Metahive Builder is a 3D planning tool for compute-site planners, so that they see whether space, power, cooling or memory runs out first, before buying hardware.
Place devices within racks, rooms, buildings and plot with sourced assumptions.
Tagged equipment and site assumptions.
One explicit revised equipment assumption; preserve the original baseline.
Place devices within racks, rooms, buildings and plot with sourced assumptions.
Fit, aisle, mass, floor load and site assumptions.
Original configuration and changed input remain identifiable.
Physical fit and site assumptions retain their source and units.
Keep distinct power bases and propagate losses, heat and cooling demand.
Equipment/power assumptions with containment context.
Keep distinct power bases and propagate losses, heat and cooling demand.
Nameplate, provisioned peak and utilization basis.
IT heat, conversion losses and facility demand.
Nameplate, provisioned peak and utilization are different power bases.
Derive memory and throughput bounds from the workload and selected equipment.
Selected capacity and placement.
Model shape, precision, context and concurrency.
Derive memory and throughput bounds from the workload and selected equipment.
Weights, workspace, KV memory, precision and concurrency.
Compute/memory bounds and selected utilization assumptions.
An infeasible memory requirement prevents an operating-energy result for that configuration.
Balance load, supply, storage and grid under the selected time basis.
Calculated facility load under the selected power basis.
Duration/weather/supply/storage assumptions.
Balance load, supply, storage and grid under the selected time basis.
Load, generation, exchange, shortfall and storage state.
Base/low/high inputs and their source quality.
Time-based supply and storage remain balanced, including shortfall.
Join physical and workload results and expose the binding constraint.
Electrical and thermal feasibility.
Memory feasibility and workload bounds.
Calculation with its time basis, not an observed operating result.
Join physical and workload results and expose the binding constraint.
Feasible, failed and unresolved results.
Units, method, input lineage and comparison basis.
Calculated feasibility is an illustrative evaluation, not commissioned hardware performance.
Tagged equipment and site assumptions.
One explicit revised equipment assumption; preserve the original baseline.
Place devices within racks, rooms, buildings and plot with sourced assumptions.
Fit, aisle, mass, floor load and site assumptions.
Original configuration and changed input remain identifiable.
Physical fit and site assumptions retain their source and units.
Equipment/power assumptions with containment context.
Keep distinct power bases and propagate losses, heat and cooling demand.
Nameplate, provisioned peak and utilization basis.
IT heat, conversion losses and facility demand.
Nameplate, provisioned peak and utilization are different power bases.
Selected capacity and placement.
Model shape, precision, context and concurrency.
Derive memory and throughput bounds from the workload and selected equipment.
Weights, workspace, KV memory, precision and concurrency.
Compute/memory bounds and selected utilization assumptions.
An infeasible memory requirement prevents an operating-energy result for that configuration.
Calculated facility load under the selected power basis.
Duration/weather/supply/storage assumptions.
Balance load, supply, storage and grid under the selected time basis.
Load, generation, exchange, shortfall and storage state.
Base/low/high inputs and their source quality.
Time-based supply and storage remain balanced, including shortfall.
Electrical and thermal feasibility.
Memory feasibility and workload bounds.
Calculation with its time basis, not an observed operating result.
Join physical and workload results and expose the binding constraint.
Feasible, failed and unresolved results.
Units, method, input lineage and comparison basis.
Calculated feasibility is an illustrative evaluation, not commissioned hardware performance.
An illustrative design calculation under explicit equipment and workload assumptions.
Additional mechanisms are described in the technical reference.
The site hierarchy feeds electrical, thermal and workload calculations. Compare the resulting limits with the time-based energy scenario to identify the binding constraint.
Configure devices, racks, rooms, buildings and a plot, with containment, aisle and floor-load checks.
Follow IT load through electrical losses and heat to cooling, facility demand and energy supply.
Evaluate compute precision, model-memory limits, inference rooflines and scenario ranges on a common basis.
Move one rack. See what else changes.
Discuss a site scenarioWhat the research says
Regulators now ask for a site’s energy performance and plan to rate it, while the industry’s headline efficiency figure has stood still. A planner has to defend those numbers before the hardware arrives. Builder computes them from the layout and shows the assumptions behind each one.
A planned rack move can change equipment fit, electrical loading and cooling across the room. The planner needs to see those consequences while comparing the layout and the workload it is meant to support.
Use Builder while evaluating a fitout, a different device mix or a changed inference workload. Start with the physical hierarchy and sourced assumptions, then inspect which constraints change as the configuration changes. Its engineering and financial scenarios retain their input quality and assumptions.
Check equipment fit, weight, clearances and electrical demand.
Follow cooling, airflow, floor loading and containment across neighboring equipment.
Connect the building and plot to power paths, land use and operating assumptions.
A compute design must support both its physical demand and its intended workload.
Nameplate rating, provisioning peak, component load and utilization-dependent energy remain separate. PSU losses, redundancy and cooling add their own requirements.
Weights, KV cache, tensor-parallel placement, batch and concurrency determine memory needs. Compute and memory rooflines bound the workload a configuration can support.
Hourly scenarios connect generation, storage, grid caps and cooling. A favorable design-point ratio cannot stand in for an annual energy quantity.
A site planner changes one thing; Builder recomputes power, cooling and inference under the same assumptions. Illustrative.
Consider replacing a rack configuration while keeping the workload and ambient scenario fixed.
Update equipment or workload and retain the comparison basis.
Inspect the electrical, thermal or memory limit that prevents the intended operation.
A result keeps its units, method, input lineage and source quality. Assumptions and defaults remain available for scrutiny.
Builder produces calculated bounds and scenarios from its declared models and inputs. Hardware benchmarking is a separate activity.
It informs technical feasibility and operating assumptions. Financial feasibility also requires independently checked costs, demand, financing and contractual terms.
The technical choices that give Metahive Builder its particular behavior.
Metahive Builder couples a 3D site hierarchy with electrical, thermal, energy and inference scenarios. A configuration change reveals its constraints, and each calculation retains its units, method and input lineage.
Device → rack. Check equipment fit, weight, clearances and electrical demand.
Rack → room. Follow cooling, airflow, floor loading and containment across neighboring equipment.
Room → site. Connect the building and plot to power paths, land use and operating assumptions.
The energy model. Nameplate rating, provisioning peak, component load and utilization-dependent energy remain separate. PSU losses, redundancy and cooling add their own requirements.
The inference model. Weights, KV cache, tensor-parallel placement, batch and concurrency determine memory needs. Compute and memory rooflines bound the workload a configuration can support.
The operating day. Hourly scenarios connect generation, storage, grid caps and cooling. A favorable design-point ratio cannot stand in for an annual energy quantity.
Change one input. Update equipment or workload and retain the comparison basis.
Find the binding constraint. Inspect the electrical, thermal or memory limit that prevents the intended operation.
Open the calculation. A result keeps its units, method, input lineage and source quality. Assumptions and defaults remain available for scrutiny.
A proposed compute-site scenario showing the relationship between equipment, energy and inference.
Follow the storyWhich limit would your next site hit first?
Discuss a site scenario