The Hidden World of AI Scrap Heat: What Happens to the Energy That Trains Large Models?
Everyone is talking about what AI can or cannot do, but very few people mention the physical reality behind it. Artificial intelligence may feel effortless, but the machinery powering it produces enormous amounts of heat. The electricity used to train large models becomes waste heat on a scale that rivals heavy industry, and in several cities, that heat is already being reused to warm homes, buildings, and even public pools.
Behind every model is an industrial-scale network of servers, power systems, cooling loops, and GPUs running at full tilt. These machines do more than process information; they generate heat in volumes large enough to influence how some cities plan their energy infrastructure. As AI continues to expand, that heat is becoming one of its most significant and least discussed by-products.
AI systems do not simply process data. They radiate heat, and the rising demand for AI is turning that heat into a resource some cities already depend on.
AI Training Is a Thermodynamic Machine, Not Digital Abstraction
When large AI models train, almost all the electrical energy consumed by GPUs is converted into heat. Research from the International Renewable Energy Agency explains that nearly all power entering data center equipment ultimately becomes thermal energy. You can read the analysis in IRENA’s waste-heat recovery report.
Individually, a GPU drawing 500 to 700 watts seems harmless. At scale, thousands of GPUs running simultaneously can generate continuous heat loads of 10 to 50 megawatts. The largest AI campuses exceed 100 megawatts, rivaling the energy footprint of a small city.
Every watt of that energy becomes heat that must be managed or reused.
Where All That Heat Normally Goes
Traditionally, data centers rejects excess heat through chillers, cooling towers, and heat exchangers. At AI scale, this becomes inefficient and expensive. The more compute required, the more valuable that “waste” heat becomes.
Some Cities Already Use Data Center Heat to Warm Homes
AI heat reuse may sound speculative, but several European cities already use data center waste heat as part of their municipal energy systems.
Finland
In Mäntsälä, Finland, a data center supplies recovered heat to roughly 2,500 homes, as Bloomberg reported.
Helsinki
Microsoft’s new data center developments in Helsinki will feed heat into the region’s largest district heating system. According to the World Economic Forum, the recovered heat could support up to 250,000 homes.
Odense, Denmark
Meta’s Odense data center recovers heat for the local utility, warming about 11,000 homes. The system was highlighted in an Equinix sustainability briefing.
Stockholm, Sweden
Stockholm’s Data Parks initiative encourages operators to connect data centers directly into the city’s district heating system. Details of the program are covered by RESET.org.
Paris, France
Several public pools in Paris now receive supplemental warmth from nearby server-cooling loops. These heat reuse systems are also documented by RESET.org.
These systems are not pilot projects. They are functioning components of public heating infrastructure.
Where Heat Recycling Works Best
Heat reuse is real, but not universal. Several conditions make it viable:
- Cold climates with long heating seasons
- Established district heating systems
- Proximity between data centers and heat customers
- Infrastructure planned during data center construction
The Open Compute Project notes these constraints in its overview of heat reuse feasibility, which you can read in the OCP Heat Reuse 101 document.
Why Heat Reuse Is Growing Quickly
1. AI Power Consumption Is Increasing Rapidly
Cooling 10 megawatts of GPUs is difficult. Cooling 100 megawatts pushes existing systems to their limits. Reuse becomes a practical, sometimes necessary, solution.
2. Cities Are Seeking Cleaner Heating Alternatives
Municipalities can replace fossil-fuel boilers with waste heat, improving sustainability while reducing carbon emissions.
3. Data Center Operators Can Monetize Excess Heat
Instead of paying to reject thermal energy, operators can sell it, turning a cost burden into a revenue stream.
A Future Where Homes Are Heated by AI Workloads
The basic thermal loop looks like this:
AI computes → GPUs generate heat → Heat is captured → Heat enters district heating → Homes stay warm.
A 2025 review in Renewable and Sustainable Energy Reviews notes that cities worldwide are now evaluating similar approaches as AI infrastructure grows.
This makes it increasingly plausible that the heat generated during large language model training will become part of future urban heating strategies.
AI Scrap Heat as Future Infrastructure
Urban planners and researchers are currently examining several potential applications for data center waste heat:
- Greenhouses warmed by high-density compute loads
- Warehouses and logistics hubs using surplus thermal energy
- Residential buildings connected to heat-recovery loops
- Industrial parks co-designed with AI compute infrastructure
As AI demand accelerates and energy standards tighten, integrating these heat sources into public systems may become both economically and environmentally necessary.
The Internet Has Always Been Physical
Despite its intangible appearance, the internet is deeply physical. Beneath every AI request is an ecosystem of substations, cooling plants, heat exchangers, pumps, and thousands of GPUs obeying the laws of thermodynamics.
AI may feel invisible, but its heat is not. As compute workloads grow, the connection between digital intelligence and physical energy systems will become increasingly visible.
Large models think, and cities grow warmer.