Turning Distributed Server Heat Into Useful Building Heating Across Switzerland
14 Swiss building digital twins · 136 liquid-cooled racks · 484 MWh/yr recovered heat · 0 kWh wasted
Basel Klingental Home
4 x 1 kW racks · 34.1% annual heat covered
Basement Heat Exchanger
80% heat transfer (55/45 C water) · 35 kWh buffer
Gartenbad St. Jakob Pool
Absorbs summer compute · 96.2% utilisation
Vaibhav Motwani · Prakash Aryan · Vilas Harapanahalli
Scan either QR code to open the live Swiss digital twin or interactive deck
[0:00-0:25] Data centers reject nearly 100% of their electricity as waste heat, while Swiss buildings still pay every winter to heat water. CalorGrid links 136 liquid-cooled servers across 14 real Swiss building twins, recovering 484 MWh of useful heat per year with zero wasted energy.
Moving servers into a single home saves heat in winter, yet breaks down across 3,816 warm summer hours
Basel 2025 Reference House: 31.2 MWh/yr demand (206 cold days), but only 12 kWh/day hot-water demand in July
Cloud facilities consume electricity to run compute, then spend extra cooling power venting that thermal energy into the air while nearby buildings burn gas or electricity for heating.
Running 4 x 1 kW servers continuously in one Basel home covers 32.7% of winter heat, but dumps 3,790 kWh/yr outdoors once the 35 kWh buffer fills in summer (raising OPEX to 0.255 CHF/srv-h).
Thermal-led pausing eliminates dumping (0.0 kWh) and cuts electricity by CHF 1,401/yr, yet servers sit idle across summer, leaving utilisation at 38.0% and full CAPEX+OPEX at 0.65 CHF/srv-h.
Basel 2025 benchmark: Open-Meteo 8,760 h (206 heating days, -5.6 C min, 31,231 kWh demand), ElCom 2026 IWB H4 (0.3325 CHF/kWh excl. VAT), 4 x 1 kW servers, 35 kWh buffer tank.
[0:25-1:05] Here is the core problem. Traditional data centers vent their heat to the sky. Installing 4 liquid-cooled servers in a Basel house performs effectively in winter, covering 34% of demand. Across 3,816 warm summer hours, uncoordinated servers reject 3,790 kWh outdoors, while pausing them leaves hardware idle and raises full cost to 65 Rappen per server-hour.
Heat stays local in the basement; compute migrates in 2-4 ms over Swiss fiber to buildings with active thermal demand
Before vs. With CalorGrid (8,760 h Engine)
| Operating Mode | Dumped | Util | OPEX / Full |
|---|---|---|---|
| Uncoordinated House | 3,790 kWh | 50.0% | 0.255 / 0.57 |
| CalorGrid Smart House | 0.0 kWh | 38-65% | 0.226 / 0.47-0.65 |
| CalorGrid + Summer Sink | 0.0 kWh | 96.2% | 0.226 / 0.39-0.43 |
| CalorGrid 14-Site Fleet | 0.0 kWh | 89.6% | 0.193 / 0.51 |
Sealed non-conductive dielectric coolant cools the servers and transfers heat via a plate exchanger (CDU) into the 35 kWh water buffer at 55/45 C, heating buildings directly without a heat pump.
Every hour, our Python controller (decide_live_step()) checks tank state-of-charge and weather, dispatching or suspending batch jobs (OCR, embeddings, LLMs) via Linux SIGSTOP/SIGCONT.
When residential buffers fill in summer, CalorGrid migrates batch compute in 2-4 ms over Swiss fiber to active summer heat sinks: pools, dairy farms, glasshouses, boreholes, hospitals and Alpine spas.
All scenarios satisfy strict hourly energy conservation (residual <= 7.3e-9 kWh). Paired sink figures span full vs zero secondary-sink heat credit; central server CAPEX = 1,400 CHF/srv-yr.
[1:05-1:50] CalorGrid solves this in three steps. First, direct liquid cooling captures 80% of server electricity as 55 degree warm water. Second, our thermal scheduler tracks the 35 kWh buffer and pauses local batch jobs before a single watt is dumped. Third, instead of building district pipes, we shift batch compute over Swiss fiber in 2 to 4 milliseconds to pools, dairy farms, boreholes and hospitals that need heat all summer, lifting utilisation to 96% and reducing full break-even from 65 down to 39 Rappen per server-hour.
Live interactive 3D web twin · one FastAPI engine powers the 3D thermal model and live Linux CPU workloads
tesseract PID in milliseconds; thermal demand resumes the same PID.
80% of electrical input returns as warm water to the buffer tank, closing the real-time energy conservation loop
Verified live day runs: Jan 15 (212.59 kWh demand, 38.40 kWh used, 0.0 kWh dumped); Jul 15 Smart (15.0 kWh elec, 12.0 kWh used, 0.0 kWh dumped) vs Baseline (48.0 kWh elec, 26.4 kWh dumped).
[1:50-3:05] Here is the live system. On the left, our Basel 3D cutaway on January 15 at minus 4.5 degrees: all 4 servers deliver 38.4 kWh straight into the house. In the middle, July 15: uncoordinated servers trigger heat rejection, while CalorGrid caps output at 0 kWh dumped. On the right, our live Linux worker executing Tesseract OCR suspends on SIGSTOP once the buffer saturates and resumes the exact same PID when demand returns. [Switch to the live CalorGrid web app for 45-second interactive demo.]
