Heat-native AI infrastructure.

Sustainable AI is an engineering problem

Every watt used by AI compute eventually becomes heat. The question is what we do with it.

Give the energy another use

High-density AI infrastructure draws substantial electrical power and produces heat of roughly the same magnitude. Cooling keeps the equipment within its operating limits. The choice of cooling system also determines where the heat goes.

Many data centres use evaporation because it removes heat effectively. Water leaves the system as vapour and must be replaced, potentially in large volumes. When heat is rejected to the atmosphere, its opportunity to warm a building or industrial process is lost.

Eldrim intends to design around a different order of priorities: capture heat at a useful temperature, find a local use, store surplus where feasible, and reject the remainder safely.

Electricity

AI compute

Useful AI workResults delivered to the customer

Recoverable heat

Direct heatingUse heat while there is demand

Thermal storageAccept surplus where feasible

Dry coolingFallback for the remaining heat

Intended heat-use priority. Compute produces useful results and recoverable heat. Heat is used directly first; surplus goes to storage, with dry cooling for heat neither can accept. These are alternative destinations, not a requirement to send all heat through every stage.

Closed-loop liquid cooling

Water is an effective heat-transfer fluid. Inside a sealed cooling loop, the same water circulates between the equipment and a heat exchanger. That circulation is different from evaporating water to remove energy from the site.

Water used to move heat ≠ water consumed to reject heat.

Our objective is warm-water direct-to-chip cooling, carrying heat from processors and accelerators into an isolated facility loop. We intend to avoid routine evaporative cooling wherever practical. Leaks, maintenance and initial filling still matter; a closed-loop design alone does not establish zero water consumption.

Preserve the temperature

Heat becomes more useful when it leaves the computer at a temperature a customer can use. Losing temperature between the chips and the heating system can mean spending more electricity to raise it again.

Eldrim intends to design compute systems around warm-water operation, with coolant temperatures, heat exchangers and pumping considered alongside memory bandwidth, networking and power. The aim is to capture heat in a useful form while keeping the hardware reliable.

Equipment operating limits and warranties set constraints. We have not selected a coolant temperature or demonstrated the heat-recovery performance of an Eldrim system.

Use heat locally

AI compute

Hot water

Building, industrial process or district heating

First priority: deliver current compute heat to a nearby customer who can use it.

Connecting a modular compute installation to district heating is an important milestone in the Eldrim roadmap. A nearby network can offer a destination for heat, provided its temperatures, demand and operating schedule fit.

Haparanda is a feasibility study for this relationship. Pipe routes, hourly heat demand, delivery temperatures, connection costs and commercial terms need evidence before we can claim a viable heat customer. The compute business must also work without assuming heat revenue.

Heat-recovery concept showing compute containers, a utility area and proposed pipe routes leaving the site — Haparanda · heat-recovery concept **Give the heat a route out.** This view explores where heat-handling equipment and a connection to a local heating system could fit. The routes and equipment are design assumptions; heat delivery and recovery performance remain unproven. View sketch fitted to your screen.

Store heat when demand does not match production

AI workloads can run throughout the year, while Nordic heating demand changes with the seasons. A tank can buffer short changes in demand. Moving summer heat into winter calls for storage on a different scale.

We are researching borehole thermal energy storage, or BTES. A field of boreholes transfers heat into a volume of rock, then allows some of that stored heat to be recovered later. Feasibility depends on geology, groundwater, temperature and losses over repeated seasons.

Summer

AI compute

Surplus heat

Rock storageCharge a borehole field

Winter

Current AI heat

Stored summer heatHeat pump if a higher temperature is needed

HeatingCurrent heat and recovered heat combine

Research concept, not an installed Eldrim system. Winter compute continues producing heat; storage supplements that supply with recovered summer heat.

When stored heat is below the required delivery temperature, a heat pump can raise it. The pump consumes electricity, which belongs in the energy account. Storage losses, pumping and useful heat delivered all need measurement; energy sent underground is not automatically energy reused.

Dry cooling as the last resort

The system needs guaranteed heat rejection even when a heat customer stops taking delivery or storage is full. Compute safety cannot depend on a heating customer always being available.

Our desired fallback is dry cooling: transfer heat to outside air without routine water evaporation. Climate, equipment operating limits and peak summer conditions determine where that is practical. Fan and pumping electricity still count toward the total.

Measure it

We want sustainability claims to be measurable. The following are intended operating metrics, not results Eldrim has achieved. Comparisons will need a defined site boundary, workload, reporting period and measurement method.

Electricity per unit of useful AI work: Energy for a stated workload and service level, with the model and hardware identified.
Litres of water consumed: Water leaving the cooling system and requiring replacement, distinguished from water circulating inside it.
Percentage of IT heat captured in liquid: The fraction of compute heat reaching the liquid loop rather than the surrounding air.
Coolant return temperature: The temperature available for heat delivery, recorded alongside load and flow.
Percentage of annual heat reused: Useful heat delivered relative to IT heat produced, with stored heat accounted for across reporting periods.
Heat sent to seasonal storage: Thermal energy entering the store, measured separately from direct delivery.
Heat recovered from storage: Energy extracted later, with losses and heat-pump electricity recorded separately.
Cooling and pumping overhead: Electricity used by the cooling system, pumps, fans and any heat upgrading.

Power usage effectiveness, PUE, compares total facility electricity with IT electricity. It is useful for understanding overhead, but covers only part of this system.

PUE alone does not tell us whether heat was useful or whether water was consumed to get rid of it.

Carbon removal, from the first payment

Most of this page describes where we are heading. One commitment applies today: Eldrim directs 1 % of its revenue to permanent carbon removal through Stripe Climate, which funds removal projects chosen with scientific advisors.

This pays for removing carbon dioxide from the atmosphere. It is separate from the emissions of the compute we sell, which we intend to measure and reduce as described above, and we do not describe Eldrim as carbon neutral.

Build toward it, one step at a time

Sell other providers' computeOur current commercial direction. Learn which workloads customers will pay for before buying hardware.
Eldrim modular computeA future container or modular installation, justified by demand, site evidence and financing.
Connect to district heatingValidate useful heat delivery with a local energy partner.
Eldrim-designed compute hardwareUse operating evidence to design around memory, networking, power and warm-water cooling.
Heat-native hardware and storageDesign compute, heat recovery and reuse together. Seasonal storage remains a research question.

Each step should fund, validate and reduce the risk of the next. Owning more of the infrastructure would give us more control over these choices, which is why this work connects to our sovereignty roadmap.

The goal is simple: turn electricity into useful compute, preserve the resulting heat, use it again, and consume as little water as possible throwing heat away.

This is the direction Eldrim is designing toward. Haparanda is a feasibility case; modular compute, district-heat integration and seasonal storage are future work.