell sees it too. Jeff Clarke, Dell’s COO, said in May that customers “are looking for more capable PCs as agentic workloads make their way to the edge”. Dell’s commercial PC demand has now grown 10 quarters in a row.
I’m just one data point, so let’s not overindex on me. But I’m experiencing firsthand how the PC landscape is shifting thanks to agents. Yes, cloud agents for now, and that’s fine, it’s just a bit slow sitting in queues waiting for my turn for tokens.
My daily driver was a Mac. It has a nice user experience and industrial design, and it’s good for programming too, because macOS is Unix underneath. But for the past month I’ve switched away from Mac and now use a Dell XPS (with Intel 18A inside!) running Omarchy, and dude, it’s amazing.
x86 laptop makers have tried for years to match the MacBook, and mostly haven’t. Windows made it harder. The XPS is the exception. The screen is excellent. The keyboard isn’t great. And Linux on the XPS isn’t new for Dell, which has shipped Linux developer editions since Project Sputnik in 2012.
Omarchy is DHH’s opinionated Linux setup (DHH created Ruby on Rails and co-founded 37signals). It’s Arch Linux with the Hyprland tiling window manager. You drive it from the keyboard. Every setting lives in a plain text config file. Minor thing, but cool implication is that you can configure it to your liking and just commit that config file to GitHub and nearly instantaneously set up new machines or have a backup of your setup.
Yeah that’s a mouthful and I didn’t know anything about Hyprland either. BUT I listened to this incredible conversation between DHH and Lex Fridman and was inspired.
Yeah there’s a learning curve. I don’t know how to take a screen grab on Linux, for example. But I’m going agent first on this bad boy, so I went to ask Claude. And then I thought, screw it, I don’t need to learn anything new, so I told Claude to map the print screen command to Cmd + Shift + 4, the Mac reflex my fingers already have. Claude added one line to my Hyprland config and it worked right away. LFG!
No, I don’t run the inference on the laptop. Actually, honestly, I run a lot of my agents over ssh (using herdr, a terminal tool for juggling agent sessions) on an “agent computer”, a small AMD-powered box tucked next to our piano that runs my agents around the clock. And even then, most of the inference is in the cloud, although I do some small models and transcription stuff on the AMD GPU.
What about my Mac? Well, I keep it plugged in and on (even with the lid closed) so it can forward iMessage to my Dell computer :) And of course there’s an open-source Linux program that makes it simple for me to feel like my XPS is in the iMessage ecosystem.
Enterprise AI
Microsoft, Google, Amazon and Meta mostly design their own racks and have ODMs like Quanta, Wiwynn and Foxconn build them. Most everyone else buys from an OEM like Dell, HPE, Supermicro or Lenovo. That’s neoclouds, sovereigns, banks, hospitals, HFT shops, insurers and retailers.
Dell is the biggest of those OEMs in AI servers. It recognized $16.4B of AI server revenue in the quarter ending July, while Supermicro’s total revenue in its June quarter was $11.1B and HPE’s AI systems revenue was about $1.6B. Dell booked $131.7B of AI server orders over the last four quarters and ended July with a $95B backlog. On its September call, it raised its FY27 revenue guide by $25B, to $192B.
Understandably, most investors focus on the buildouts supporting Anthropic and OpenAI and overlook enterprise AI. But this trend is long in the making; for a few years now Nvidia continues to comment on the growth of AI in the enterprise. And given that the lion’s share of accelerators are Nvidia, they obviously have a good read on the situation.
In fact, Nvidia now splits its data center business out in a way to give us a bit clearer read on enterprise AI. In August it broke $89B of data center revenue into hyperscale ($49B, up 13% sequentially) and what it calls ACIE ($40B, up 25% sequentially and 138% year over year). ACIE is neoclouds, sovereigns, industrial and enterprise customers.
Jensen on the call:
Most people see just hyperscalers. And that’s half of the picture. The other half of the picture is what we call ACIE and that’s all the enterprise, the neoclouds, the sovereign AIs, that part of the world is invisible to everybody.
Colette Kress put a number on enterprise too, saying Nvidia’s trailing-12-month on-prem revenue was $8B from automotive plus $7B from financial services, manufacturing and healthcare combined.
So Nvidia confirms what Dell sees. But... ACIE is a bit of a muddy read on enterprise, because neoclouds are bucketed in there too. And given that hyperscalers rent a lot of that neocloud capacity... it’s not a pure enterprise read-through.
Yes, Dell’s AI server revenue has the same “muddiness” since a lot of its revenue comes from those same neoclouds. But Dell also shares an AI customer count most quarters (4,000+ in February, 6,500+ in September); clearly an expanding long tail, even if many are small buyers. And its traditional server line, up 122% last quarter, is “primarily our historical enterprise customers”. And yes, those CPUs aren’t all just driven by the AI buildout, but include traditional upgrades. Yet we still can look at Dell and say enterprises are buying AI racks.
And frankly, we’re early and those racks are still quite expensive, in large part due to the accelerators and memory inside.
So here’s what I’ll dig into:
Why are enterprises buying their own AI hardware instead of renting tokens?
Can Dell’s enterprise growth tell us how fast open-weight models are catching on?
Does on-prem AI have to mean a rack?
Why is an enterprise AI customer worth more to Dell than a neocloud?
Why do agents drive demand for CPUs and storage, and not just GPUs?
How much of this demand is customers buying ahead of memory price increases?
What happens if three neocloud customers make up most of Dell’s AI server number?
Does Dell moving memory from PCs to servers hurt the agentic PC story?
Let’s get into it.