--- title: "Do this before you buy an AI server" description: "Before you buy an AI server, name the workload, the owner, the expected use and the room it will live in. The steps that decide whether it pays off." url: "https://www.lmtek.com/source-code/do-this-before-you-buy-an-ai-server" locale: "en" updated: "2026-10-09" --- # Do this before you buy an AI server By Glen Rhodes · Strategic Partnership Manager, LM TEK · Published 9 Oct 2026 ## Key points - The steps that decide whether a first AI server pays off happen before the purchase order, not after delivery. - Four of them: a workload that already runs, a named owner, enough steady use to beat renting, and a room ready for the machine. Call it the Monday Morning Test. - For some teams, renting a while longer is still the right call. Treat that as a plan with a trigger, not a default. The meeting usually goes the same way. The budget has been approved. The quote is on the table, signed off by everyone who needed to sign it. The purchase order is one click away. Then someone asks a simple question: what will the machine actually be running on its first Monday, and who will be looking after it? The room goes quiet. Not because anyone has been careless. Because until that moment, the whole conversation was about which box to buy. I've sat in some version of that meeting more often than I can count. We build these machines, so read what follows accordingly. This piece is about the work that makes buying one pay off, and in my experience nearly all of it happens before the order goes out. For some teams, that work will show the right answer is to rent a while longer, and that's a good outcome too. ## Most buyers compare boxes Buyers spend weeks on the hardware question: how many GPUs, how much memory, which configuration. Those questions matter, and getting them wrong is expensive. But when a first deployment struggles, the box is almost never the problem. The problem is what was missing around it on day one: a job for it to do, a person responsible for it, enough work to keep it busy, or a room that can cope with it. None of those is solved by a better quote. All four can be solved before you sign one. ## The Monday Morning Test Picture the machine's first Monday. It's racked, powered and switched on. Four things decide whether that Monday is the start of something useful or the start of an expensive wait. **1. A workload that already runs.** The strongest first deployments I see move a job that is already running somewhere else, usually on rented compute, with real users. The team knows the volumes. It knows what a good answer looks like. It knows which parts of the job matter and which don't. That is your sizing data. Without it, every hardware decision is a guess, and guesses tend to be either too big or too small. **2. A named owner.** Ask who gets the call when the machine stops answering at eight on a Monday morning. If the answer is "IT, probably" or a long pause, the machine has no owner yet. Owning AI infrastructure is ongoing work: patching, monitoring, updates, and the occasional bad morning. It needs a person whose job includes it, not a person who will fit it in. Who that should be is the subject of *Who feeds the machine: staffing your first AI server*. **3. Enough steady use to beat the rent.** An owned machine earns its keep by being busy. A server that runs one report a week is an expensive paperweight. Whether yours will be busy enough is arithmetic, and the inputs are knowable before you buy: what the machine and its owner cost, how many hours a day the workload runs, what the electricity costs, and how long you expect to keep the machine. Run those numbers against what you pay to rent the same work today. *Break-even: when a GPU pays for itself* walks through the method. **4. A room ready for the machine.** An AI server turns nearly all the power it draws into heat, and it draws a lot. ([TechTarget, 2024](https://www.techtarget.com/searchdatacenter/tip/How-to-calculate-data-center-cooling-requirements)) A machine that was easy to approve can be hard to power, cool and fit into a server room built for the 4 to 6 kW racks that are still the most common size. ([Uptime Institute, 2024](https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf)) Check the power supply, the cooling and the floor space before delivery, not after. Our pieces on the physical layer go into the detail. Workload, owner, use, room. That's the Monday Morning Test. ## When renting a while longer is right Some teams should not buy yet, and I'd rather say so plainly than have them find out the expensive way. If the work is still exploratory and you don't yet know which use cases will stick, rent. If demand comes in bursts rather than steadily, rent; metered pricing suits spiky work. If nobody can own the machine yet, rent while you sort that out. A rented service is a perfectly good laboratory. What I'd push back on is renting as a default nobody revisits. Renting for now works best as a plan with a trigger. Decide today what would make you buy: a data-protection requirement, a monthly bill above a certain size, or a workload that now runs most of the day. Write it down. Then the decision makes itself when the moment comes. ## The decision rule Run the Monday Morning Test before the purchase order. - **All four in place:** the order is the easy part. Buy, and the machine has work waiting on its first Monday. - **One missing:** fix it first. The server can wait a few weeks; an idle server costs more than a delayed one. - **Two or more missing:** keep renting while you close the gaps, and set the trigger for when you'll look again. Buy the server the week you pass the Monday Morning Test. Not before, and not much after. ## Frequently asked questions ### What should I do before buying an AI server? Four things: move or identify a workload that already runs with real users, name a person who owns the machine day to day, check that the workload will keep the machine busy enough to beat renting, and confirm that the room can supply the power, cooling and space it needs. ### When is renting AI compute better than buying a server? When the work is still exploratory, when demand comes in bursts, when nobody can own the machine yet, or when volumes are low. Treat renting as a plan with a trigger, such as a bill size or a data requirement, rather than a default. ### Why do first AI server deployments struggle? Rarely because of the hardware itself. More often the machine arrives without a defined workload, without a named owner, without enough use to keep it busy, or into a room that can't power or cool it properly. ## About the author Glen Rhodes is Strategic Partnership Manager at LM TEK, working with the system integrators and partners who build and deploy its platforms. He previously spent 15 years as a senior product manager in gaming licensing and system integration. ## Read next - [Renting intelligence is the new outsourcing mistake](https://www.lmtek.com/source-code/renting-intelligence-is-the-new-outsourcing-mistake): Why core, steady AI work shouldn't stay rented by default. Free to cite with attribution: Rhodes, G. (2026). Do this before you buy an AI server. LM TEK Source Code. https://www.lmtek.com/source-code/do-this-before-you-buy-an-ai-server