# The energy landscape

Digital growth is also a physical infrastructure problem: power, cooling, capacity and time.

Type: Research · Reviewed: 4 October 2026

## Computation has a physical footprint

Every digital request becomes work performed by chips, memory, storage and networking equipment. The facility also needs power delivery and environmental control. An efficient accelerator is therefore only one part of an efficient service.

The IEA describes demand across servers, storage, networking and cooling, with a composition that varies between facilities. Its 2025 outlook highlights a timing mismatch: compute deployments can move faster than the energy infrastructure needed to support them.

Source: [Energy and AI: energy demand](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai) — IEA · 2025. Data-centre demand estimates and scenario-based forecasts. Sector totals include workloads beyond intelligence.

## Growth and efficiency are different variables

Total demand can rise even while individual operations become cheaper. If energy per task falls by 20% but task volume doubles, combined energy consumption rises by 60%, before changes in overhead. This is arithmetic, not a project forecast.

Track both intensity and scale: energy per accepted result and energy consumed across all results. A system can improve the first metric while worsening the second.

## Keep four objectives separate

| Objective | What improves it | What it does not establish |
| --- | --- | --- |
| Energy efficiency | Less electricity for equivalent work | Lower emissions in every location. |
| Carbon intensity | Cleaner electricity at the time of use | Lower absolute energy demand. |
| Reliability | Resilience and sufficient capacity | Resource efficiency. |
| Economics | Lower delivered cost and better utilization | Lower lifecycle impact. |

## Flexibility is a resource

Some work must finish immediately; other work can move within a deadline. Radovanović and colleagues describe scheduling flexible compute around changing grid carbon intensity while respecting capacity constraints. This is an operational mechanism rather than a replacement for adequate power supply.

Our design inference is to separate interactive work from flexible batch work before assigning resources. Scheduling decisions need explicit deadlines, location restrictions and data-transfer costs.

Source: [Carbon-Aware Computing for Datacenters](https://arxiv.org/abs/2106.11750) — Radovanović et al. · 2021 paper. Scheduling flexible compute around time-varying grid carbon intensity.

## Implications for SpaceXSI

- Make energy an input to workload decisions, not only a dashboard after execution.
- Compare equivalent service levels rather than headline chip specifications.
- Report absolute demand alongside per-task savings.
- Treat available power, local impacts and supply constraints as design inputs.