How we compute thermal output across real server models, rendered in Isaac Sim 6.2 RTX and calibrated on 5 public datasets
How We Derive Thermal Output From Any Server
1. Electrical load: P_elec(t) = P_idle + (P_peak - P_idle) * u(t) calibrated from HEATWISE 5-min steps & MLPerf power logs.
2. Two-loop heat exchange (80%): Primary dielectric coolant transfers Q_water = 0.80 * P_elec via the Liebert plate CDU into 55/45 C building water (m_dot = 1.15 L/min per 1 kW at dT = 10 K).
3. Basement air fraction (20%): Q_room = 0.20 * P_elec warms the lumped room node (strict conservation: residual <= 7.3e-9 kWh).
| Real Server Class & Dataset | Cooling | Idle -> Peak | Water Heat (80%) | Role in CalorGrid |
|---|---|---|---|---|
| Edge Mini PC (Ryzen 5 5560 · MLPerf) | Air | 27 -> 34 W | 0.027 kW_th | Edge baseline (0.155 J/img) |
| 9x OCP Winterfell (18x Xeon · HEATWISE) | Hybrid 54.9 C | 0.67 -> 2.93 kW | 2.34 kW_th | Measured idle draw & temp cap |
| CalorGrid 4x 1 kW Rack (Aquasar class) | Liquid 55/45 C | 0 -> 4.00 kW | 3.20 kW_th | Modular building heating unit |
| Lenovo 8x NVIDIA B200 (MLPerf v5.1) | Liquid W45 | 7.47 -> 10.4 kW | 8.32 kW_th | AI cost (0.0063 CHF/1M tok) |
3D OEM CAD (NVIDIA Isaac Sim 6.2 RTX, CC BY 4.0): Vertiv VR3307 42U Rack, 16OCP ORV3 Rack, 33 kVA Power Shelf, Liebert XDU1350 CDU; NVIDIA Edge Node & 1,000 L Buffer; ForgeCAD Valve.
[3:05-3:45] Every thermal parameter is empirically grounded. On the left are OEM 3D server models rendered in NVIDIA Isaac Sim: an Edge PC, a Vertiv VR3307 enterprise rack and a Vertiv ORV3 liquid rack feeding a Liebert XDU1350 heat exchanger and 1,000-liter buffer tank. We calibrate electrical draw and 80% liquid thermal output directly from five SHA-256 pinned datasets: HEATWISE, MLPerf v5.1 up to 8x NVIDIA B200, Open-Meteo 2025, ElCom 2026 tariffs and federal swisstopo 3D geometry.
136 liquid-cooled racks linked across 10 cantons · 484 MWh/yr useful heat · 0 kWh dumped nationwide
Compact Utility Closet
4 x 0.25 kW Edge · 0.8 kW heat (45 C)
Basel Concrete Basement
4 x 1.0 kW HydroNode · 3.2 kW heat (55 C)
Emmental Dairy Utility Room
3 x 2.9 kW Winterfell · 6.9 kW heat (52 C)
St. Jakob Pool Plant Room
2 x 10.4 kW B200 AI · 16.6 kW heat (45 C)
All 14 Swiss building exteriors are rendered in Isaac Sim 6.2 RTX from real swisstopo 3D geometry; each interior digital twin displays live liquid-cooled server sizing, Liebert CDU exchange, buffer tank and flowing thermal load.
[3:45-4:25] When we scale from one Basel home to 14 photorealistic digital twins across 10 Swiss cantons, the economics transform. In January every site absorbs 100% of its heat locally. In summer, warm homes in Bern, Geneva and Lugano shift nearly 40,000 server-hours per year to outdoor pools, Emmental dairy farms, Thurgau glasshouses and St. Gallen hospital, cutting 31.7 MWh of summer waste to zero and lowering fleet OPEX break-even to 19.3 Rappen per server-hour.
Every real-world tariff, idle-power and thermal boundary is engineered directly into our deployment sequence
Anchoring first deployments on commercial C6 tariffs (0.253 CHF/kWh vs retail H4 0.333) saves 8.0 Rp/kWh, keeping 100% of the displaced heat credit intact and achieving 0.386 CHF/srv-h incl. CAPEX.
HEATWISE measurements show 21-27% idle power under OS pause alone. CalorGrid combines hardware sleep states with 35-45 C warm-water sinks so servers stay cool (54.9 C) and reach 96-100% utilisation.
Built-in 4-category 3D Provenance Mode labels live OS telemetry, pinned datasets and OEM CAD models transparently, backed by 77/77 unit tests and 125/125 Chrome E2E tests.
Attributions: (c) swisstopo; (c) OpenStreetMap contributors, ODbL; (C) buildingSMART International Ltd. (CC BY 4.0); NVIDIA, Vertiv and ForgeCAD SimReady (CC BY 4.0). Literature: ETH Zurich Aquasar (2012).
[4:25-5:00] Because we modelled Swiss tariffs, HEATWISE idle draw and ASHRAE W45 temperatures from the outset, our rollout path is directly actionable. Phase 1 launches at year-round commercial hosts in Basel at 38.6 Rappen all-in. Phase 2 adds borehole-paired residential homes at 96% utilisation. Phase 3 scales our fiber scheduler nationwide. Scan the QR code to try the live CalorGrid platform yourself. We are Team Invincibles. We are ready for your questions.