For compute buyers
GPU capacity from infrastructure that sells its heat instead of dumping it.
HeatLoop runs liquid-cooled NVIDIA accelerators inside Estonian buildings that buy the recovered heat. Same silicon, EU jurisdiction, and a cost structure with one fewer expense in it — because the cooling is somebody else's heating.
There is no capacity to sell today. The first production rack is scheduled for commercial operation in the first quarter of 2027, and we are building a waitlist to size it against real demand. If you tell us what you need now, it shapes what we buy — and waitlist members get first refusal and pilot pricing.
What makes this different
The heat is a product, not a cost
Conventional facilities spend capital and power on rejecting heat. We sell ours to the building the rack sits in. That removes a cost line rather than passing it through to you.
Inside the EU
Hardware, operations and company are Estonian. For teams with data-residency requirements, that is a shorter compliance conversation than a non-EU region flag.
You talk to the operator
We are small and intend to stay direct. Capacity questions, scheduling and incidents go to the people who run the fleet, not to a ticket queue.
Hardware
Racks are populated with NVIDIA workstation-class RTX PRO accelerators and data-centre-class DGX systems, sourced new and certified-refurbished through authorised channel partners.
Each rack draws about 40 kW fully populated and is liquid-cooled end to end. Exact configuration for the first production rack will be set by what the waitlist actually asks for, which is the point of running one.
- Model training and fine-tuning
- Multi-day runs on dedicated hardware with persistent storage, reserved for the length of the job.
- Inference
- Steady-state serving on committed capacity, with latency inside the EU for European users.
- Rendering and simulation
- Frame and batch queues from studios and engineering teams — naturally parallel, naturally schedulable.
- Checkpointable batch
- Jobs that can pause and resume run on interruptible capacity at a lower rate. This is the same flexibility we sell to the grid, which is why it is cheaper.
Two ways to buy
Rates are set once the first rack's configuration and power contract are fixed. We would rather quote you a real number later than an invented one now.
Committed capacity
Reserved GPUs for a fixed term. Predictable performance and predictable cost, for training runs and production inference.
- Dedicated hardware
- Fixed monthly rate
- Persistent storage
- Term from one month
Interruptible capacity
Lower rate in exchange for jobs that checkpoint and resume. Suited to rendering queues and batch training where wall-clock time is flexible.
- Reduced rate
- Advance notice before pause
- Checkpoint-and-resume required
- Billed per GPU-hour used
Join the pilot waitlist
Tell us what you would run and we will tell you honestly whether the first rack will suit it. No obligation, and it genuinely influences what we buy.
- What you would run — training, inference, rendering or batch
- Roughly how many GPUs, and for how long
- Whether your jobs can checkpoint and resume
- Any data-residency or compliance requirements
- When you would need capacity
info@heatloop.com