--- title: "AI is a heat problem wearing a software costume" description: "Every watt an AI server draws becomes heat in the room. Why cooling, not the model, often sets the ceiling on owned AI, and how to size the room first." url: "https://www.lmtek.com/source-code/ai-is-a-heat-problem-wearing-a-software-costume" locale: "en" updated: "2026-10-07" --- # AI is a heat problem wearing a software costume By Marko Kranjac Došen · Technical Product Manager, LM TEK · Published 7 Oct 2026 ## Key points - The AI conversation is about models, prompts and agents. The limit on owned AI is often physics: every watt a server draws ends up as heat in the room. - In the cloud, heat is someone else's problem, priced into the bill. Own the machine and it moves into your building. - A GPU that can't shed its heat protects itself by slowing down. The room, not the datasheet, sets the ceiling. Size the room before you size the model. A 5 kW AI server is a 5 kW heater. It runs all day and all night, in whatever room you put it in, and it never takes a weekend off. That isn't how anyone talks about AI. The conversation is about models, prompts, agents and benchmarks, because the interesting part of AI looks like software. But the moment you own the machine, the software is the costume. Underneath is a box that turns electricity into heat. ## Where the watts go Start with the physics, because it's short. The electricity a server draws doesn't disappear. A tiny fraction leaves as signals on network cables. Practically all the rest becomes heat, in the room where the server sits. A machine drawing a given number of kilowatts is, for the room's purposes, a heater of the same number of kilowatts. ([TechTarget, 2024](https://www.techtarget.com/searchdatacenter/tip/How-to-calculate-data-center-cooling-requirements)) The arithmetic is simple enough to do on a napkin. Take an illustrative AI server drawing 5 kW under load. (The number is round for the sake of the maths, not a spec.) | Quantity | Value | How | |---|---|---| | Power drawn | 5 kW | illustrative | | Heat into the room | 5 kW | power in = heat out | | Same heat in BTU per hour | about 17,000 BTU/h | 5 × 3,412 | | Heat per day, running continuously | 120 kWh | 5 kW × 24 h | | Household comparison | two or three 2 kW fan heaters, on all day and night | 5 ÷ 2 | Nobody would put two or three fan heaters in a cupboard, close the door and expect it to stay cool. With an AI server, that's roughly what a first deployment can look like if nobody runs the napkin maths. And the number only goes one way as the model grows. A larger model needs more GPU memory, more memory means more cards, and more cards draw more power. [*How much GPU does a model actually need? The VRAM math*](/source-code/how-much-gpu-does-a-model-actually-need) works through that chain. Every step along it adds kilowatts, and every kilowatt arrives in the room as heat. ## Heat is in someone else's invoice If you rent AI through an API, you're right to ignore all of this. The heat is real, but it lives in someone else's building. Your provider pays for the power, the cooling and the facility, and you pay for them indirectly through the price per token. For a renter, heat is a line inside someone else's invoice. For a large share of today's AI users, cooling really isn't their problem. It becomes their problem the day they decide to own. Bringing AI in-house brings the physics in-house with it. The power supply, the cooling and the floor space were someone else's engineering yesterday. Today they're yours. ## Hot GPUs don't fail. They slow down. A server in a room that can't remove its heat usually doesn't fail dramatically. It just quietly gets slower. GPUs protect themselves. When they get too hot, they reduce their clock speeds to stay within safe temperatures, a behaviour usually called thermal throttling. NVIDIA's own monitoring library reports it as a "thermal slowdown", and the hardware version cuts core clocks by a factor of two or more. ([NVIDIA NVML documentation](https://docs.nvidia.com/deploy/nvml-api/group__nvmlClocksThrottleReasons.html)) The model still runs. The answers are still correct. They just take longer, and the machine you sized for a given number of users serves fewer of them. That's what makes it expensive. Unless someone is watching the clock-speed telemetry, nothing alerts you. Meanwhile you're getting less of the performance you paid for, every day, because the air around the machine is too warm. Call it the silent performance tax. You paid for the full machine; the room decides how much of it you get to use. The room sets the ceiling. Not the GPU. ## Size the room before you size the model So reverse the usual order. Before the model, before the GPU count, before the quote, answer three questions about the room. 1. **How much power can the room supply** to this machine, on its own circuit, with headroom? 2. **How much heat can the room remove,** continuously, on the hottest day of the year? Write it in kilowatts, the same unit as the power. 3. **Is the second number at least as big as the server's power draw,** with margin for everything else already in the room? If the answer to the third question is no, you have three ways forward. Improve the air handling in the room. Move the heat out more efficiently, which is where liquid cooling earns its place, because liquid carries heat away far more effectively than air. Lawrence Berkeley National Laboratory puts it in volumes: about 1.5 m³ of water holds as much heat as about 5,380 m³ of air for the same temperature rise. ([LBNL, 2021](https://datacenters.lbl.gov/sites/default/files/Sartor%20IEEE%20Madison%20052621%20%281%29.pdf)) Or put the machine somewhere that can take it. Which of those fits is the subject of *Air vs liquid cooling: which is better for your AI hardware?*. The room you have is probably smaller than the machine you want. In the Uptime Institute's 2024 survey, the most common data-centre racks run at 4 to 6 kW, and the average typical rack is below 8 kW. ([Uptime Institute, 2024](https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf)) One illustrative 5 kW server can use most of that budget on its own. What doesn't work is the fourth option: buy the machine, put it in the room you have, and hope. Before you ask how big a model you can run, ask how many kilowatts your room can carry away. That number is your real ceiling. ## Frequently asked questions ### How much heat does an AI server produce? Practically all the electrical power an AI server draws becomes heat in the room. A server drawing 5 kW adds about 5 kW of heat, roughly 17,000 BTU per hour, for as long as it runs. ### What happens if an AI server gets too hot? GPUs protect themselves by lowering their clock speeds when they get too hot, which is usually called thermal throttling. The server keeps working, but more slowly, so it serves fewer users or takes longer per answer than its specification suggests. ### Do I need liquid cooling for an AI server? It depends on how much heat the machine produces and how much the room can remove. Work out the server's power draw and the room's cooling capacity in kilowatts first. Where air can't keep up, liquid cooling moves heat away more effectively; better air handling or a different location are the other options. ## About the author Marko Kranjac Došen is the Technical Product Manager at LM TEK, where he is responsible for platform development projects. ## Read next - [How much GPU does a model actually need? The VRAM math](https://www.lmtek.com/source-code/how-much-gpu-does-a-model-actually-need): The memory chain that decides how many cards, and so how many kilowatts, a model needs. Free to cite with attribution: Kranjac Došen, M. (2026). AI is a heat problem wearing a software costume. LM TEK Source Code. https://www.lmtek.com/source-code/ai-is-a-heat-problem-wearing-a-software-costume